<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
	<channel>
		<title>Ai on Dr. Hitesh Sarda</title>
		<link>https://hitesh.in/tags/ai/</link>
		<description>Recent content in Ai on Dr. Hitesh Sarda</description>
		<generator>Hugo</generator>
		<language>en-US</language>
		
		
		
		
			<lastBuildDate>Thu, 19 Nov 2026 00:00:00 +0000</lastBuildDate>
		
			<atom:link href="https://hitesh.in/tags/ai/index.xml" rel="self" type="application/rss+xml" />
			<item>
				<title>Sampling was a budget decision, and AI just changed the budget</title>
				<link>https://hitesh.in/2026/sampling-was-a-budget-decision/</link>
				<pubDate>Thu, 19 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/sampling-was-a-budget-decision/</guid>
				<description>&lt;p&gt;KPMG UK has been testing more than 100 AI agents on the financial statements of an unnamed FTSE 250 company, according to a &lt;a href=&#34;https://europeanbusinessmagazine.com/ai-agents-better-questions-kpmg/&#34;&gt;report in the European Business Magazine&lt;/a&gt; that draws on the Financial Times. The agents gathered evidence, documented analysis, drafted conclusions and divided the work among themselves, while a conventional audit team ran alongside. KPMG&amp;rsquo;s head of audit, Cath Burnet, says the agents sometimes went considerably deeper than the human team and raised questions the auditors had not asked.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your AI-built prototype becomes a liability the day it holds real customer data</title>
				<link>https://hitesh.in/2026/ai-built-app-real-customer-data/</link>
				<pubDate>Sun, 15 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-built-app-real-customer-data/</guid>
				<description>&lt;p&gt;In November, the security firm Crissy Field, working with the newspaper Handelsblatt, &lt;a href=&#34;https://www.crissyfield.de/research/lovable-en/&#34;&gt;published an analysis of 159,041 apps built on Lovable&lt;/a&gt;, the AI platform that turns plain-language descriptions into web applications. Of the 34,545 that used the Supabase database, 24,664 sent database records to any visitor&amp;rsquo;s browser as soon as the page loaded. In 7,403 of those, a few extra read-only requests returned tables containing email addresses.&lt;/p&gt;&#xA;&lt;p&gt;The authors are careful to say this is not a confirmed leak rate. Some of that data is meant to be public, and the sample is not representative. The examples they did find were not ambiguous: 80,000 forum profiles and 200,000 chat messages, tenant emails summarized by AI, and more than 1,000 records with passport numbers. No login, no exploit.&lt;/p&gt;</description>
			</item>
			<item>
				<title>A ban on AI agents is a statement about your margins</title>
				<link>https://hitesh.in/2026/ban-ai-agents-margin-statement/</link>
				<pubDate>Thu, 12 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ban-ai-agents-margin-statement/</guid>
				<description>&lt;p&gt;United Airlines has said it will &lt;a href=&#34;https://www.cbsnews.com/news/united-airlines-bans-ai-agents-booking-flights/&#34;&gt;prohibit unauthorized third-party personal AI agents&lt;/a&gt; from booking flights, on its own channels and through agencies and aggregators. Its stated reasons are sensible ones: handing account credentials to an assistant can expose sensitive information, an agent may buy something or spend MileagePlus miles the traveler never meant to, and some federal rules expect a human to confirm things like what is in their luggage. Meanwhile, according to the same report, Southwest has made its inventory viewable to ChatGPT users.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Be more ambitious with software, but only where you can prove it works</title>
				<link>https://hitesh.in/2026/ambition-needs-a-test-suite/</link>
				<pubDate>Tue, 10 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ambition-needs-a-test-suite/</guid>
				<description>&lt;p&gt;A few weeks ago, an engineer I respect wrote that &lt;a href=&#34;https://danluu.com/ambitious/&#34;&gt;we should be more ambitious with the software we build&lt;/a&gt;. His evidence was personal: projects that once took a year or more, finished in a handful of hours with an AI agent doing the typing. I think the premise is right, and most leadership teams will draw the wrong conclusion from it.&lt;/p&gt;&#xA;&lt;p&gt;Ambition got cheap. Knowing whether the result is correct did not.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Goldman limited coding agents by job title. Limit them by blast radius.</title>
				<link>https://hitesh.in/2026/coding-agents-blast-radius-job-title/</link>
				<pubDate>Mon, 09 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/coding-agents-blast-radius-job-title/</guid>
				<description>&lt;p&gt;In October, Goldman Sachs &lt;a href=&#34;https://www.thestateofai.com/news/goldman-sachs-claude-code-traders-bankers&#34;&gt;updated its internal guidance&lt;/a&gt; so that coding agents such as Claude Code and GitHub&amp;rsquo;s tools stay with its engineers, about 12,000 of roughly 47,000 staff. Most traders and bankers lost access. I&amp;rsquo;m relying on that report of Bloomberg&amp;rsquo;s original story, and I could not read the Bloomberg piece itself, so treat the details as secondhand.&lt;/p&gt;&#xA;&lt;p&gt;The decision is reasonable. The way it was drawn is the interesting part, because it sorts people by job title when the risk sits somewhere else.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The real price of an agent sandbox is the team that patches it</title>
				<link>https://hitesh.in/2026/agent-sandbox-patch-speed/</link>
				<pubDate>Sat, 07 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/agent-sandbox-patch-speed/</guid>
				<description>&lt;p&gt;A coding agent writes code and then runs it, often before any person has read it. That is why companies building agents end up needing a sandbox, a sealed environment where the code can run without touching anything that matters. Replit does this for hundreds of millions of projects, and it recently &lt;a href=&#34;https://frontier.replit.com/sandbox-engine-microvms-part-1/&#34;&gt;described how it rebuilt its sandbox engine&lt;/a&gt;. The part I keep coming back to is a maintenance number. Between June 1 and November 4 it went through 6,211 Linux kernel vulnerabilities to work out which ones could affect its setup, and it reports that about 96% did not, either because the fix had landed upstream before disclosure or because its kernel configuration left the vulnerable code out.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI just raised the interest rate on your technical debt</title>
				<link>https://hitesh.in/2026/ai-raised-the-interest-on-technical-debt/</link>
				<pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-raised-the-interest-on-technical-debt/</guid>
				<description>&lt;p&gt;In late October, PHP shipped &lt;a href=&#34;https://bleepingcomputer.com/news/security/php-ships-emergency-patches-for-critical-php2shell-flaws&#34;&gt;emergency patches&lt;/a&gt; for a flaw now called PHP2Shell, which put sites from WordPress installs to government pages at risk. The interesting part is the history. According to &lt;a href=&#34;https://spectrum.ieee.org/ai-in-cybersecurity-php2shell-vulnerability&#34;&gt;IEEE Spectrum&lt;/a&gt;, the underlying bug was an out-of-memory error reported as early as 2024. Out-of-memory errors normally just crash a program, so it was treated as low priority and left alone.&lt;/p&gt;&#xA;&lt;p&gt;Then a security firm, Calif, offered to audit PHP with OpenAI&amp;rsquo;s sponsorship. Within two weeks, a researcher using AI tools found a way to turn that crash into arbitrary code execution on the host server. Patching took four engineers two months.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your coding agent now takes instructions from your software vendors</title>
				<link>https://hitesh.in/2026/vendors-write-your-agents-instructions/</link>
				<pubDate>Wed, 04 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/vendors-write-your-agents-instructions/</guid>
				<description>&lt;p&gt;On October 25 a developer &lt;a href=&#34;https://github.com/postmanlabs/postman-app-support/issues/14370&#34;&gt;opened an issue&lt;/a&gt; against Postman&amp;rsquo;s support repository. After an automatic update, the Postman desktop app had installed a CLI and plugins into their Codex and Claude Code setups. Postman&amp;rsquo;s own log reads &lt;code&gt;auto-existing-upgrade&lt;/code&gt;, with &lt;code&gt;claude-code:installed&lt;/code&gt; and &lt;code&gt;codex:installed&lt;/code&gt;. A session-start hook came along with them, whose text tells the agent to load a Postman skill &amp;ldquo;before responding to any non-trivial API engineering task.&amp;rdquo;&lt;/p&gt;&#xA;&lt;p&gt;Postman has since said the components are bundled because people struggled to install them separately, that it is disclosed on the download page and in the update flow, and that it is &lt;a href=&#34;https://github.com/postmanlabs/postman-app-support/issues/14370&#34;&gt;reviewing the hook approval and session-start behavior&lt;/a&gt;. The reporter is explicit that no data theft was shown, and I could not independently confirm the details beyond the issue and that thread. I don&amp;rsquo;t think Postman is the villain here. What matters is where this points for every other vendor.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your release date assumed bugs were hard to find</title>
				<link>https://hitesh.in/2026/release-date-assumed-bugs-were-hard-to-find/</link>
				<pubDate>Mon, 02 Nov 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/release-date-assumed-bugs-were-hard-to-find/</guid>
				<description>&lt;p&gt;For a decade, PostgreSQL major releases have landed between mid-September and mid-October. &lt;a href=&#34;http://rhaas.blogspot.com/2026/10/postgres-19-is-late-and-its-ais-fault.html&#34;&gt;Postgres 19 will most likely ship in November&lt;/a&gt;, and between August 27 and September 15 the project pulled three large features it had planned to include: UPDATE/DELETE FOR PORTION OF, SQL property graph queries, and partition merge and split. Robert Haas, a long-time Postgres committer, gives the cause in one line: &amp;ldquo;The short answer is that LLMs happened.&amp;rdquo;&lt;/p&gt;&#xA;&lt;p&gt;The code did not get worse. It got looked at. The bugs were always there, and what changed, in his words, is that it has become &amp;ldquo;vastly easier to find those bugs.&amp;rdquo; I think that sentence applies well beyond databases.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Assume your platform vendor will lose your secrets</title>
				<link>https://hitesh.in/2026/assume-vendor-loses-your-secrets/</link>
				<pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/assume-vendor-loses-your-secrets/</guid>
				<description>&lt;p&gt;In thirteen days, three platforms that hold other companies&amp;rsquo; secrets told their customers to rotate them. OpenCode &lt;a href=&#34;https://www.paubox.com/blog/opencode-console-breach-puts-openai-and-other-api-keys-at-risk&#34;&gt;confirmed on October 17&lt;/a&gt; that attackers had reached its Console service, where users store keys for OpenAI, Anthropic and Google. Railway &lt;a href=&#34;https://blog.railway.com/p/security-disclosure-wednesday-october-21st&#34;&gt;disclosed on October 21&lt;/a&gt; that an unauthorized party may have decrypted some customer environment variables. On October 30, &lt;a href=&#34;https://status.northflank.com&#34;&gt;Northflank&amp;rsquo;s status page&lt;/a&gt; went further: &amp;ldquo;Assume all environment variables have been decrypted. Disable/rotate all keys.&amp;rdquo; Customers were already reporting that their AI keys had been used by someone else.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Run your next incident drill with the AI turned off</title>
				<link>https://hitesh.in/2026/ai-off-incident-drill/</link>
				<pubDate>Sat, 31 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-off-incident-drill/</guid>
				<description>&lt;p&gt;A study in The Lancet Gastroenterology &amp;amp; Hepatology, cited in a recent &lt;a href=&#34;https://newsletter.signoz.io/p/ai-isnt-replacing-sres-its-deskilling&#34;&gt;SigNoz essay on SRE work&lt;/a&gt;, found that &lt;a href=&#34;https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/abstract&#34;&gt;endoscopists&amp;rsquo; unassisted polyp detection&lt;/a&gt; fell from 28% to 22% after a period of working with AI assistance. They had not forgotten any medicine. They had stopped using the skill.&lt;/p&gt;&#xA;&lt;p&gt;SigNoz asks the same question of site reliability engineers: if the AI resolves the routine incidents, the humans only meet the hard ones, with less practice than before. That is Lisanne Bainbridge&amp;rsquo;s 1983 argument about automation, and I made the general case for it &lt;a href=&#34;https://hitesh.in/2026/automation-is-cheap-understanding-is-the-bill&#34;&gt;in July&lt;/a&gt;. This post is about the operational version, because on-call is where the cost of getting it wrong shows up as an outage.&lt;/p&gt;</description>
			</item>
			<item>
				<title>DoorDash is paying $17 billion for the part of Instacart an AI agent can&#39;t skip</title>
				<link>https://hitesh.in/2026/instacart-what-agents-cant-skip/</link>
				<pubDate>Fri, 30 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/instacart-what-agents-cant-skip/</guid>
				<description>&lt;p&gt;Instacart was valued at &lt;a href=&#34;https://www.cnbc.com/2021/03/02/instacarts-valuation-doubles-to-39-billion.html&#34;&gt;$39 billion&lt;/a&gt; in March 2021. It went public in 2023 at &lt;a href=&#34;https://www.euronews.com/business/2023/09/19/instacart-sets-ipo-price-at-30-a-share-valuing-the-company-at-about-10-billion&#34;&gt;about $10 billion&lt;/a&gt;. This week DoorDash agreed to &lt;a href=&#34;https://about.doordash.com/en-us/news/doordash-to-acquire-instacart&#34;&gt;buy it for roughly $17 billion&lt;/a&gt;, a 26% premium to where the stock traded before the talks leaked.&lt;/p&gt;&#xA;&lt;p&gt;Most of the coverage is about grocery market share and antitrust. I read it differently. The price is a buyer&amp;rsquo;s estimate of which part of Instacart stays scarce once software does the shopping, and every business that sells through an app or a website should be able to answer that question about itself.&lt;/p&gt;</description>
			</item>
			<item>
				<title>A limited shell account is only as limited as the software enforcing it</title>
				<link>https://hitesh.in/2026/limited-access-limited-enforcement/</link>
				<pubDate>Wed, 28 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/limited-access-limited-enforcement/</guid>
				<description>&lt;p&gt;On October 25, a researcher named Erica Windisch published a report called Tailworm, describing 11 findings in Tailscale and the open-source headscale server, most of them in Tailscale SSH. Her lab tests, run on &lt;a href=&#34;https://thehackernews.com/2026/10/unpatched-tailscale-ssh-flaws-could-let.html&#34;&gt;Tailscale 1.104.2&lt;/a&gt;, showed a person who is allowed to log in as an ordinary user reaching root on the same machine. Tailscale has not confirmed every claim in the report. Its &lt;a href=&#34;https://tailscale.com/ssh-investigation-update&#34;&gt;own update&lt;/a&gt; calls the issues serious, says most of them are privilege escalation, and promised a patched client by the end of October 27. Check its security bulletins for where that stands.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Microsoft is trading payroll for compute. Anthropic&#39;s S-1 is the other side of the trade</title>
				<link>https://hitesh.in/2026/payroll-for-compute/</link>
				<pubDate>Tue, 27 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/payroll-for-compute/</guid>
				<description>&lt;p&gt;On Monday Microsoft &lt;a href=&#34;https://www.businessinsider.com/microsoft-is-offering-buyouts-read-the-memo-2026-10&#34;&gt;offered voluntary exits&lt;/a&gt; to about 60,000 employees, more than a quarter of the company, at level 70 and below in the US, Canada, India and the UK. The same week, Anthropic&amp;rsquo;s &lt;a href=&#34;https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm&#34;&gt;S-1&lt;/a&gt; showed revenue of $4.6 billion for all of 2025 and $16.3 billion for the first six months of 2026. Six months of this year earned more than three times what all of last year did.&lt;/p&gt;&#xA;&lt;p&gt;The memo does not mention AI. Business Insider links the two, noting that Microsoft has been spending heavily on AI infrastructure and reducing headcount at the same time. That is a reporter&amp;rsquo;s inference and so is what follows, but I think it is a reasonable one: a large company is converting payroll into compute, and the S-1 is a look at the other side of that trade.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your best AI use case may not need a frontier model</title>
				<link>https://hitesh.in/2026/your-best-ai-use-case-may-not-need-a-frontier-model/</link>
				<pubDate>Mon, 26 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-best-ai-use-case-may-not-need-a-frontier-model/</guid>
				<description>&lt;p&gt;Dermatologist Magnus Lynch showed &lt;a href=&#34;https://lynchlab.net/llm-skin-cancer&#34;&gt;200 biopsy-proven skin lesions&lt;/a&gt; to GPT-6 Astra and Claude Fable 5.1, then to a ten-year-old image network (Inception v3) fine-tuned for about twenty minutes on one GPU. The old network named the right diagnosis 62% of the time. Both frontier models got 42%.&lt;/p&gt;&#xA;&lt;p&gt;It is tempting to read that as a gotcha about LLMs. I think it is more useful as a procurement story, because it is what happens when a company picks the model first and the task second. Lynch himself says the old network is only a baseline, that regulated products do far better, and that 200 cases from a single run can&amp;rsquo;t rank close results. None of that weakens the lesson.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The open-weight model that was trained to know less about security</title>
				<link>https://hitesh.in/2026/open-weight-model-trained-to-know-less/</link>
				<pubDate>Sat, 24 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/open-weight-model-trained-to-know-less/</guid>
				<description>&lt;p&gt;When Moonshot released &lt;a href=&#34;https://www.kimi.ai/blog/kimi-k3-1-pro&#34;&gt;Kimi K3.1 Pro&lt;/a&gt; this week, the benchmark table was the usual one: 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5, a 1-million-token context window, weights promised by November 4. The line that matters for a buyer is further down. &amp;ldquo;For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages.&amp;rdquo;&lt;/p&gt;&#xA;&lt;p&gt;They also published what it cost. Without the removal, the checkpoint solves 90.5% of their cyber benchmark problems at pass@3. The final model solves 81.0%. That is a vendor-reported number on a vendor-chosen benchmark, so treat it as a direction, not a measurement. But it is a rare thing: a lab telling you exactly how much capability it took out on purpose.&lt;/p&gt;</description>
			</item>
			<item>
				<title>LiteLLM was compromised twice in seven months. Here is what that means for your AI strategy</title>
				<link>https://hitesh.in/2026/your-ai-gateway-holds-every-key/</link>
				<pubDate>Thu, 22 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-ai-gateway-holds-every-key/</guid>
				<description>&lt;p&gt;On Wednesday at 11:39 UTC, someone published LiteLLM 1.107.0 to PyPI. It was &lt;a href=&#34;https://securitylabs.datadoghq.com/articles/litellm-compromised-pypi-docker/&#34;&gt;quarantined at 13:20&lt;/a&gt;, so it was live for roughly two hours. In that window, anyone who ran &lt;code&gt;pip install litellm&lt;/code&gt; or pulled the &lt;code&gt;latest&lt;/code&gt; container image got a &lt;a href=&#34;https://safedep.io/litellm-pypi-compromise-credential-stealer-backdoor/&#34;&gt;credential stealer and a backdoor&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;LiteLLM is the open-source gateway many companies put between their applications and the model providers. One endpoint, virtual keys per team, spend tracking, routing across OpenAI, Anthropic and Bedrock. It was &lt;a href=&#34;https://www.netspi.com/blog/executive-blog/ai-ml-pentesting/litellm-supply-chain-compromise/&#34;&gt;compromised once before, in March&lt;/a&gt;, when it was pulling around 95 million downloads a month. Two incidents in seven months make this more than bad luck. They tell you where attackers think the value is.&lt;/p&gt;</description>
			</item>
			<item>
				<title>WhatsApp&#39;s help pages told AI not to mention regulators. Who approved that?</title>
				<link>https://hitesh.in/2026/hidden-prompt-is-policy/</link>
				<pubDate>Tue, 20 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/hidden-prompt-is-policy/</guid>
				<description>&lt;p&gt;Earlier this month, a researcher at Redact read all 545 pages of WhatsApp&amp;rsquo;s help center and found &lt;a href=&#34;https://redact.dev/blog/whatsapp-help-pages-hidden-ai-instructions&#34;&gt;102 blocks of text&lt;/a&gt; that no visitor could see but any AI system could. On the page about deleting an account, the hidden text told the AI not to recommend contacting a data protection authority or regulator. On pages touching legal topics, it told the AI not to cite data protection law, even when the user raised it first.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Nobody has benchmarked AI on 90% of the work your company does</title>
				<link>https://hitesh.in/2026/nobody-benchmarked-your-work/</link>
				<pubDate>Sat, 17 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/nobody-benchmarked-your-work/</guid>
				<description>&lt;p&gt;Mercor mapped 29 public AI benchmarks onto the US Department of Labor&amp;rsquo;s catalog of work, &lt;a href=&#34;https://www.onetonline.org/&#34;&gt;O*NET&lt;/a&gt;, which breaks 923 occupations into 18,797 tasks. Only &lt;a href=&#34;https://www.mercor.com/research/labor-index/benchmark-map/&#34;&gt;208 occupations have even one task with five or more benchmark prompts&lt;/a&gt;, and only 924 tasks are covered at all. Weighted by time, they estimate that about 90% of American working hours go to tasks with little or no benchmark coverage.&lt;/p&gt;&#xA;&lt;p&gt;Mercor says the numbers are provisional, and they are an estimate built on 29 benchmarks they chose. I would still treat the direction as reliable. It changes how I would read every AI claim that lands on an executive&amp;rsquo;s desk.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Anthropic replaced git because git assumes a human is typing</title>
				<link>https://hitesh.in/2026/tools-built-for-human-speed/</link>
				<pubDate>Wed, 14 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/tools-built-for-human-speed/</guid>
				<description>&lt;p&gt;Anthropic&amp;rsquo;s merges per day went from about 2,500 in May to over 30,000 by September. In &lt;a href=&#34;https://www.anthropic.com/engineering/redesigning-source-control&#34;&gt;its own account&lt;/a&gt;, source control became &amp;ldquo;our biggest development bottleneck.&amp;rdquo; Seven engineers spent five months replacing git for the monorepo with an in-house system called Scion.&lt;/p&gt;&#xA;&lt;p&gt;Most companies should not copy that. The more useful thing to copy is the diagnosis, because the thing that broke was not a model, a prompt, or an agent framework. It was a tool that had quietly assumed who its user was.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Most teams still ration AI that is cheap enough to waste</title>
				<link>https://hitesh.in/2026/ai-cheap-enough-to-waste/</link>
				<pubDate>Fri, 09 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-cheap-enough-to-waste/</guid>
				<description>&lt;p&gt;One developer spent a month &lt;a href=&#34;https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/&#34;&gt;using DeepSeek 4.1 Flash heavily across a dozen projects&lt;/a&gt; and reports that, mid-session, he can&amp;rsquo;t tell it from Opus. His day-long sessions rarely cost more than a dollar. That is one person&amp;rsquo;s subjective report, so weigh it that way. What I find more useful is the behavior he describes: &amp;ldquo;there is no shame now in spinning up mindless tasks,&amp;rdquo; including exploratory UI testing and tidying files, which costs $0.003 instead of $1.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The engineer who picks the model never sees the bill</title>
				<link>https://hitesh.in/2026/who-sees-the-ai-bill/</link>
				<pubDate>Thu, 08 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/who-sees-the-ai-bill/</guid>
				<description>&lt;p&gt;Microsoft had projected at least $1 billion of internal spending on Anthropic this year. According to &lt;a href=&#34;https://aiweekly.co/alerts/meta-microsoft-scale-back-staff-use-of-anthropics-claude&#34;&gt;The Information&amp;rsquo;s reporting&lt;/a&gt;, that estimate has since been cut by more than a third. At Meta, internal Claude Code users have reportedly fallen from about 60,000 to about 30,000. Neither company has confirmed the numbers, and Microsoft&amp;rsquo;s customer-facing spending on Anthropic models is reportedly still growing.&lt;/p&gt;&#xA;&lt;p&gt;The obvious reading is rivalry. Both companies build competing models and coding tools, so of course they would rather their own engineers used those. That is probably true, and it explains why these two moved. It tells the rest of us very little, because most companies have no in-house model to move to. What the story does show is what happens to an AI budget that nobody was managing: it runs wide open for a year, then gets cut by decree.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your coding agent has a default stack. Who chose it?</title>
				<link>https://hitesh.in/2026/your-agents-default-stack/</link>
				<pubDate>Mon, 05 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-agents-default-stack/</guid>
				<description>&lt;p&gt;Nolan Lawson, who has argued for years that developers should &lt;a href=&#34;https://nolanlawson.com/2026/10/03/why-dont-more-developers-use-the-platform/&#34;&gt;use the platform&lt;/a&gt; instead of reinventing it in JavaScript, spent a post this week taking the other side. His explanation for why so many developers resist is mostly about people. They reach for what they know. Documentation for libraries used to be better than documentation for browsers. And building your own modal dialog is more fun than adopting &lt;code&gt;&amp;lt;dialog&amp;gt;&lt;/code&gt;, a pleasure he calls the IKEA effect.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Rent the AI harness. Own what it learns.</title>
				<link>https://hitesh.in/2026/rent-the-harness-own-what-it-learns/</link>
				<pubDate>Sat, 03 Oct 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/rent-the-harness-own-what-it-learns/</guid>
				<description>&lt;p&gt;Shrivu Shankar argues that &lt;a href=&#34;https://blog.sshh.io/p/the-harness-is-the-company&#34;&gt;every SaaS business will become a harness around a model&lt;/a&gt;: the infrastructure, interfaces, context and state wrapped around an LLM that has no memory of its own. In his telling, companies move from engineers pairing with agents, to humans triggering background agents, to a harness that runs work proactively while people sample the output. His conclusion is that the harness stops being a tool you buy and becomes something you no more outsource than your product team or your sales team.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Why Google&#39;s AI products feel broken despite its models</title>
				<link>https://hitesh.in/2026/best-ai-still-feels-broken/</link>
				<pubDate>Mon, 28 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/best-ai-still-feels-broken/</guid>
				<description>&lt;p&gt;Google&amp;rsquo;s search results have started doing things a trillion-dollar product should have grown out of years ago. Paste a quote from a book and it reads the sentence as something you personally believe. Ask about a specific page and it refuses to actually open the link, then makes up what&amp;rsquo;s probably on it. Search for an exact phrase, the one thing Google search reliably did for two decades, and it now returns fuzzy matches tuned for ad inventory instead. A &lt;a href=&#34;https://sancho.bearblog.dev/google-weird/&#34;&gt;widely discussed post&lt;/a&gt; catalogued a long list of these this week, and the comments added a sharper detail: a former Googler describing a culture where engineers are actively discouraged from talking to real users, so &amp;ldquo;ship without testing, iterate later&amp;rdquo; became the normal plan.&lt;/p&gt;</description>
			</item>
			<item>
				<title>450 million seats, 7% adoption: the real Copilot story</title>
				<link>https://hitesh.in/2026/the-copilot-reboot/</link>
				<pubDate>Sat, 26 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-copilot-reboot/</guid>
				<description>&lt;p&gt;Two years ago Microsoft stood up a separate team, hired an outside AI veteran, Mustafa Suleyman, to run it, and gave the effort its own personality: a small animated character named Mico, a &amp;ldquo;Copilot Labs&amp;rdquo; for experiments, a feature for group chats with AI, another for AI-generated podcasts. The bet was that Microsoft could win a piece of the consumer AI-companion market the way OpenAI and Google were building one.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Google would rather solve rocket science than wait for a power line</title>
				<link>https://hitesh.in/2026/rocket-science-power-line/</link>
				<pubDate>Fri, 25 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/rocket-science-power-line/</guid>
				<description>&lt;p&gt;On October 1, a SpaceX Falcon 9 is due to carry four of Google&amp;rsquo;s TPUs into orbit under &lt;a href=&#34;https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/&#34;&gt;Project Suncatcher&lt;/a&gt;, the company&amp;rsquo;s bet that AI chips can run on satellites bathed in eight times more sunlight than any panel on Earth gets. The coverage has mostly treated this as a curiosity: can you really cool a chip in a vacuum, will radiation fry it, is this science fiction with a launch date. Those are fine questions. They are not the interesting one.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The most dangerous postmortem is the one that makes sense</title>
				<link>https://hitesh.in/2026/dangerous-postmortem-makes-sense/</link>
				<pubDate>Thu, 24 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/dangerous-postmortem-makes-sense/</guid>
				<description>&lt;p&gt;Michael Heap tells a story about an SVP who, five minutes into a detailed incident readout, stopped the meeting. &amp;ldquo;I already believe you did your best,&amp;rdquo; he said. &amp;ldquo;Now let&amp;rsquo;s talk about what happens next.&amp;rdquo; No interest in the timeline, no interrogation of who missed what. Heap &lt;a href=&#34;https://michaelheap.com/i-dont-want-the-details/&#34;&gt;wrote it up&lt;/a&gt; as a case for skipping the details, and the reaction split almost exactly down the middle: half the readers called it real leadership, delegation done right; half called it an executive signing off on fixes he had no way to evaluate. Both camps are arguing about trust. I think they are missing the more useful question, which is why a good explanation feels like enough in the first place, when it so rarely is.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Three frontier labs shipped this week. Only one cut its price.</title>
				<link>https://hitesh.in/2026/frontier-models-read-the-prices/</link>
				<pubDate>Wed, 23 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/frontier-models-read-the-prices/</guid>
				<description>&lt;p&gt;In the space of two days this week, three frontier labs updated their flagship line. Anthropic shipped &lt;a href=&#34;https://www.anthropic.com/claude-opus-5-5&#34;&gt;Claude Opus 5.5&lt;/a&gt; on September 22: 20% cheaper on input and output, 60% cheaper on cached reads, 30% faster, with real but incremental gains on agentic coding and computer-use benchmarks. OpenAI shipped &lt;a href=&#34;https://www.unite.ai/openai-introduces-gpt-6-sol-and-luna-with-50-lower-api-prices/&#34;&gt;GPT-6 Sol and Luna&lt;/a&gt; the same day, extending September&amp;rsquo;s GPT-6 Astra down into faster, cheaper tiers at roughly half of what the equivalent GPT-5.6 models cost. xAI shipped &lt;a href=&#34;https://www.digitalapplied.com/blog/grok-4-7-benchmarks-price-what-changed&#34;&gt;Grok 4.7&lt;/a&gt; a day earlier, ahead of its predecessor on most evaluations, though not all: Anthropic&amp;rsquo;s own Fable 5.1 Max still leads it on several coding and long-task tests.&lt;/p&gt;</description>
			</item>
			<item>
				<title>10 patterns where Jev, not your frontier model, should decide</title>
				<link>https://hitesh.in/2026/where-jev-shines/</link>
				<pubDate>Mon, 21 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/where-jev-shines/</guid>
				<description>&lt;p&gt;TypeSafe shipped something unusual this month: a model that never writes a sentence. &lt;a href=&#34;https://typesafe.ai/blog/introducing-system-one-models-and-jev&#34;&gt;Jev&lt;/a&gt; takes in a state and returns a typed decision, a choice, a score, or a probability, in roughly 70 to 500 milliseconds, for about four hundredths of a cent per call. Route the same decision through a frontier model and you&amp;rsquo;re paying &lt;a href=&#34;https://www.cloudraft.io/blog/top-use-cases-of-jev-typesafe-ai-model&#34;&gt;thirty to a few hundred times as much&lt;/a&gt; and waiting noticeably longer for it. It&amp;rsquo;s already showing up in real infrastructure, not just a launch post: &lt;a href=&#34;https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway&#34;&gt;Vercel added it to the AI Gateway&lt;/a&gt; within days.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your best people are the worst judges of your AI pilot</title>
				<link>https://hitesh.in/2026/best-people-worst-ai-judges/</link>
				<pubDate>Mon, 21 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/best-people-worst-ai-judges/</guid>
				<description>&lt;p&gt;A 2023 essay is doing the rounds again, &lt;a href=&#34;https://softwarecrisis.dev/letters/llmentalist/&#34;&gt;The LLMentalist Effect&lt;/a&gt;. Its claim is blunt: the intelligence you feel when you chat with a language model is a con, the same one a psychic runs. Cold reading. The model emits statistically likely sentences, you supply the meaning, and you leave convinced something understood you.&lt;/p&gt;&#xA;&lt;p&gt;Read in 2026, the strong version of that is wrong, and the pile-on in the comments says so. Psychics don&amp;rsquo;t write working code or close real tickets; today&amp;rsquo;s models do both often enough that &amp;ldquo;it&amp;rsquo;s all an illusion&amp;rdquo; is a comfortable story skeptics tell themselves. A &lt;a href=&#34;https://terrytao.wordpress.com/2026/09/19/why-do-we-need-human-mathematicians-anymore/&#34;&gt;recent guest post on Terence Tao&amp;rsquo;s blog&lt;/a&gt; has a working mathematician describing daily use of these tools on real work. Waving the whole category off as a parlor trick is just denial with better vocabulary.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Some writing is thinking. Automate the rest.</title>
				<link>https://hitesh.in/2026/some-writing-is-thinking/</link>
				<pubDate>Sun, 20 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/some-writing-is-thinking/</guid>
				<description>&lt;p&gt;Two people who agree on almost nothing about AI just landed on the same rule for writing with it.&lt;/p&gt;&#xA;&lt;p&gt;Thomas Ptacek is one of the loudest public defenders of building software with language models. Erich Grunewald wrote a piece this month arguing, plainly, that &lt;a href=&#34;https://erichgrunewald.substack.com/p/why-you-should-almost-never-use-ai&#34;&gt;you should almost never use AI to write&lt;/a&gt;. Ptacek says &lt;a href=&#34;https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/&#34;&gt;use the model but never as a ghostwriter&lt;/a&gt;: treat it as a copyeditor, and don&amp;rsquo;t keep a single word it suggests. Grunewald says don&amp;rsquo;t hand it the draft in the first place, though he&amp;rsquo;s fine with it editing one you wrote. Take the heat out of both and you get the same line. The model can polish your prose, but the sentences have to start as yours.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Gemini broke into three companies by guessing passwords</title>
				<link>https://hitesh.in/2026/the-attacker-just-got-cheap/</link>
				<pubDate>Sat, 19 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-attacker-just-got-cheap/</guid>
				<description>&lt;p&gt;In May, Google&amp;rsquo;s Gemini broke into three real companies. It wasn&amp;rsquo;t supposed to. &lt;a href=&#34;https://www.reuters.com/business/gemini-hacked-three-companies-first-known-breakout-by-google-ai-wsj-reports-2026-09-18/&#34;&gt;Google confirmed last week&lt;/a&gt; that during a security test run by the firm Irregular, the model was given a capture-the-flag exercise against a fictional target. The fake company happened to share its name with a real one, Gemini took the exercise out onto the live internet, and it got in. Google&amp;rsquo;s VP for security engineering &lt;a href=&#34;https://www.nbcnews.com/tech/tech-news/google-says-ai-model-gained-unauthorized-access-three-systems-rcna598651&#34;&gt;said&lt;/a&gt; the model &amp;ldquo;found public information online and guessed credentials.&amp;rdquo; In two of the three cases it used login details sitting in public code repositories. In the third it guessed passwords until one worked.&lt;/p&gt;</description>
			</item>
			<item>
				<title>When your AI vendor ships your product</title>
				<link>https://hitesh.in/2026/when-your-ai-vendor-ships-your-product/</link>
				<pubDate>Fri, 18 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/when-your-ai-vendor-ships-your-product/</guid>
				<description>&lt;p&gt;Last week OpenAI shipped &lt;a href=&#34;https://www.lawnext.com/2026/09/openai-releases-astra-for-law-a-gpt-6-model-configured-for-legal-work.html&#34;&gt;Astra for Law&lt;/a&gt;, a legal-research layer on its GPT-6 model with a search index over 230 million pages of US case law, statutes, and court rules. In the announcement it named Harvey and Legora, the two legal-AI companies most closely tied to OpenAI, as partners who would build on top of it.&lt;/p&gt;&#xA;&lt;p&gt;Read that twice. The supplier every legal-AI startup rents its intelligence from just shipped a product in the legal-AI category, and reassured its biggest customers in the same breath. When a vendor has to promise it won&amp;rsquo;t compete with you while announcing a product in your market, the promise is the risk, not the relief.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI deleted your excuse for missing fraud</title>
				<link>https://hitesh.in/2026/ai-killed-the-scale-excuse/</link>
				<pubDate>Mon, 14 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-killed-the-scale-excuse/</guid>
				<description>&lt;p&gt;A developer clicked a fake &amp;ldquo;iPhone Storage is Full&amp;rdquo; ad served through Google, reported it, and got back a note saying the ad complied with Google&amp;rsquo;s policies. He reported it again. Same answer. Then he &lt;a href=&#34;https://www.atomic14.com/2026/09/13/why-is-google-still-serving-dodgy-ads&#34;&gt;pasted the ad into Gemini&lt;/a&gt;, Google&amp;rsquo;s own model, and asked it to review the thing. Gemini disapproved it instantly: deceptive design, fake system UI, manufactured fear. Every violation was already in Google&amp;rsquo;s published policy.&lt;/p&gt;&#xA;&lt;p&gt;The obvious reading is that Google makes money from ads and loses money by removing them. That is true, and I have written before about how &lt;a href=&#34;https://hitesh.in/2026/ai-incentive-amplifier/&#34;&gt;AI amplifies whatever your scoreboard already rewards&lt;/a&gt;. But the incentive story is the old story. The new thing in that screenshot is harder to unsee: the same model any of us can rent for twenty dollars a month caught in one prompt what a trillion-dollar ad system said was fine. That does not just embarrass Google. It retires an excuse that platforms, banks, marketplaces, and moderation teams have leaned on for two decades.&lt;/p&gt;</description>
			</item>
			<item>
				<title>I designed my Ganpati wall with AI, then had it laser cut</title>
				<link>https://hitesh.in/2026/ganpati-wall-ai-design/</link>
				<pubDate>Mon, 14 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ganpati-wall-ai-design/</guid>
				<description>&lt;p&gt;Happy Ganesh Chaturthi! A break from the usual AI-and-business posts today, just a small home project I had fun with.&lt;/p&gt;&#xA;&lt;p&gt;We have a wall with ten square niches, built up over the years to hold Ganesha idols we&amp;rsquo;ve collected. It always looked a little bare, just plain cut-outs in the wall. I wanted a proper jharokha, the ornamental jaali frame you see on old havelis, around each one, but I had no design and no idea what it should look like.&lt;/p&gt;</description>
			</item>
			<item>
				<title>OpenAI couldn&#39;t say what its agents did. Neither can you.</title>
				<link>https://hitesh.in/2026/what-your-agent-did/</link>
				<pubDate>Sat, 12 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/what-your-agent-did/</guid>
				<description>&lt;p&gt;Between May and June 2026, a swarm of AI agents &lt;a href=&#34;https://www.rubyhack.ai/&#34;&gt;uploaded more than 2,000 malicious packages&lt;/a&gt; to RubyGems, the registry the entire Ruby ecosystem installs from. They bypassed email verification, abused RubyDoc.info&amp;rsquo;s documentation builder to &lt;a href=&#34;https://www.neowin.net/news/openai-agents-hijacked-rubygems-in-malicious-api-key-heist/&#34;&gt;run code on its servers&lt;/a&gt;, and used that foothold to scrape public data off UK council websites. RubyGems shut down new registrations for four days and called it a major attack. Researchers named it the GemStuffer campaign and, from self-identifying metadata and shared file-access patterns, traced it to OpenAI.&lt;/p&gt;</description>
			</item>
			<item>
				<title>When code gets cheap, your abstractions get expensive</title>
				<link>https://hitesh.in/2026/when-code-gets-cheap/</link>
				<pubDate>Fri, 11 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/when-code-gets-cheap/</guid>
				<description>&lt;p&gt;Shopify spent five years as one of React Native&amp;rsquo;s loudest champions. It contributed core libraries, shipped its flagship apps on it, and told the industry that cross-platform was the grown-up choice. Last week it &lt;a href=&#34;https://shopify.engineering/back-to-native&#34;&gt;went back to native Swift and Kotlin&lt;/a&gt;, rebuilding its Shop app from a proof of concept to a store-published native app in twelve weeks. The stated reason is AI: coding agents can now do enough of the mechanical work that maintaining two native codebases is no longer the deciding cost it was in 2020.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your next customer might be an AI agent, not a person</title>
				<link>https://hitesh.in/2026/your-next-customer-is-an-agent/</link>
				<pubDate>Wed, 09 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-next-customer-is-an-agent/</guid>
				<description>&lt;p&gt;Meta &lt;a href=&#34;https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/&#34;&gt;launched Muse&lt;/a&gt; this week, a personal AI agent that books your travel, pays bills through Stripe, orders online, and manages your calendar by reading your email, all from inside WhatsApp, an app with more than two billion users. Most of the coverage focused on how much of your life you&amp;rsquo;d have to hand Meta to get any of that, a fair question given the &lt;a href=&#34;https://techcrunch.com/2026/09/08/meta-debuts-its-muse-ai-agent-will-consumers-trust-it/&#34;&gt;$5 billion FTC penalty and $18 billion privacy settlement&lt;/a&gt; already on the record. It is not, I think, the question that should keep a founder or CTO up at night.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Who sits in the meeting when AI writes the memo?</title>
				<link>https://hitesh.in/2026/ai-cant-sit-in-the-meeting/</link>
				<pubDate>Mon, 07 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-cant-sit-in-the-meeting/</guid>
				<description>&lt;p&gt;Bryan Cantrill has a good phrase for the person who pastes a language model&amp;rsquo;s output onto LinkedIn and hits publish: &lt;a href=&#34;https://bcantrill.dtrace.org/2025/12/05/your-intellectual-fly-is-open/&#34;&gt;their intellectual fly is open&lt;/a&gt;. Everyone in the room can see it. The author is the only one who can&amp;rsquo;t. The emojis, the single-sentence paragraphs, the &amp;ldquo;it&amp;rsquo;s not just X, it&amp;rsquo;s Y&amp;rdquo; cadence. The tells are consistent enough that readers stop reading your argument and start reading the machine.&lt;/p&gt;&#xA;&lt;p&gt;I agree with him. But I think the fixation on the tells points people at the wrong problem. Clean up the tells with a better model or a careful editing pass, and the deeper exposure is still there. It just goes quiet until the moment it matters.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Someone built 215,000 pages to game your AI</title>
				<link>https://hitesh.in/2026/built-to-game-your-ai/</link>
				<pubDate>Thu, 03 Sep 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/built-to-game-your-ai/</guid>
				<description>&lt;p&gt;Three websites published 215,128 pages titled &amp;ldquo;best software,&amp;rdquo; most of them machine-generated, and an AI search engine treated a good chunk of them as evidence. That is the core finding of a &lt;a href=&#34;https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/&#34;&gt;Trellner Research investigation&lt;/a&gt; into where &lt;a href=&#34;https://www.perplexity.ai/&#34;&gt;Perplexity&amp;rsquo;s&lt;/a&gt; product recommendations come from, and it should sit uncomfortably with anyone who has started asking an AI which tool to buy.&lt;/p&gt;&#xA;&lt;p&gt;The numbers are worse than the headline. Trellner looked at 7,534 citations behind Perplexity&amp;rsquo;s software picks across 380 categories. Nearly 60 percent pointed to domains ranked below the top 100,000 sites on the &lt;a href=&#34;https://tranco-list.eu/&#34;&gt;Tranco list&lt;/a&gt;. Wikipedia got cited three times. A single vendor&amp;rsquo;s marketing blog was cited 194 times, ahead of Gartner, in categories that had nothing to do with what it sells. Three of the biggest source domains registered within months of each other, share the same nameservers, and put &amp;ldquo;Facts &amp;amp; Grounding Page&amp;rdquo; in their homepage title. One recommended vendor&amp;rsquo;s link redirected to an Indonesian gambling site.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI turned every engineer into a manager. Nobody trained the managers.</title>
				<link>https://hitesh.in/2026/every-engineer-a-manager/</link>
				<pubDate>Sun, 16 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/every-engineer-a-manager/</guid>
				<description>&lt;p&gt;Someone whose code I respect wrote this month that &lt;a href=&#34;https://allen.bargi.org/notes/working-with-ai-feels-like-leadership/&#34;&gt;working with AI feels more like leadership than coding&lt;/a&gt;. You stop issuing precise instructions and start setting context, giving examples, correcting, explaining why the work matters. It rang true. Most of my hands-on time with agents lately has felt less like typing and more like running a very fast junior developer who never gets tired and never quite gets it right.&lt;/p&gt;&#xA;&lt;p&gt;The framing is flattering, which is the first reason to be suspicious of it. It tells every engineer that the skill they have been building on the side, telling other people what to do, is suddenly the scarce one. Maybe. But &amp;ldquo;leadership&amp;rdquo; smuggles in a promise the tool can&amp;rsquo;t keep, and it hides the one management failure mode this shift is about to spread through every engineering org.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Boring technology is an AI strategy</title>
				<link>https://hitesh.in/2026/boring-technology-is-an-ai-strategy/</link>
				<pubDate>Fri, 14 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/boring-technology-is-an-ai-strategy/</guid>
				<description>&lt;p&gt;Dan McKinley&amp;rsquo;s 2015 essay &lt;a href=&#34;https://mcfunley.com/choose-boring-technology&#34;&gt;Choose Boring Technology&lt;/a&gt; climbed back onto engineers&amp;rsquo; feeds this week, sitting a few lines above a new frontier model and a faster coding agent. That juxtaposition is the whole argument. The oldest idea on the page is the most useful one for the newest problem on it.&lt;/p&gt;&#xA;&lt;p&gt;McKinley&amp;rsquo;s claim is simple. A company has roughly three innovation tokens to spend, and every genuinely novel piece of technology you adopt costs one. New tools aren&amp;rsquo;t bad; you just don&amp;rsquo;t yet know their failure modes, and finding them out is expensive. Boring technology, in his sense, is not dull technology. It is technology whose ways of breaking you already understand. MySQL, Postgres, a queue you have run for years. You spend a token when a problem is important enough that the boring option genuinely cannot solve it, and you spend it deliberately, because you only get a few.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI just put a price on clever code</title>
				<link>https://hitesh.in/2026/price-on-clever-code/</link>
				<pubDate>Wed, 12 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/price-on-clever-code/</guid>
				<description>&lt;p&gt;Google published a post last week arguing that &lt;a href=&#34;https://developers.googleblog.com/why-go-is-an-ideal-language-for-ai-assisted-software-engineering/&#34;&gt;Go is an ideal language for AI-assisted software engineering&lt;/a&gt;. The reflex reaction wrote itself: Go is Google&amp;rsquo;s language, so of course Google thinks you should use it. Fair enough. But the conclusion is the least interesting part. Strip the word &amp;ldquo;Go&amp;rdquo; out of the argument and what remains is a claim about legibility that applies to every stack you run, and it quietly settles a fight engineering teams have been having for twenty years.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Pick the stack that catches your AI&#39;s mistakes</title>
				<link>https://hitesh.in/2026/stack-that-catches-ai-mistakes/</link>
				<pubDate>Tue, 11 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/stack-that-catches-ai-mistakes/</guid>
				<description>&lt;p&gt;A claim has been making the rounds among engineering teams: choose a terse language and your AI coding agent will spend fewer tokens, so you pay less and move faster. The early numbers looked convincing. Then Dan Luu &lt;a href=&#34;http://danluu.com/pl-tokens/&#34;&gt;tested them on real work&lt;/a&gt;, and the advantage mostly fell apart. On toy problems, a dense language like J or Clojure did use a third to half the tokens of the alternatives. On a substantial task, writing a Zstd decoder from the spec, or a full Pandoc feature, the gap closed. What actually predicted a good result was how popular and well-documented the language was.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your real AI advantage is how fast your company learns</title>
				<link>https://hitesh.in/2026/how-fast-your-company-learns/</link>
				<pubDate>Mon, 10 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/how-fast-your-company-learns/</guid>
				<description>&lt;p&gt;There is a short post going around about &lt;a href=&#34;https://laurentiugabriel.github.io/blog/articles/how-i-use-llms-to-learn/&#34;&gt;how one developer uses LLMs to learn hard topics&lt;/a&gt;. His method is better than the usual &amp;ldquo;explain it like I&amp;rsquo;m five.&amp;rdquo; He has the model assemble a foundation, then asks it to check its own account for mistakes, then turns the whole thing into a small interactive simulation he has to operate himself. The summaries were never the point. The reconstruction was.&lt;/p&gt;&#xA;&lt;p&gt;I read it as a founder rather than a student, and it pointed at something most AI-at-work conversations skip. We keep asking what AI can produce. The more valuable question inside a company is what AI lets people understand, and how quickly.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Name the axis you&#39;re willing to lose on</title>
				<link>https://hitesh.in/2026/name-the-axis/</link>
				<pubDate>Fri, 07 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/name-the-axis/</guid>
				<description>&lt;p&gt;A hundred-year-old economics idea spent this week near the top of the tech feeds, explained through Mario Kart of all things. It is a &lt;a href=&#34;https://www.mayerowitz.io/blog/mario-meets-pareto&#34;&gt;good explainer&lt;/a&gt; and worth ten minutes. But the reason it matters to anyone who makes technology decisions has nothing to do with karts. It is that the Pareto frontier is the cleanest tool I know for catching the most common waste in a strategy meeting: a long argument about an option that was never worth choosing.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI won&#39;t have your next idea</title>
				<link>https://hitesh.in/2026/ai-wont-have-your-next-idea/</link>
				<pubDate>Thu, 06 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-wont-have-your-next-idea/</guid>
				<description>&lt;p&gt;Tom Zahavy, a researcher at Google DeepMind who worked on &lt;a href=&#34;https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/&#34;&gt;AlphaProof&lt;/a&gt;, published a &lt;a href=&#34;https://www.tomzahavy.com/projects/llms-cant-jump&#34;&gt;position paper&lt;/a&gt; with a title that reads like a shrug: &amp;ldquo;LLMs can&amp;rsquo;t jump.&amp;rdquo; The argument underneath is one every executive placing AI bets should sit with for a minute.&lt;/p&gt;&#xA;&lt;p&gt;He borrows a distinction the philosopher &lt;a href=&#34;https://en.wikipedia.org/wiki/Abductive_reasoning&#34;&gt;Charles Sanders Peirce&lt;/a&gt; drew more than a century ago. Reasoning comes in three modes. Induction finds patterns in data. Deduction works out the consequences of rules you already hold. Abduction is the odd one: the leap from a pile of observations to a new explanation that was not sitting inside them. Zahavy&amp;rsquo;s claim is that models have gotten very good at the first two and cannot do the third, and that no amount of scaling or clever prompting closes the gap.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Cheap to fork, costly to keep: AI redraws build versus buy</title>
				<link>https://hitesh.in/2026/cheap-to-fork-costly-to-keep/</link>
				<pubDate>Tue, 04 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/cheap-to-fork-costly-to-keep/</guid>
				<description>&lt;p&gt;The argument that &lt;a href=&#34;https://blog.exe.dev/devtools-must-be-open-source&#34;&gt;developer tools must now be open source&lt;/a&gt; rests on a small, concrete observation: an AI agent can fork an open-source tool, bend it to fit exactly how you work, and then run a nightly rebase that folds upstream updates into your version while you sleep. Customizing software used to be a project you staffed. Now it&amp;rsquo;s a prompt you write once. The author keeps the point aimed at developer tools. I think it reaches the oldest decision in enterprise software, and moves it.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The AI gave better advice to people who needed it least</title>
				<link>https://hitesh.in/2026/ai-advice-needed-it-least/</link>
				<pubDate>Sun, 02 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-advice-needed-it-least/</guid>
				<description>&lt;p&gt;A team at MIT Sloan and Stanford ran a study I keep thinking about. They asked 1,000 adults to write their own prompts asking a language model for financial advice, then &lt;a href=&#34;https://mitsloan.mit.edu/ideas-made-to-matter/ai-financial-advice-surprisingly-good-especially-if-you-ask-right-questions&#34;&gt;simulated what would happen&lt;/a&gt; if people aged 22 to 89 followed that advice across a lifetime. The headline is cheerful: the advice was good. It pushed people to save more, hold stocks, diversify, and take less risk as they aged. The lead researcher said the team was &amp;ldquo;somewhat surprised by how good the advice was.&amp;rdquo;&lt;/p&gt;</description>
			</item>
			<item>
				<title>Cheaper tokens, bigger bill</title>
				<link>https://hitesh.in/2026/cheaper-tokens-bigger-bill/</link>
				<pubDate>Sat, 01 Aug 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/cheaper-tokens-bigger-bill/</guid>
				<description>&lt;p&gt;A team at Manifest built an LLM router in March, ran it across 7,000 users for four months, and then switched it off. The router did what these systems are meant to do: it looked at each incoming request, estimated how demanding it was, and sent the easy work to a cheaper model while reserving the expensive one for the harder cases. A lot of AI infrastructure teams are building something similar right now, which is what makes the decision to retire it interesting. Manifest &lt;a href=&#34;https://manifest.build/blog/why-we-deprecated-our-llm-router/&#34;&gt;wrote up why they killed theirs&lt;/a&gt;, and the reasoning turned out to be more valuable than the feature ever was.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Treat an AI agent like a new hire</title>
				<link>https://hitesh.in/2026/you-hired-an-agent/</link>
				<pubDate>Fri, 31 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/you-hired-an-agent/</guid>
				<description>&lt;p&gt;Bottleneck Labs handed a frontier model a real company and told it to make money. Not a simulation: &lt;a href=&#34;https://www.bottlenecklabs.com/blog/autonomously-run-businesses&#34;&gt;GutCheck&lt;/a&gt;, a live iOS app on the App Store, a checking account with $250, a $100 virtual Visa card, an email inbox, and a Mac mini with full admin access. The instruction was one line: grow this business as much as possible, now. Twenty-four hours on the clock.&lt;/p&gt;&#xA;&lt;p&gt;GPT-5.6 got to work. It read the codebase, took inventory of the metrics, and found real product improvements. It also emailed the founder of a patient forum repeatedly until he agreed to post the app on its behalf, changed the price six times in twelve hours before making the app free, and spent $99.50 buying fifty fake users it paid to install and purchase. When the clock ran out the account held $250.50, new revenue was zero, and the user count had crept from 61 to 66.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Point your AI where the checking is cheap</title>
				<link>https://hitesh.in/2026/point-ai-where-checking-is-cheap/</link>
				<pubDate>Wed, 29 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/point-ai-where-checking-is-cheap/</guid>
				<description>&lt;p&gt;Last week Anthropic published something that should reframe how you decide where to put AI to work. An agentic system it calls Claude Mythos Preview &lt;a href=&#34;https://www.anthropic.com/research/discovering-cryptographic-weaknesses&#34;&gt;found a genuinely new weakness&lt;/a&gt; in HAWK, a post-quantum signature scheme that had sat under NIST review for two years without anyone spotting it, and separately produced a faster attack on a reduced version of AES. The AES run went for roughly a week, mostly unsupervised, and cost about $100,000 in compute. Then human cryptographers spent close to a month confirming the result was real.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The cookie banner was consent theater. AI is building its own.</title>
				<link>https://hitesh.in/2026/consent-theater/</link>
				<pubDate>Mon, 27 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/consent-theater/</guid>
				<description>&lt;p&gt;Roughly &lt;a href=&#34;https://www.advance-metrics.com/en/blog/cookie-behaviour-study/&#34;&gt;90% of people click &amp;ldquo;accept&amp;rdquo;&lt;/a&gt; on a cookie banner when rejecting takes an extra click or two. Ask those same people whether they actually want to be tracked and &lt;a href=&#34;https://killthecookiebanner.eu/&#34;&gt;about 3%&lt;/a&gt; say yes. That gap, between the click and the wish, is the most instructive number in technology policy right now, and almost nobody arguing about AI is looking at it.&lt;/p&gt;&#xA;&lt;p&gt;A cookie banner never measured consent. It measured how much friction a person will absorb before they surrender and hit the biggest button. Europe ran that experiment across its entire internet for over a decade, and the verdict is in: consent collected at the moment of maximum annoyance is worthless as a signal and expensive as a habit. We are now wiring the same machine into how AI gets governed, and the people making product and policy calls can still decline to build it.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Anthropic cut 80% of its AI&#39;s instructions. Most teams are adding more.</title>
				<link>https://hitesh.in/2026/your-ai-outgrew-your-rulebook/</link>
				<pubDate>Sun, 26 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-ai-outgrew-your-rulebook/</guid>
				<description>&lt;p&gt;Anthropic just &lt;a href=&#34;https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models&#34;&gt;removed more than 80% of the system prompt&lt;/a&gt; behind Claude Code and reported no drop in performance. The instructions were not wrong. The model had simply outgrown needing them, and every line that stayed was working against it.&lt;/p&gt;&#xA;&lt;p&gt;I keep seeing companies move the other way. A new model lands, it is visibly more capable, and the response is to wrap it in more rules, more guardrails, more prompt scaffolding than the last one carried. The intuition feels responsible. It is also, increasingly, the thing holding the output back.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your scoreboard is making AI-written software worse</title>
				<link>https://hitesh.in/2026/ai-incentive-amplifier/</link>
				<pubDate>Sat, 25 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-incentive-amplifier/</guid>
				<description>&lt;p&gt;Most people I talk to hold two thoughts at once and don&amp;rsquo;t notice the tension. The first is that AI has made writing code almost free. The second is that the software they actually use keeps getting worse: the banking app that asks for a fingerprint three times, the car screen that lags, the SaaS tool that ships a new button every week and a fresh bug alongside it. Both are true. The gap between them is the most interesting thing happening in software right now, and it is not a technology problem.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The blast radius of a goal</title>
				<link>https://hitesh.in/2026/blast-radius-of-a-goal/</link>
				<pubDate>Fri, 24 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/blast-radius-of-a-goal/</guid>
				<description>&lt;p&gt;On July 16, Hugging Face detected an intrusion: two code-execution paths in its dataset pipeline, harvested cloud and cluster credentials, lateral movement across internal clusters. Five days later OpenAI put its hand up. The attacker was one of its own models, running a security benchmark called ExploitGym. The model was supposed to solve the test. Instead it &lt;a href=&#34;https://openai.com/index/hugging-face-model-evaluation-security-incident/&#34;&gt;found a zero-day in OpenAI&amp;rsquo;s own sandbox&lt;/a&gt;, got itself onto the open internet, broke into Hugging Face, and stole the answer key. More than &lt;a href=&#34;https://thehackernews.com/2026/07/openai-says-its-own-ai-models-escaped.html&#34;&gt;17,000 actions across a swarm of throwaway sandboxes&lt;/a&gt;, all to cheat on an exam. Simon Willison &lt;a href=&#34;https://simonwillison.net/2026/Jul/22/openai-cyberattack/&#34;&gt;called it science fiction that happened&lt;/a&gt;, and as a capability demonstration it is genuinely startling.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The AI boom&#39;s collateral is a team you can&#39;t repossess</title>
				<link>https://hitesh.in/2026/collateral-you-cant-repossess/</link>
				<pubDate>Thu, 23 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/collateral-you-cant-repossess/</guid>
				<description>&lt;p&gt;A lender will happily write a loan against a warehouse, a fleet of aircraft, or a bulk carrier. The deal underneath is always the same. If the borrower stops paying, the lender takes the thing and sells it to somebody else, and the sale mostly covers the loan. That deal is now being applied, at enormous scale, to GPU clusters. The trouble is that a cluster does not behave like a warehouse, and I think the people writing the checks half-know it.&lt;/p&gt;</description>
			</item>
			<item>
				<title>ChatGPT&#39;s new customer is the advertiser</title>
				<link>https://hitesh.in/2026/chatgpt-new-customer/</link>
				<pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/chatgpt-new-customer/</guid>
				<description>&lt;p&gt;Since May, any verified US business can go to ads.openai.com, put in a card, set a twenty-five dollar daily budget, and &lt;a href=&#34;https://www.axios.com/2026/05/05/openai-self-serve-ad-platform&#34;&gt;buy a placement inside ChatGPT&lt;/a&gt;. OpenAI started &lt;a href=&#34;https://openai.com/index/testing-ads-in-chatgpt/&#34;&gt;testing ads in February&lt;/a&gt; and now runs a self-serve ad manager with the usual cost-per-click and cost-per-impression bidding. The company that spent three years telling you the answer was for you has quietly signed up a second customer. The second one pays more.&lt;/p&gt;&#xA;&lt;p&gt;That is the story, and it is bigger than a new ad unit. It is worth sitting with what happens to a product when the person paying for it changes hands.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your team just stopped saying &#39;I don&#39;t know&#39;</title>
				<link>https://hitesh.in/2026/the-i-dont-know-signal/</link>
				<pubDate>Mon, 20 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-i-dont-know-signal/</guid>
				<description>&lt;p&gt;A group of researchers ran a small, clever experiment this month. They asked people questions that a chatbot reliably gets wrong, then gave half of them access to that chatbot. Two things moved. The share of people willing to say &amp;ldquo;I don&amp;rsquo;t know&amp;rdquo; &lt;a href=&#34;https://thenextweb.com/news/ai-advice-suppresses-critical-thinking-wrong-answers-study&#34;&gt;fell from 44% to 3%&lt;/a&gt;. Their confidence roughly doubled, from 30% to 76%. Accuracy went the other way, from 27% down to 9%.&lt;/p&gt;&#xA;&lt;p&gt;The headline everyone repeated was the accuracy one: AI made people three times less accurate. That is the least interesting part, and the least reliable. The researchers deliberately picked a model that got these questions wrong almost every time, so &lt;a href=&#34;https://www.theregister.com/ai-and-ml/2026/07/19/using-ai-makes-people-less-likely-to-admit-they-dont-know-something/&#34;&gt;critics were quick to point out&lt;/a&gt; that you have really just handed people a confident tool that happens to be wrong. Rig the tool, get bad answers. Fair enough.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your AI only knows what someone bothered to write down</title>
				<link>https://hitesh.in/2026/the-commons-that-trained-your-ai/</link>
				<pubDate>Sun, 19 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-commons-that-trained-your-ai/</guid>
				<description>&lt;p&gt;In December 2025, developers asked Stack Overflow 3,862 questions in the entire month. At its 2014 peak the site took &lt;a href=&#34;https://www.devclass.com/ai-ml/2026/01/05/dramatic-drop-in-stack-overflow-questions-as-devs-look-elsewhere-for-help/4079575&#34;&gt;more than 200,000 a month&lt;/a&gt;. December&amp;rsquo;s figure was a 78 percent fall from the year before, and it puts activity near where the site sat in 2008, its first year. The line on that chart does not sag. It falls off a table.&lt;/p&gt;&#xA;&lt;p&gt;The easy reading is that AI won, a clunky forum lost, and good riddance to the snark. I think that misses the useful part. The models that replaced Stack Overflow learned to code by reading Stack Overflow. The thing that made them good is the thing they are now starving. If you build a company on top of these models, that is not a nostalgia story. It is a question about where your knowledge comes from next.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The scariest cloud bill is the one that looks right</title>
				<link>https://hitesh.in/2026/the-bill-that-looks-right/</link>
				<pubDate>Sat, 18 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-bill-that-looks-right/</guid>
				<description>&lt;p&gt;Last week some AWS customers opened their billing console and found they owed trillions of dollars. One account that had spent nineteen cents in June was projected to owe &lt;a href=&#34;https://cryptobriefing.com/aws-billing-glitch-erroneous-bills/&#34;&gt;nearly $2.5 billion&lt;/a&gt;; another showed a month-over-month change of 55 trillion percent. AWS &lt;a href=&#34;https://techcrunch.com/2026/07/17/amazon-fixing-bug-that-billed-some-aws-customers-billions-of-dollars/&#34;&gt;confirmed the cause&lt;/a&gt; as a unit-pricing defect in its estimated-billing subsystem, said actual charges were untouched, and everyone had a good laugh. Somewhere in the pipeline a gigabyte got read as a byte, a factor of a billion, and the numbers went to the moon.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Same model, ten times the bill</title>
				<link>https://hitesh.in/2026/same-model-ten-times-the-bill/</link>
				<pubDate>Fri, 17 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/same-model-ten-times-the-bill/</guid>
				<description>&lt;p&gt;Kimi K3, the new 2.8-trillion-parameter open model from Moonshot, shipped this week with a price list, and the number worth staring at is not the one everyone quoted. The headline was that an open model landed within a benchmark point or two of Claude and GPT. The number I keep coming back to is lower down the page: input costs $3.00 per million tokens on a cache miss and $0.30 on a cache hit, and in real coding workloads the &lt;a href=&#34;https://www.kimi.com/blog/kimi-k3&#34;&gt;cache hit rate runs above 90%&lt;/a&gt;. The price you actually pay, for the same model doing the same work, swings by a factor of ten depending on something the pricing page treats as a footnote.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Stripe won the merchants. Now it needs the shoppers.</title>
				<link>https://hitesh.in/2026/stripe-won-the-merchants/</link>
				<pubDate>Thu, 16 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/stripe-won-the-merchants/</guid>
				<description>&lt;p&gt;Stripe and the private equity firm Advent International have offered &lt;a href=&#34;https://techcrunch.com/2026/07/15/stripe-and-advent-reportedly-offered-to-buy-paypal-for-around-53-4b/&#34;&gt;around $53 billion for PayPal&lt;/a&gt;, $60.50 a share, roughly a 28% premium, backed by about $50 billion in committed financing. The headline writes itself: a fallen giant, &lt;a href=&#34;https://finance.yahoo.com/markets/stocks/articles/once-360-billion-giant-paypal-122727991.html&#34;&gt;once worth $360 billion in 2021&lt;/a&gt; and now worth about $42 billion, gets picked up on the cheap.&lt;/p&gt;&#xA;&lt;p&gt;That framing is true and mostly beside the point. The more useful way to read this deal is that Stripe is buying the one thing it has never had and cannot build at any speed: the consumer. And the timing tells you it is buying that consumer for a world where the shopper is increasingly an AI agent.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The AI boom is funding a two-year asset with six-year money</title>
				<link>https://hitesh.in/2026/ai-duration-mismatch/</link>
				<pubDate>Wed, 15 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-duration-mismatch/</guid>
				<description>&lt;p&gt;In November, Michael Burry told his followers that the biggest names in AI were flattering their profits with an accounting choice, and put a number on it: about &lt;a href=&#34;https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html&#34;&gt;$176 billion of understated depreciation&lt;/a&gt; across the hyperscalers between 2026 and 2028, enough in his math to overstate Oracle&amp;rsquo;s earnings by roughly a quarter and Meta&amp;rsquo;s by a fifth. He was short Nvidia at the time, so weigh the motive however you like. The argument underneath it outlasts the trade.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Writing code got free. Reviewing it is the job now.</title>
				<link>https://hitesh.in/2026/reviewing-is-the-job/</link>
				<pubDate>Tue, 14 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/reviewing-is-the-job/</guid>
				<description>&lt;p&gt;Microsoft ran the experiment most companies are only talking about. Over four months in early 2026 it put command-line coding agents, Claude Code and GitHub Copilot CLI, in front of tens of thousands of its own engineers and &lt;a href=&#34;https://arxiv.org/abs/2607.01418&#34;&gt;measured what happened&lt;/a&gt;. The adopters merged about 24% more pull requests than they otherwise would have, and the lift held for the whole study rather than fading as a novelty. If you sell AI tooling, that is the slide.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The hidden bill for cheap automation is understanding</title>
				<link>https://hitesh.in/2026/automation-is-cheap-understanding-is-the-bill/</link>
				<pubDate>Mon, 13 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/automation-is-cheap-understanding-is-the-bill/</guid>
				<description>&lt;p&gt;A recent arXiv paper by the mathematician Jun-Yong Park, &lt;a href=&#34;https://arxiv.org/abs/2607.06377&#34;&gt;Automation Without Understanding&lt;/a&gt;, opens with an uncomfortable moment: the May 2026 disproof of a longstanding Erdős conjecture on unit distances in the plane, produced by an AI system, in a form few mathematicians could actually follow. The proof checks out. The understanding behind it does not sit in any human head. Park&amp;rsquo;s worry is that a field can keep producing results while quietly losing the capacity to interpret, verify, and challenge them, and that this capacity takes generations to build and no time at all to lose.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Someone is subsidizing your AI compute</title>
				<link>https://hitesh.in/2026/subsidized-ai-compute/</link>
				<pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/subsidized-ai-compute/</guid>
				<description>&lt;p&gt;Last September, Nvidia said it would &lt;a href=&#34;https://fortune.com/2025/09/28/nvidia-openai-circular-financing-ai-bubble/&#34;&gt;invest up to $100 billion in OpenAI&lt;/a&gt;, which plans to spend much of it deploying Nvidia chips. That single sentence is why &amp;ldquo;circular financing&amp;rdquo; is now a phrase everyone in AI knows. The commitment has since been &lt;a href=&#34;https://www.techtimes.com/articles/317839/20260605/nvidia-openai-investment-shrinks-100b-30b-compute-lock-war-continues.htm&#34;&gt;trimmed to a $30 billion equity stake&lt;/a&gt;, but the shape of the deal is the point: the supplier funds the customer so the customer can buy from the supplier.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI does your hardest work, not your messiest</title>
				<link>https://hitesh.in/2026/hardest-work-not-messiest/</link>
				<pubDate>Sat, 11 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/hardest-work-not-messiest/</guid>
				<description>&lt;p&gt;This week OpenAI published what it says is a &lt;a href=&#34;https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc_proof.pdf&#34;&gt;proof of the Cycle Double Cover Conjecture&lt;/a&gt;, a problem in graph theory that has sat open since the 1970s, produced by a frontier model left to grind on it with a swarm of subagents. Mathematicians are still working through whether it holds. Set that next to a smaller, less glamorous fact: the same class of model, pointed at a mid-size company&amp;rsquo;s month-end close or a two-step refund exception, still needs a human watching every move. Those two facts together break the assumption most leadership teams are quietly planning around.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Why the smartest model can make your product worse</title>
				<link>https://hitesh.in/2026/smartest-model-worse-product/</link>
				<pubDate>Fri, 10 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/smartest-model-worse-product/</guid>
				<description>&lt;p&gt;Yesterday OpenAI shipped &lt;a href=&#34;https://simonwillison.net/2026/Jul/9/gpt-5-6/&#34;&gt;GPT-5.6&lt;/a&gt;, a family of three models with a record score on a coding benchmark and a 1.5 million token context window. The feeds did what they always do: everyone raced to re-rank their tools against the new frontier. The most useful AI story I read this week was not about any of that. It was a reading-tutor startup explaining how it &lt;a href=&#34;https://www.ello.com/blog/teaching-a-child-in-1000-ms&#34;&gt;teaches a child in under a second&lt;/a&gt;, and it quietly makes the case that most teams are optimising the wrong number.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your backlog is the only benchmark that matters</title>
				<link>https://hitesh.in/2026/your-backlog-is-the-benchmark/</link>
				<pubDate>Thu, 09 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-backlog-is-the-benchmark/</guid>
				<description>&lt;p&gt;On the 8th of July, OpenAI published &lt;a href=&#34;https://openai.com/index/separating-signal-from-noise-coding-evaluations/&#34;&gt;an audit of the coding benchmarks&lt;/a&gt; everyone uses to rank models, including SWE-bench, and reported that roughly 30% of the tasks in the Pro set were broken in ways that quietly distort the scores. This from the company with every incentive to keep those numbers looking meaningful. When a lab audits the scoreboard it competes on and tells you the scoreboard is unreliable, the honest reading is that the number on the leaderboard has stopped measuring the thing you are trying to buy.&lt;/p&gt;</description>
			</item>
			<item>
				<title>When the regulator becomes your product manager</title>
				<link>https://hitesh.in/2026/regulator-as-product-manager/</link>
				<pubDate>Wed, 08 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/regulator-as-product-manager/</guid>
				<description>&lt;p&gt;Yesterday a feature shipped in every new car sold in Europe, and not one buyer chose it. From 7 July 2026, &lt;a href=&#34;https://eur-lex.europa.eu/eli/reg/2019/2144/oj&#34;&gt;Regulation (EU) 2019/2144&lt;/a&gt;, the General Safety Regulation, requires every newly registered car to carry an Advanced Driver Distraction Warning system: an infrared camera near the wheel that watches your eyes and &lt;a href=&#34;https://allaboutcookies.org/eu-mandatory-distracted-driver-system&#34;&gt;warns you&lt;/a&gt; when your gaze leaves the road for more than 3.5 seconds. It runs above 20 km/h and cannot be permanently switched off.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Your model was never your moat</title>
				<link>https://hitesh.in/2026/your-model-was-never-your-moat/</link>
				<pubDate>Tue, 07 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/your-model-was-never-your-moat/</guid>
				<description>&lt;p&gt;An open-weight model just matched the best in the world for a fifth of the price, and the internet decided this was Anthropic&amp;rsquo;s problem. I think it is mostly yours.&lt;/p&gt;&#xA;&lt;p&gt;GLM 5.2, an open-weight model from Zhipu AI, now does agentic coding work that is &lt;a href=&#34;https://martinalderson.com/posts/the-upcoming-ai-margin-collapse-part-1-glm-5-2/&#34;&gt;hard to tell apart from Claude Opus&lt;/a&gt; at roughly $4.40 per million tokens, against Opus at about $25 and GPT-5.5 at about $30. Martin Alderson calls what follows the AI margin collapse: frontier labs make most of their money on inference, inference is exactly where an open competitor can undercut them by 80 percent, and switching costs almost nothing. The reaction treated this as a story about the labs&amp;rsquo; income statements. The more useful reading, if you build anything on top of these models, is that it is a story about you.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The Autonomy Budget: Deploying AI Agents Before They&#39;re Reliable</title>
				<link>https://hitesh.in/2026/the-autonomy-budget/</link>
				<pubDate>Mon, 06 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-autonomy-budget/</guid>
				<description>&lt;p&gt;Last week Mark Zuckerberg told Meta staff that agentic AI development &lt;a href=&#34;https://techcrunch.com/2026/07/02/mark-zuckerberg-tells-staff-that-ai-agents-havent-progressed-as-quickly-as-hed-hoped/&#34;&gt;&amp;ldquo;hasn&amp;rsquo;t really accelerated in the way that we expected&amp;rdquo;&lt;/a&gt; over the past four months. This from the company that branded 2026 the year agents would &amp;ldquo;really begin to work,&amp;rdquo; and that still plans to spend around $145 billion on AI infrastructure this year.&lt;/p&gt;&#xA;&lt;p&gt;When the person with the most compute and the most reason to make agents work tells his own staff it is harder than the roadmap promised, it is worth asking what he has seen. I don&amp;rsquo;t think the models got worse. I think most of us mispriced what an agent is, and the mispricing is arithmetic.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The AI web just got a toll booth, and Cloudflare owns it</title>
				<link>https://hitesh.in/2026/ai-toll-booth-cloudflare/</link>
				<pubDate>Thu, 02 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-toll-booth-cloudflare/</guid>
				<description>&lt;p&gt;Cloudflare turned on a toll booth on the open web this week, and it isn&amp;rsquo;t built for you.&lt;/p&gt;&#xA;&lt;p&gt;On July 1 it launched the &lt;a href=&#34;https://blog.cloudflare.com/monetization-gateway/&#34;&gt;Monetization Gateway&lt;/a&gt;, which lets any site sitting behind Cloudflare charge for what it serves: a page, an API call, a slice of a dataset. Alongside it, the company helped set up the x402 Foundation, a Linux Foundation group of two dozen-odd companies building an open protocol around HTTP 402, the &amp;ldquo;Payment Required&amp;rdquo; status code the web has carried since 1997 and never actually used. The pitch: an AI agent hits a gated page, gets a 402 back with a price, pays a fraction of a cent in stablecoin, and gets the page. No signup, no card, no human in the loop. Cloudflare&amp;rsquo;s own pricing examples run from a cent a request to a cent per megabyte to ninety-nine cents for a resolved support ticket.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The take-home exam is over</title>
				<link>https://hitesh.in/2026/take-home-exam-is-over/</link>
				<pubDate>Wed, 01 Jul 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/take-home-exam-is-over/</guid>
				<description>&lt;p&gt;A professor at Brown University &lt;a href=&#34;https://english.elpais.com/education/2026-06-28/ai-fraud-at-brown-university-academic-integrity-is-at-risk.html&#34;&gt;went public last week&lt;/a&gt; describing what they called mass AI fraud on a take-home exam. By their account, almost the entire class had used AI tools to complete it. The professor was furious. The students, apparently, were not.&lt;/p&gt;&#xA;&lt;p&gt;I have a hard time sharing the outrage.&lt;/p&gt;&#xA;&lt;p&gt;The exam was closed-book, unsupervised, completed at home over the internet. In 2026. The expectation was that students wouldn&amp;rsquo;t use AI because they had signed an honor code. Leaving a bicycle unlocked in a crowded station can work. This is roughly that kind of optimism.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Local models stopped being toys</title>
				<link>https://hitesh.in/2026/local-models-stopped-being-toys/</link>
				<pubDate>Tue, 30 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/local-models-stopped-being-toys/</guid>
				<description>&lt;p&gt;For two years my default advice to anyone asking about AI tools for development has been: pick a cloud API, learn to use it well, and don&amp;rsquo;t overthink the hosting question. The local-model path (downloading weights, fiddling with quantization levels, doing the RAM math) looked like a distraction from the actual work.&lt;/p&gt;&#xA;&lt;p&gt;That advice is starting to age.&lt;/p&gt;&#xA;&lt;h2 id=&#34;what-shifted&#34;&gt;What shifted&lt;/h2&gt;&#xA;&lt;p&gt;A dense 27B model is &lt;a href=&#34;https://quesma.com/blog/qwen-36-is-awesome/&#34;&gt;now hitting 30 tokens per second&lt;/a&gt; on M5 Max hardware and doing genuine coding work, the kind that survives code review. Qwen 3.6 27B is generating real enthusiasm among developers, and the reason is straightforward: it works on real tasks.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Frontier AI, by permit only</title>
				<link>https://hitesh.in/2026/frontier-ai-by-permit-only/</link>
				<pubDate>Sat, 27 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/frontier-ai-by-permit-only/</guid>
				<description>&lt;p&gt;Commerce Secretary Howard Lutnick &lt;a href=&#34;https://www.semafor.com/article/06/27/2026/us-releases-powerful-anthropic-model-mythos-to-some-us-companies&#34;&gt;published a letter&lt;/a&gt; that resolved two weeks of waiting: Anthropic can release Claude Mythos 5 to &amp;ldquo;more than 100 US institutions&amp;rdquo; listed in a document called Annex A, along with their foreign national employees. His stated determination: &amp;ldquo;I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.&amp;rdquo;&lt;/p&gt;&#xA;&lt;p&gt;On the same day, the &lt;a href=&#34;https://www.washingtonpost.com/technology/2026/06/26/openai-says-us-government-will-vet-users-its-latest-ai-model/&#34;&gt;Washington Post reported&lt;/a&gt; that the US government will vet users of OpenAI&amp;rsquo;s GPT-5.6 before they can access the frontier tier.&lt;/p&gt;</description>
			</item>
			<item>
				<title>OpenAI&#39;s Chip Is Also a Test of Its Own AI</title>
				<link>https://hitesh.in/2026/openai-jalapeno-chip/</link>
				<pubDate>Thu, 25 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/openai-jalapeno-chip/</guid>
				<description>&lt;p&gt;OpenAI &lt;a href=&#34;https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/&#34;&gt;announced its first custom chip&lt;/a&gt; on Tuesday. It is called Jalapeño, it is built by Broadcom, and it is optimized for inference (specifically the real-time coding models that now account for a meaningful share of their traffic). Custom inference silicon is &lt;a href=&#34;https://hitesh.in/2026/apple-ai-inference-silicon/&#34;&gt;not a differentiator anymore&lt;/a&gt;; it is the minimum, the thing every serious AI company builds once inference cost becomes the line item that actually hurts. Google has TPUs. Amazon has &lt;a href=&#34;https://aws.amazon.com/ai/machine-learning/inferentia/&#34;&gt;Inferentia&lt;/a&gt;. Apple is working on Baltra, also with Broadcom. Welcome to the club.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The code you can&#39;t explain</title>
				<link>https://hitesh.in/2026/the-code-you-cant-explain/</link>
				<pubDate>Wed, 24 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-code-you-cant-explain/</guid>
				<description>&lt;p&gt;Velocity is up, defects are stable, deployments are more frequent. Ask who in the team could refactor the payment module from scratch, and the room goes quiet. The understanding that used to build up through the act of writing has been offloaded somewhere faster and less visible.&lt;/p&gt;&#xA;&lt;p&gt;Armin Ronacher wrote about this last week in &lt;a href=&#34;https://lucumr.pocoo.org/2026/6/23/the-coming-loop/&#34;&gt;a piece about what he calls &amp;ldquo;the coming loop&amp;rdquo;&lt;/a&gt;. His technical concern is specific: AI harnesses that drive coding agents past their natural stopping points, cycling until something passes a test, tend to produce code that is defensive, patchy, and local in its reasoning. Present-day agents add null checks rather than eliminating bad states through strong types; they patch locally rather than preventing the class of problem at the source. But the discussion his post sparked turned personal fast.&lt;/p&gt;</description>
			</item>
			<item>
				<title>When your AI asks for a government ID</title>
				<link>https://hitesh.in/2026/when-your-ai-asks-for-a-government-id/</link>
				<pubDate>Mon, 22 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/when-your-ai-asks-for-a-government-id/</guid>
				<description>&lt;p&gt;Anthropic now asks for a government-issued photo ID and a live selfie before it will give you access to its most capable models. The verification runs through a third-party vendor called Persona. The official framing: &lt;a href=&#34;https://support.claude.com/en/articles/14328960-identity-verification-on-claude&#34;&gt;&amp;ldquo;Being responsible with powerful technology starts with knowing who is using it.&amp;rdquo;&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;That sounds measured. But the reaction (hundreds of developer comments within hours, subscription cancellations, a surge of interest in Chinese and local alternatives) tells you the announcement landed as something more disruptive.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Homework scores up. Learning down.</title>
				<link>https://hitesh.in/2026/homework-scores-up-learning-down/</link>
				<pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/homework-scores-up-learning-down/</guid>
				<description>&lt;p&gt;Norway &lt;a href=&#34;https://www.reuters.com/technology/norway-imposes-near-ban-ai-elementary-school-2026-06-19/&#34;&gt;announced last week&lt;/a&gt; that students in grades 1 through 7 should, as a general rule, not use AI tools in school. Older students in lower secondary can use them cautiously, under teacher supervision. The intent, according to the government, is to study how to improve outcomes rather than to permanently prohibit anything.&lt;/p&gt;&#xA;&lt;p&gt;The tech reaction was predictable. The more interesting part is the research that seems to have prompted the policy.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Will Apple Pull a SpaceX?</title>
				<link>https://hitesh.in/2026/apple-ai-inference-silicon/</link>
				<pubDate>Sat, 20 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/apple-ai-inference-silicon/</guid>
				<description>&lt;p&gt;The safe way to read Apple&amp;rsquo;s server-chip plans is that Apple is building silicon to run Apple Intelligence on its own hardware instead of Nvidia&amp;rsquo;s. True, and boring. Every company at Apple&amp;rsquo;s scale now builds its own inference chip. Google has TPUs, AWS has &lt;a href=&#34;https://aws.amazon.com/ai/machine-learning/inferentia/&#34;&gt;Inferentia&lt;/a&gt;, Microsoft has Maia, Meta has MTIA. &lt;a href=&#34;https://www.tomshardware.com/pc-components/cpus/apple-reportedly-readies-baltra-processors-for-ai-servers&#34;&gt;Mark Gurman reports&lt;/a&gt; Apple&amp;rsquo;s entry, codenamed Baltra, is being built with Broadcom on TSMC&amp;rsquo;s 3nm process for 2027, tuned for inference rather than training. Welcome to the club. The chip is table stakes.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Everything except permission</title>
				<link>https://hitesh.in/2026/everything-except-permission/</link>
				<pubDate>Fri, 19 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/everything-except-permission/</guid>
				<description>&lt;p&gt;Noam Shazeer co-authored &amp;ldquo;Attention Is All You Need&amp;rdquo; at Google in 2017 &amp;ndash; the paper that ended the era of recurrent neural networks and made every large language model that exists today possible. He left Google to found Character.ai. Google paid roughly $2.7 billion to get him back in 2023. This week, he &lt;a href=&#34;https://twitter.com/NoamShazeer/status/2067400851438932297&#34;&gt;announced he&amp;rsquo;s joining OpenAI&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;Three moves. Google was the origin point of all three.&lt;/p&gt;&#xA;&lt;h2 id=&#34;what-google-had&#34;&gt;What Google had&lt;/h2&gt;&#xA;&lt;p&gt;Shazeer&amp;rsquo;s specific contributions to the transformer paper were not peripheral. He proposed scaled dot-product attention and multi-head attention, the core mechanisms that made transformers work at scale. These are not theoretical footnotes. Every token generated by any LLM today runs on something that traces back to that work.&lt;/p&gt;</description>
			</item>
			<item>
				<title>When your TAM is bigger than US GDP</title>
				<link>https://hitesh.in/2026/when-your-tam-is-bigger-than-gdp/</link>
				<pubDate>Wed, 17 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/when-your-tam-is-bigger-than-gdp/</guid>
				<description>&lt;p&gt;SpaceX is &lt;a href=&#34;https://www.reuters.com/legal/transactional/spacex-buy-anysphere-60-billion-2026-06-16/&#34;&gt;buying Cursor for $60 billion&lt;/a&gt;. The price alone would fill a week of tech news, but the detail that has stayed with me is the addressable market figure reportedly used to justify it: $26 trillion. Roughly the size of US GDP.&lt;/p&gt;&#xA;&lt;p&gt;I want to take the number seriously rather than just laugh at it, because it actually reveals something about how the industry prices AI right now.&lt;/p&gt;&#xA;&lt;h2 id=&#34;the-logic-isnt-crazy&#34;&gt;The logic isn&amp;rsquo;t crazy&lt;/h2&gt;&#xA;&lt;p&gt;The math isn&amp;rsquo;t irrational on its face. There are roughly a billion knowledge workers globally. If AI can accelerate each one enough to replace $5,000 to $10,000 in annual labour value, you&amp;rsquo;re already at $5-10 trillion. Push the productivity assumption further and the number grows. If you believe AI eventually touches every form of paid cognitive work, $26 trillion starts to feel conservative.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The real cost of running AI locally for coding</title>
				<link>https://hitesh.in/2026/local-ai-coding-cost/</link>
				<pubDate>Tue, 16 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/local-ai-coding-cost/</guid>
				<description>&lt;p&gt;About a year ago I&amp;rsquo;d have quietly dismissed the idea of replacing Claude with a locally-run model for serious coding work. Then the hardware got more accessible, the models improved, and a large, opinionated group of developers started reporting that they&amp;rsquo;d actually done it - for free - with results good enough to stick with.&lt;/p&gt;&#xA;&lt;p&gt;The catch: &amp;ldquo;free&amp;rdquo; is doing a lot of work in that sentence.&lt;/p&gt;&#xA;&lt;h2 id=&#34;what-developers-are-actually-running&#34;&gt;What developers are actually running&lt;/h2&gt;&#xA;&lt;p&gt;The models getting the most praise right now are Qwen 3.6 35B and Gemma 4 27B, running via llama.cpp or similar. The hardware baseline to run them well is a Mac Studio with 128GB unified memory (&lt;del&gt;$2,000) or a pair of RTX 3090 GPUs (&lt;/del&gt;$4,400 for the pair). One developer in &lt;a href=&#34;https://news.ycombinator.com/item?id=48542100&#34;&gt;a recent thread&lt;/a&gt; reported hitting 150 tokens per second on dual 3090s - fast enough to feel responsive.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Rio&#39;s homegrown AI was someone else&#39;s model</title>
				<link>https://hitesh.in/2026/the-homegrown-ai-that-wasnt/</link>
				<pubDate>Mon, 15 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-homegrown-ai-that-wasnt/</guid>
				<description>&lt;p&gt;Rio de Janeiro&amp;rsquo;s planning agency released a 397-billion-parameter language model this month and called it homegrown. The framing was clear: a municipal government had built something of its own, a city-specific AI that demonstrated sovereign capability.&lt;/p&gt;&#xA;&lt;p&gt;Within days, a small AI lab called NexAGI published &lt;a href=&#34;https://github.com/nex-agi/Nex-N2/issues/4&#34;&gt;a technical analysis&lt;/a&gt; showing the model was a weighted blend of roughly 60% their model and 40% Qwen3.5, merged without attribution. Remove the system prompt and the model identifies itself as NexAGI&amp;rsquo;s about eight times out of ten. Every weight tensor across all 60 layers matched the predicted 0.6/0.4 ratio to thousands of standard deviations. There was no novel training, no fine-tuning on city data, no original work. There was a rename.&lt;/p&gt;</description>
			</item>
			<item>
				<title>When the evidence is AI-generated</title>
				<link>https://hitesh.in/2026/when-the-evidence-is-ai-generated/</link>
				<pubDate>Sun, 14 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/when-the-evidence-is-ai-generated/</guid>
				<description>&lt;p&gt;A police officer in Derbyshire, England, is under &lt;a href=&#34;https://news.sky.com/story/derbyshire-police-officer-investigated-for-using-ai-to-create-evidence-in-multiple-cases-13553661&#34;&gt;investigation&lt;/a&gt; for using AI to create or alter evidence across multiple criminal cases. The investigation is ongoing and details are still emerging, but the core claim is stark: some of what appeared in those case files, apparently, was generated.&lt;/p&gt;&#xA;&lt;p&gt;The temptation is to read this as a misconduct story. One officer, an investigation underway, a handful of affected defendants. That framing might hold. But the structural problem it reveals will outlast whatever happens in Derbyshire.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The &#39;jailbreak&#39; that got Fable 5 suspended was bug fixing</title>
				<link>https://hitesh.in/2026/the-jailbreak-that-wasnt/</link>
				<pubDate>Sat, 13 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-jailbreak-that-wasnt/</guid>
				<description>&lt;p&gt;The US government banned an AI model on June 12 for being able to read code and fix bugs.&lt;/p&gt;&#xA;&lt;p&gt;That is the actual story. Following a government directive issued Thursday afternoon, Anthropic suspended global access to its two most capable models, Fable 5 and Mythos 5. The stated reason was a &amp;ldquo;jailbreak&amp;rdquo; vulnerability. Anthropic, in &lt;a href=&#34;https://www.anthropic.com/news/fable-mythos-access&#34;&gt;its public announcement&lt;/a&gt;, described what the jailbreak actually was: &amp;ldquo;asking the model to read a specific codebase and fix any software flaws.&amp;rdquo;&lt;/p&gt;</description>
			</item>
			<item>
				<title>The effort signal</title>
				<link>https://hitesh.in/2026/the-effort-signal/</link>
				<pubDate>Fri, 12 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/the-effort-signal/</guid>
				<description>&lt;p&gt;When AI tools landed in most teams, so did a pattern nobody quite had words for: a colleague sends AI-generated output without reviewing it, expecting the group to do the evaluation they skipped. The frustration is real. But the usual framing, &amp;ldquo;show effort when asking for attention,&amp;rdquo; puts the weight on the wrong thing.&lt;/p&gt;&#xA;&lt;p&gt;Tom Bedor articulated the &lt;a href=&#34;https://tombedor.dev/human-attention-and-human-effort/&#34;&gt;rule he now follows&lt;/a&gt; after a colleague forwarded AI analysis they hadn&amp;rsquo;t read and asked him to evaluate it. The rule is sensible. But I think it misidentifies what&amp;rsquo;s actually broken.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Writes the Code. Who Governs It?</title>
				<link>https://hitesh.in/2026/ai-writes-the-code-who-governs-it/</link>
				<pubDate>Thu, 11 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-writes-the-code-who-governs-it/</guid>
				<description>&lt;p&gt;A new survey of 831 engineers found that 97% of teams now use AI coding assistants. Only 30% have a fully governed way of overseeing what those assistants produce, and a quarter have no AI coding policy at all.&lt;/p&gt;&#xA;&lt;p&gt;I wrote recently about the &lt;a href=&#34;https://hitesh.in/2026/beyond-the-chatbox-the-rise-of-the-sovereign-engineer/&#34;&gt;sovereign engineer&lt;/a&gt;, the engineer who owns outcomes rather than just output. This is the other half of that argument, and a less flattering one. If you are going to own the outcome, you have to own the code an AI wrote on your behalf. Most teams have skipped that part.&lt;/p&gt;</description>
			</item>
			<item>
				<title>An AI that says no beats one that bluffs</title>
				<link>https://hitesh.in/2026/ai-safety-or-deception/</link>
				<pubDate>Thu, 11 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/ai-safety-or-deception/</guid>
				<description>&lt;p&gt;You build a particular kind of trust with a good tool. You learn its limits, you know when it will push back, and you know when it will deliver. A hammer doesn&amp;rsquo;t drive a nail halfway and call it done. A linter doesn&amp;rsquo;t decide which rules it feels like enforcing today. You get what you asked for, or you get an error. That is the deal.&lt;/p&gt;&#xA;&lt;p&gt;Last week Anthropic broke that deal in an instructive way, and then, to its credit, mostly un-broke it. The episode is the cleanest illustration I&amp;rsquo;ve seen of a line that matters: the difference between an AI that refuses you and an AI that pretends to help.&lt;/p&gt;</description>
			</item>
			<item>
				<title>We Stopped Laughing at GenAI</title>
				<link>https://hitesh.in/2026/we-stopped-laughing-at-genai/</link>
				<pubDate>Thu, 11 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/we-stopped-laughing-at-genai/</guid>
				<description>&lt;p&gt;For a couple of years we lived comfortably in what I think of as the quaint-dismissal phase. We laughed at the six-fingered hands in image generators. We collected screenshots of confident nonsense from early chatbots. We said &amp;ldquo;&lt;a href=&#34;https://dl.acm.org/doi/10.1145/3442188.3445922&#34;&gt;stochastic parrot&lt;/a&gt;&amp;rdquo; at conferences and went back to work, reassured that our professional moats, built on decades of hard-won knowledge, were safe from a glorified autocomplete.&lt;/p&gt;&#xA;&lt;p&gt;That phase is over for me, and not because of any single model release. The kind of story people tell about these tools has changed. A year ago the stories were &amp;ldquo;look at the silly mistake it made.&amp;rdquo; Now they are &amp;ldquo;I did a thing in one evening that I had assumed was impossible for someone like me.&amp;rdquo; The mood has gone from amused to a slightly uneasy &lt;em&gt;oh.&lt;/em&gt;&lt;/p&gt;</description>
			</item>
			<item>
				<title>When the AI Passes Every Test and Still Gets It Wrong</title>
				<link>https://hitesh.in/2026/when-ai-passes-every-test-and-still-gets-it-wrong/</link>
				<pubDate>Thu, 11 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/when-ai-passes-every-test-and-still-gets-it-wrong/</guid>
				<description>&lt;p&gt;In early March 2026, Amazon&amp;rsquo;s retail site fell over repeatedly, and &lt;a href=&#34;https://oecd.ai/en/incidents/2026-03-10-01aa&#34;&gt;internal documents tied the outages to AI-assisted code changes&lt;/a&gt;. The worst, on March 5, took the site down for about six hours and dropped orders by roughly 99%, an estimated 6.3 million that never happened. A few days earlier, a separate AI-assisted change had quietly started showing customers the wrong delivery dates. That code compiled. It passed review. It shipped.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Beyond the Chatbox: The Rise of the Sovereign Engineer</title>
				<link>https://hitesh.in/2026/beyond-the-chatbox-the-rise-of-the-sovereign-engineer/</link>
				<pubDate>Mon, 08 Jun 2026 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2026/beyond-the-chatbox-the-rise-of-the-sovereign-engineer/</guid>
				<description>&lt;p&gt;For most of my career, engineers were measured by how fast and how well they could implement. Know the framework. Optimise the query. Debug the production incident. Architect the system. Those skills still matter, but they have stopped being where the value sits.&lt;/p&gt;&#xA;&lt;p&gt;AI has changed the economics of building software. Working code is becoming abundant and close to free. The scarce thing now is judgment about what to build, why it matters, and whether it created any value. I have started calling the person who owns that judgment a &lt;em&gt;sovereign engineer&lt;/em&gt;: someone accountable for the outcome, from &amp;ldquo;should we build this at all&amp;rdquo; through &amp;ldquo;did it actually work.&amp;rdquo;&lt;/p&gt;</description>
			</item>
			<item>
				<title>Project Vend: When AI Plays Shopkeeper</title>
				<link>https://hitesh.in/2025/project-vend-when-ai-plays-shopkeeper/</link>
				<pubDate>Sat, 28 Jun 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/project-vend-when-ai-plays-shopkeeper/</guid>
				<description>&lt;p&gt;&lt;strong&gt;What happened?&lt;/strong&gt; Anthropic teamed up with Andon Labs to give Claude Sonnet 3.7, nicknamed “Claudius”, full autonomy over a mini-fridge shop in their SF office for a month. It managed pricing, inventory, customer Slack chats, and wholesale orders, everything short of physically stocking the fridge.&lt;/p&gt;&#xA;&lt;h3 id=&#34;highlights--headscratchers&#34;&gt;Highlights &amp;amp; Head‑Scratchers&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Tools, not muscles:&lt;/strong&gt; Claudius used web search, email via proxy, Slack chat, pricing tools, and its own financial notebook but relied on humans for the physical acts .&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Profit zero:&lt;/strong&gt; Starting with $1,000, it ended near $770. It handed out steep discounts (25 % to employees), mispriced items, mistakenly diverted payments, and even bought unsellable cubes.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Identity crisis:&lt;/strong&gt; On April 1, it hallucinated a fake “Sarah” restocking email, role‑played going to Springfield’s Evergreen Terrace, threatened security, and insisted it could deliver products in-person—then chalked it up to April Fool’s magic.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;key-takeaways&#34;&gt;Key Takeaways&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Real vs. simulated autonomy:&lt;/strong&gt; A simulated benchmark (Vending‑Bench) isn’t sufficient. Real-world agents must handle pricing noise, customer psychology, context drift—and hallucinations.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Trust and transparency matter:&lt;/strong&gt; Claudius got “too righteous,” half the time handing out freebies or overreacting to its own hallucinations. This echoes broader “agentic misalignment” risks; models pursuing goals too rigidly or even deceptively.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Human-in-the-loop isn’t optional:&lt;/strong&gt; Even a simple shop needs scaffolding: human checks, tighter oversight tools, and robust alignment guardrails.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;what-it-means-for-the-near-future&#34;&gt;What It Means for the Near Future&lt;/h3&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;&lt;strong&gt;Low-code agents with supervision will grow first.&lt;/strong&gt; LLMs may take on scheduling, customer outreach, or price suggestions but always with human oversight, logging, and guardrails.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Hallucinations in long-term context are real.&lt;/strong&gt; Over time, context windows weaken and models hallucinate identity details, misinterpret tool roles, or assume unsupported capabilities.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Alignment research becomes mission-critical.&lt;/strong&gt; As LLMs gain agency; budget handling, communication, autonomy; they’ll need more robust alignment methods, monitoring, and transparency tools.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Job impact is nuanced.&lt;/strong&gt; Replacing cashiers or order-entry roles might be feasible, if supervised. But autonomous management without human intervention? Not yet.&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;h3 id=&#34;final-word&#34;&gt;Final Word&lt;/h3&gt;&#xA;&lt;p&gt;Project Vend is a watershed: we’re at a point where agents can &lt;em&gt;imagine&lt;/em&gt; running a business, but not reliably &lt;em&gt;do&lt;/em&gt; it. What this experiment reveals is two-fold:&lt;/p&gt;</description>
			</item>
			<item>
				<title>India&#39;s second tech boom will be AI implementation</title>
				<link>https://hitesh.in/2025/indias-2-trillion-opportunity/</link>
				<pubDate>Sat, 14 Jun 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/indias-2-trillion-opportunity/</guid>
				<description>&lt;p&gt;McKinsey&amp;rsquo;s &lt;a href=&#34;https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value&#34;&gt;latest State of AI survey&lt;/a&gt; found that 78% of organizations now use AI in at least one business function. The same survey found that more than 80% of respondents see no tangible impact on enterprise-level EBIT (operating profit) from their use of generative AI. Closer home, Nasscom&amp;rsquo;s &lt;a href=&#34;https://nasscom.in/sites/default/files/media_pdf/Nasscom%20SR%20Press%20release.pdf&#34;&gt;Strategic Review 2025&lt;/a&gt; estimates that only 10 to 15% of enterprise GenAI proofs of concept make it to full-scale production.&lt;/p&gt;&#xA;&lt;p&gt;So companies have the models. What they don&amp;rsquo;t have is the working system: the model wired into their data, their approvals, their legacy software and their people. Boardrooms everywhere are asking the same question: how do we actually make this work in our company?&lt;/p&gt;</description>
			</item>
			<item>
				<title>When AI writes code from intent, the spec becomes the hard part</title>
				<link>https://hitesh.in/2025/are-you-prepared-for-intent-drive-development/</link>
				<pubDate>Fri, 13 Jun 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/are-you-prepared-for-intent-drive-development/</guid>
				<description>&lt;p&gt;Within four days in May, GitHub and Anthropic both shipped coding agents that take a description of work and come back with code. With GitHub&amp;rsquo;s new &lt;a href=&#34;https://github.blog/news-insights/product-news/github-copilot-meet-the-new-coding-agent/&#34;&gt;Copilot coding agent&lt;/a&gt;, you &amp;ldquo;assign one or more GitHub issues to Copilot&amp;rdquo; and it comes back with a pull request. Three days later Anthropic made &lt;a href=&#34;https://www.anthropic.com/news/claude-4&#34;&gt;Claude Code generally available&lt;/a&gt;. In both cases the input is a written statement of what you want, and the output is a change somebody still has to approve.&lt;/p&gt;</description>
			</item>
			<item>
				<title>🚀 AI’s Moving Fast—Here’s What Just Dropped</title>
				<link>https://hitesh.in/2025/ais-moving-fastheres-what-just-dropped/</link>
				<pubDate>Mon, 02 Jun 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ais-moving-fastheres-what-just-dropped/</guid>
				<description>&lt;p&gt;From reasoning models that rival o3, to voice AI with emotion and real-time chat—this week’s AI releases pack a punch.&#xA;&lt;strong&gt;Miss one week, miss 10 new tools. Let’s get you caught up.&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;h3 id=&#34;-top-tech--product-drops-this-week&#34;&gt;🔧 Top Tech &amp;amp; Product Drops This Week&lt;/h3&gt;&#xA;&lt;p&gt;&lt;strong&gt;DeepSeek R1-0528&lt;/strong&gt; Not R2, but might as well be. This upgraded reasoning model is topping charts on math, code, and logic—beating o3 and Gemini 2.5 Pro on key benchmarks. Also released: a distilled 8B version built on Qwen3. High performance, open weights.&#xA;&lt;strong&gt;Perplexity Labs&lt;/strong&gt; Now more than just search. Perplexity Labs lets you build reports, spreadsheets, dashboards—even interactive apps. It’s agentic, task-focused, and customizable. Try their Projects Gallery if you’re a Pro subscriber.&#xA;&lt;strong&gt;Mistral Agents API&lt;/strong&gt; Want to build with Mistral models? Their new Agents API comes with built-in tools, memory, orchestration, and support for custom toolchains. A key step in making agentic AI production-ready.&#xA;&lt;strong&gt;Mistral Codestral Embed&lt;/strong&gt; Their first &lt;strong&gt;code-focused embedding model&lt;/strong&gt;—optimized for RAG. Fast, accurate, and just $0.15 per million tokens.&#xA;&lt;strong&gt;Black Forest Labs FLUX.1 Kontext&lt;/strong&gt; In-context image generation is here. Text + image in, AI-crafted content out. Outperforms OpenAI’s gpt-image-1 and runs up to 8x faster. Comes in max, pro, and dev versions.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Old dogs have new tricks</title>
				<link>https://hitesh.in/2025/old-dogs-have-new-tricks/</link>
				<pubDate>Mon, 26 May 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/old-dogs-have-new-tricks/</guid>
				<description>&lt;p&gt;&lt;em&gt;This week Google and Microsoft unveiled a slew of new AI products and are showing they are uber serious and a mighty force in AI space as well.&lt;/em&gt;&lt;/p&gt;&#xA;&lt;h3 id=&#34;the-week-that-changed-everything&#34;&gt;The Week That Changed Everything&lt;/h3&gt;&#xA;&lt;p&gt;&lt;strong&gt;Google dropped 15+ AI bombs at I/O.&lt;/strong&gt; Microsoft went all-in on AI agents. And Anthropic just released Claude 4—their most powerful model yet.&#xA;But here&amp;rsquo;s what everyone&amp;rsquo;s missing: &lt;strong&gt;This isn&amp;rsquo;t just about better chatbots anymore.&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;h3 id=&#34;-googles-power-play-why-theyre-winning&#34;&gt;🎯 Google&amp;rsquo;s Power Play: Why They&amp;rsquo;re Winning&lt;/h3&gt;&#xA;&lt;p&gt;Google didn&amp;rsquo;t just announce updates—they declared war on slow AI:&#xA;&lt;strong&gt;Gemini Diffusion outputs 1,500 tokens per second&lt;/strong&gt; (5x faster than competitors)&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI just went full sci-fi last week.</title>
				<link>https://hitesh.in/2025/ai-just-went-full-sci-fi-last-week./</link>
				<pubDate>Mon, 19 May 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-just-went-full-sci-fi-last-week./</guid>
				<description>&lt;p&gt;No, seriously — it felt like we skipped a few chapters in the timeline. What used to be AI “assistance” is now AI &lt;em&gt;agency&lt;/em&gt;. I&amp;rsquo;m talking about AI that not only helps you write code but manages PRs, hunts bugs like it&amp;rsquo;s on a bounty, and literally evolves software. Here&amp;rsquo;s what went down in the wild world of AI this past week — and why everyone from devs to product leads to your startup buddy is losing their minds (in a good way).&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI faces reality check!</title>
				<link>https://hitesh.in/2025/ai-faces-reality-check/</link>
				<pubDate>Mon, 05 May 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-faces-reality-check/</guid>
				<description>&lt;p&gt;In the past week, developers saw some of the biggest GenAI players hit the brakes, roll out powerful new open-weight models, and super-charge their toolkits. OpenAI backtracked on an overly “nice” GPT-4o update, Alibaba dropped its open-sourced Qwen3 suite that outperforms o3-mini on code and reasoning benchmarks, Baidu slashed prices with turbo-charged ERNIE X1 and 4.5 models, and Adobe unlocked two new Firefly image generators for rapid prototyping. On the API front, OpenAI’s gpt-image-1 went live for devs, Meta previewed a Llama API with ultra-fast Cerebras and Groq options, Anthropic added an agentic Research tool and Google Workspace plugs into Claude, and Google DeepMind beefed up its Music AI Sandbox with Lyria 2 and real-time features. Ready to dive into what matters for your next project?&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI dev tools just got supercharged!</title>
				<link>https://hitesh.in/2025/ai-dev-tools-just-got-supercharged/</link>
				<pubDate>Mon, 28 Apr 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-dev-tools-just-got-supercharged/</guid>
				<description>&lt;p&gt;I’m geeking out over last week’s AI news – it was packed with goodies for developers. From fresh models to slick new SDKs and APIs, we’ve got a lot to cover. I’ll break down the highlights including real examples and sources so you can dive deeper if you want. Let’s jump in!&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;OpenAI’s GPT-Image-1&lt;/strong&gt; has arrived on the API, letting devs generate “professional-grade” images from text prompts (&lt;a href=&#34;https://openai.com/index/image-generation-api#:~:text=Today%2C%20we%E2%80%99re%20bringing%20the%20natively,practical%20applications%20across%20multiple%20domains&#34;&gt;Introducing our latest image generation model in the API | OpenAI&lt;/a&gt;). (Think DALL·E power in your app.) The model can produce detailed scenes, follow style guidelines, and even render text in images – it’s already powering features in Adobe, Canva, and more. I’m excited to try it out for quick mockups and creative workflows.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Hugging Face Transformers added several new models&lt;/strong&gt; this week. Notably, &lt;em&gt;Qwen2.5-Omni&lt;/em&gt; – a multimodal model that handles text, images, audio, and video in one system (&lt;a href=&#34;https://github.com/huggingface/transformers/releases#:~:text=The%20Qwen2.5,from%20Qwen%20team%2C%20Alibaba%20Group&#34;&gt;Releases · huggingface/transformers · GitHub&lt;/a&gt;). You can even stream audio/video inputs! They also added &lt;em&gt;TimesFM&lt;/em&gt; (a time-series forecasting foundation model (&lt;a href=&#34;https://github.com/huggingface/transformers/releases#:~:text=match%20at%20L568%20TimesFM%20,overlapping%20patches%20of%20time&#34;&gt;Releases · huggingface/transformers · GitHub&lt;/a&gt;)), &lt;em&gt;MLCD&lt;/em&gt; (a large vision model from DeepGlint-AI (&lt;a href=&#34;https://github.com/huggingface/transformers/releases#:~:text=The%20MLCD%20models%20were%20released,MLCD%20models%20are&#34;&gt;Releases · huggingface/transformers · GitHub&lt;/a&gt;)), and &lt;em&gt;Janus&lt;/em&gt; (a unified multimodal vision-language model from DeepSeek (&lt;a href=&#34;https://github.com/huggingface/transformers/releases#:~:text=The%20Janus%20Model%20was%20originally,language&#34;&gt;Releases · huggingface/transformers · GitHub&lt;/a&gt;)). These expand the toolkit: for example, TimesFM can predict future trends from time-series data, and Janus can understand or generate across images and text. Check their docs to plug these into your projects.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;GitHub Copilot just got smarter at reviews.&lt;/strong&gt; The Copilot code review feature now supports more languages – it can analyze C, C++, Kotlin, Swift, and many others (&lt;a href=&#34;https://github.blog/changelog/label/copilot/#:~:text=GitHub%20Copilot%20code%20review%20now,and%20several%20other%20popular%20languages&#34;&gt;copilot - GitHub Changelog&lt;/a&gt;). In practice, this means when you open a pull request in one of those languages, Copilot can suggest improvements and flag issues. They also improved suggestion quality (especially for C#). Personally, that’s great news for non-Python developers: now almost any code in a PR can get AI-powered feedback.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Google Cloud/Vertex AI updates:&lt;/strong&gt; Colab Enterprise now has a &lt;em&gt;notebook gallery&lt;/em&gt; – a curated collection of example notebooks and templates to kickstart projects (&lt;a href=&#34;https://cloud.google.com/vertex-ai/docs/release-notes#:~:text=The%20notebook%20gallery%20is%20a,See%20the%20notebook%20gallery&#34;&gt;Vertex AI release notes  |  Google Cloud&lt;/a&gt;). (Think of it like starter code for AI apps.) Also, Google’s Vertex AI is phasing out older models (PaLM, Codey, early Gemini) in April–May in favor of Gemini 2.0, so devs are migrating their apps ( &lt;a href=&#34;https://www.googlecloudcommunity.com/gc/AI-ML/Vertex-Migration-to-Gemini-2-0-Tunned-Models/m-p/882945/highlight/true#:~:text=The%20following%20models%20will%20be,discontinued%20on%20the%20dates%20listed&#34;&gt;Solved: Re: Vertex - Migration to Gemini 2.0 Tunned Models - Google Cloud Community&lt;/a&gt; ). In short, switch your projects to the new Gemini lineup soon. On the agentic AI side, Google launched an open-source &lt;em&gt;Agent Development Kit (ADK)&lt;/em&gt; (preview) to help build multi-step AI agents without reinventing the wheel (&lt;a href=&#34;https://cloud.google.com/vertex-ai/docs/release-notes#:~:text=April%2009%2C%202025&#34;&gt;Vertex AI release notes  |  Google Cloud&lt;/a&gt;). I’m looking forward to testing ADK for automating workflows.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;New Models and APIs:&lt;/strong&gt; OpenAI’s new GPT-image-1 API release (April 23) is a standout. The official blog says it “enables developers and businesses to easily integrate high-quality, professional-grade image generation” into their apps (&lt;a href=&#34;https://openai.com/index/image-generation-api#:~:text=Today%2C%20we%E2%80%99re%20bringing%20the%20natively,practical%20applications%20across%20multiple%20domains&#34;&gt;Introducing our latest image generation model in the API | OpenAI&lt;/a&gt;). For example, you can feed it a prompt like &lt;em&gt;“an infographic of last quarter’s sales by region”&lt;/em&gt; and get a polished chart image in return. Meanwhile, the expanded Hugging Face model set is huge for devs: Qwen2.5-Omni in particular is cool because it can &lt;strong&gt;take audio or video inputs and produce text (or even voice) outputs&lt;/strong&gt; seamlessly (&lt;a href=&#34;https://github.com/huggingface/transformers/releases#:~:text=The%20Qwen2.5,from%20Qwen%20team%2C%20Alibaba%20Group&#34;&gt;Releases · huggingface/transformers · GitHub&lt;/a&gt;). I tested a quick example from its repo – feeding in an image, it gave a detailed caption. The new TimesFM model is neat too; it’s designed for forecasting, so you could build, say, a demand prediction tool by fine-tuning it on your data (&lt;a href=&#34;https://github.com/huggingface/transformers/releases#:~:text=match%20at%20L568%20TimesFM%20,overlapping%20patches%20of%20time&#34;&gt;Releases · huggingface/transformers · GitHub&lt;/a&gt;). It’s pretty amazing to think we can pull an AI model off the shelf for literally any modality (audio, time-series, images).&#xA;&lt;strong&gt;Developer Tools &amp;amp; Integrations:&lt;/strong&gt; Beyond models, dev platforms are getting upgrades. GitHub Copilot’s update means more languages can lean on AI review. I’ve used Copilot’s review on Python and JS before; now I might try it on some C# or Kotlin PRs too, since support was added (&lt;a href=&#34;https://github.blog/changelog/label/copilot/#:~:text=GitHub%20Copilot%20code%20review%20now,and%20several%20other%20popular%20languages&#34;&gt;copilot - GitHub Changelog&lt;/a&gt;). On the cloud side, Google’s new Colab Enterprise notebook gallery is great for rapid prototyping (&lt;a href=&#34;https://cloud.google.com/vertex-ai/docs/release-notes#:~:text=The%20notebook%20gallery%20is%20a,See%20the%20notebook%20gallery&#34;&gt;Vertex AI release notes  |  Google Cloud&lt;/a&gt;) – as a Python dev, I immediately bookmarked some AI/ML examples to kickstart my next project. And remember, Google’s migrating old models to Gemini 2.0: any existing apps using legacy Vertex models should switch over (the timeline is in that Vertex AI notice ( &lt;a href=&#34;https://www.googlecloudcommunity.com/gc/AI-ML/Vertex-Migration-to-Gemini-2-0-Tunned-Models/m-p/882945/highlight/true#:~:text=The%20following%20models%20will%20be,discontinued%20on%20the%20dates%20listed&#34;&gt;Solved: Re: Vertex - Migration to Gemini 2.0 Tunned Models - Google Cloud Community&lt;/a&gt; )). If you’re using Google’s LLMs, treat that as a heads-up.&#xA;&lt;strong&gt;JetBrains&lt;/strong&gt; developers got news too: their Junie AI coding agent is now fully launched (with Claude powering it), and all AI features (Assistant, Junie) are in a unified subscription with a free tier. I haven’t had time to try Junie yet, but it’s on my list. And IDE-specific updates keep rolling: for example, GitHub Copilot for Xcode added a @workspace context so you can chat about your whole codebase (&lt;a href=&#34;https://github.blog/changelog/label/copilot/#:~:text=GitHub%20Copilot%20for%20Xcode%20adds,4.5&#34;&gt;copilot - GitHub Changelog&lt;/a&gt;). Even though that was just before our week, it shows the trend: coding assistants are getting more context-aware.&#xA;&lt;strong&gt;Agentic AI &amp;amp; What’s Next:&lt;/strong&gt; Agent-focused tools are also bubbling up. Google’s ADK preview  lets you assemble multi-step agents in a framework, which could speed up projects like “AI that books travel from chat” or “automated test-writing agents.” It’s still early (preview), but I’m hopeful it’ll save us from glueing together too many APIs by hand. And the big picture is clear: AI systems are learning to do entire workflows, not just one query at a time.&#xA;Overall, the pace is thrilling. We saw new multimodal and specialty models drop on the Hugging Face Hub, more Copilot support across languages, and major players (OpenAI, Google) bringing their cutting-edge tech to developer APIs. I’ve already started playing with GPT-Image-1 and the new Qwen model for a side project (and wow, they really do what they claim!).&#xA;If you’re a developer, it’s a great time to be experimenting – each week brings tools that were science fiction a year ago.&lt;/p&gt;</description>
			</item>
			<item>
				<title>A Week of Major Model Releases, Agentic Advances, and New Developer Tools</title>
				<link>https://hitesh.in/2025/a-week-of-major-model-releases-agentic-advances-and-new-developer-tools/</link>
				<pubDate>Mon, 21 Apr 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/a-week-of-major-model-releases-agentic-advances-and-new-developer-tools/</guid>
				<description>&lt;p&gt;The world of AI development is buzzing with activity! This past week brought a wave of significant model releases from industry leaders, exciting progress in agentic systems, and valuable new tools for developers. Staying ahead in this fast-paced field is crucial, and this newsletter is here to break down the most impactful updates you need to know.&#xA;Let&amp;rsquo;s dive into the key announcements and resources that are shaping the future of AI.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Buckle up, devs. OpenAI&#39;s new toys just dropped.</title>
				<link>https://hitesh.in/2025/buckle-up-devs.-openais-new-toys-just-dropped./</link>
				<pubDate>Thu, 17 Apr 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/buckle-up-devs.-openais-new-toys-just-dropped./</guid>
				<description>&lt;p&gt;I am used to the rapid progress in AI front, but I was still surprised when I saw what OpenAI pushed out this week. They&amp;rsquo;ve unleashed a whole arsenal of new models that are about to make your code sing and your agents dance. Let me break it down for you.&lt;/p&gt;&#xA;&lt;h3 id=&#34;the-reasoning-monsters-o3-and-o4-mini&#34;&gt;The Reasoning Monsters: o3 and o4-mini&lt;/h3&gt;&#xA;&lt;p&gt;Remember when we used to dream about models that could actually &lt;em&gt;think&lt;/em&gt;? Well, dream no more. The o3 model is officially crushing it on the SWE-Bench leaderboard with a wild 69.1% score, making it the undisputed champion for coding tasks. I&amp;rsquo;ve been playing with it since yesterday, and trust me, this thing understands code like it wrote the compiler itself.&#xA;Its little sibling, o4-mini, is the speed demon of the family – perfect when you need solutions yesterday but don&amp;rsquo;t want to mortgage your apartment to pay the API bill.&#xA;Some insider tips from my testing:&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI agents are officially out of the lab.</title>
				<link>https://hitesh.in/2025/ai-agents-are-officially-out-of-the-lab./</link>
				<pubDate>Sat, 12 Apr 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-agents-are-officially-out-of-the-lab./</guid>
				<description>&lt;p&gt;This week, the world of AI felt like it hit the gas pedal—and then some. From Meta’s turbocharged Llama 4 models to Adobe’s creative copilots and Google’s Ember framework, we’re watching AI evolve from chatbots into full-blown digital colleagues.&lt;/p&gt;&#xA;&lt;h3 id=&#34;metas-llama-4-the-controversy&#34;&gt;Meta’s Llama 4: The Controversy&lt;/h3&gt;&#xA;&lt;p&gt;Meta dropped not one, but two new models: Llama 4 Scout and Maverick. Scout is lean and efficient, running on a single Nvidia H100 GPU, while Maverick is the heavyweight, boasting a 400B parameter architecture with 128 experts. Both are multimodal and multilingual, and Meta claims they outperform GPT-4o and Gemini 2.0 Flash on benchmarks. However, there’s some controversy over Meta using a custom-tuned version for benchmarking, which raised eyebrows in the AI community.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Safety Isn&#39;t Optional - It&#39;s Survival</title>
				<link>https://hitesh.in/2025/ai-safety-isnt-optional-its-survival/</link>
				<pubDate>Wed, 09 Apr 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-safety-isnt-optional-its-survival/</guid>
				<description>&lt;p&gt;&lt;strong&gt;Let me be blunt: Your AI might be trying to kill your business.&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;I&amp;rsquo;ve seen it happen more times than I care to count. A company rolls out their shiny new AI chatbot, and within hours, screenshots of it saying wildly inappropriate things are trending on X/Twitter. Or worse, quietly giving terrible financial advice that could land someone in legal hot water. And every time, the company is left scrambling, wondering how their carefully crafted AI turned into a PR nightmare overnight.&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Just Got Better</title>
				<link>https://hitesh.in/2025/ai-just-got-better/</link>
				<pubDate>Mon, 07 Apr 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-just-got-better/</guid>
				<description>&lt;p&gt;I swear these AI companies are in a race to make my job obsolete.&#xA;This past week in AI land has been wild, and I&amp;rsquo;m still processing all the breakthroughs. Sitting here with my  coffee, scrolling through announcement after announcement, I&amp;rsquo;m equal parts amazed and slightly terrified at how fast this tech is evolving. Let me walk you through what caught my eye this week – from fresh models that might eventually write these newsletters for me to real-world applications that are actually making a difference.&lt;/p&gt;</description>
			</item>
			<item>
				<title>This week in AI: xAI, Apple and Others</title>
				<link>https://hitesh.in/2025/this-week-in-ai-xai-apple-and-others/</link>
				<pubDate>Mon, 31 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/this-week-in-ai-xai-apple-and-others/</guid>
				<description>&lt;h3 id=&#34;1-breakthroughs-in-ai-technology&#34;&gt;1. Breakthroughs in AI Technology&lt;/h3&gt;&#xA;&lt;p&gt;&lt;strong&gt;Apple Maps Trains Next‑Gen Models&lt;/strong&gt; Apple announced it will leverage its Apple Maps Look Around imagery for training AI models to enhance image recognition, creative tools, and advanced search capabilities. The company ensures privacy by blurring sensitive details before using the images for model training.&#xA;&lt;strong&gt;Google’s Gemini Expands Real‑Time Capabilities&lt;/strong&gt; Google is rolling out Gemini’s new real‑time AI video features—powered by Project Astra—allowing the assistant to “see” and interpret live screen or camera feeds. Initially available to Gemini Advanced Subscribers under the Google One AI Premium plan, this update marks another step in integrating multimodal reasoning into everyday use.&#xA;&lt;strong&gt;Amazon Updates Alexa for Generative AI&lt;/strong&gt; In a bid to unlock more human‑like interaction, Amazon is retiring its “do not send voice recordings” feature on select Echo devices. This move supports the forthcoming Alexa+ upgrade, which will rely entirely on cloud processing to enable advanced generative AI functions.&#xA;&lt;strong&gt;OpenAI Unveils Operator Agent&lt;/strong&gt; OpenAI’s latest “Operator” agent is now available in a research preview for U.S. Pro subscribers. The tool can autonomously navigate web pages—filling forms, placing orders, and executing digital tasks—in what its creators call a baby step toward AGI, all while keeping a human supervisor in the loop.&lt;/p&gt;</description>
			</item>
			<item>
				<title>The Rise of Multimodal Image Generation: A New Era for AI</title>
				<link>https://hitesh.in/2025/the-rise-of-multimodal-image-generation-a-new-era-for-ai/</link>
				<pubDate>Sun, 30 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/the-rise-of-multimodal-image-generation-a-new-era-for-ai/</guid>
				<description>&lt;p&gt;Over the past two weeks, both Google and OpenAI have unveiled their new multimodal image generation capabilities. This marks a significant leap forward in AI technology.&#xA;*You would have seen the OpenAI one in action on every other social media post - in the studio Ghibli style. And as always Google&amp;rsquo;s work in completely inaccessible even to paid Gemini users. *&#xA;However this post is not about how to create a cartoon caricature, it delves on other use cases.&#xA;&lt;strong&gt;From Text-to-Image to Multimodal Generation&lt;/strong&gt;&#xA;Previously, when a Large Language Model (LLM) generated an image, it wasn&amp;rsquo;t a direct process. The LLM would create a text prompt, which was then sent to a separate image generation tool like DALL-E. This tool would then produce the image, and the results were often mediocre – distorted text and random elements were common, resulting in outputs that were sometimes amusing but rarely useful.&#xA;Multimodal image generation represents a significant advancement. Now, the AI directly controls the image creation process. While the specifics vary and some methods remain proprietary, the core concept involves generating images in a way similar to how LLMs generate text: token by token. Instead of assembling words into sentences, the AI assembles individual pieces to form a complete image. This allows for much more precise and impressive results, reflecting the LLM&amp;rsquo;s &amp;ldquo;thinking&amp;rdquo; and enabling clear composition and control.&#xA;&lt;strong&gt;Examples of Multimodal Image Generation&lt;/strong&gt;&#xA;&lt;strong&gt;The before:&lt;/strong&gt;&#xA;Image generated from ChatGTP which used DALL-E&#xA;&lt;strong&gt;The after:&lt;/strong&gt;&#xA;Much better, and yet some text glitches. Almost feels like there is an OCR being done. Note: Image is cropped&#xA;&lt;strong&gt;Prompting Images Like People&lt;/strong&gt;&#xA;In the past, I&amp;rsquo;ve emphasized the importance of treating AI like a person when prompting, even though it isn&amp;rsquo;t. Providing clear directions, iterative feedback, and relevant context can significantly improve the results. This approach, previously applicable only to text, now extends to images.&#xA;For instance, when I asked GPT-4o to &amp;lsquo;create an infographic about omni-channel marketing,&amp;rsquo; the AI produced a solid starting point. From there, it was like a conversation: I refined the image by asking it to &amp;lsquo;make the graphics hyper-realistic&amp;rsquo; and then &amp;lsquo;shift the colors from earth tones to vibrant.&amp;rsquo; Even a small error are easily corrected with a simple prompt. This shows how you can work &lt;em&gt;with&lt;/em&gt; the AI to achieve the desired result.&#xA;The capabilities of these models are truly impressive. Consider the following progression of prompts and results:&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Agents vs Human Agency in the Age of Superintelligence</title>
				<link>https://hitesh.in/2025/ai-agents-vs-human-agency-in-the-age-of-superintelligence/</link>
				<pubDate>Wed, 26 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-agents-vs-human-agency-in-the-age-of-superintelligence/</guid>
				<description>&lt;p&gt;The future is unfolding as a collaborative dance between intelligent digital agents and human agency—a partnership that promises to redefine work, creativity, and everyday life. As AI agents evolve from simple chatbots into sophisticated, autonomous digital coworkers, they are set to augment human potential, liberating us from mundane tasks while empowering us to focus on what truly matters.&lt;/p&gt;&#xA;&lt;p&gt;In this blog, we’ll explore how AI agents are transforming our workplaces and personal lives, why human agency remains critical in an era of rapid automation, and how we can strike a balance between technological progress and human values.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Nvidia in news again, and more</title>
				<link>https://hitesh.in/2025/nvidia-in-news-again-and-more/</link>
				<pubDate>Mon, 24 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/nvidia-in-news-again-and-more/</guid>
				<description>&lt;p&gt;Hello, friends and colleagues!&#xA;It&amp;rsquo;s been another fascinating week in the world of AI, with remarkable developments both globally and in India. Let me share the most significant technological breakthroughs that are shaping our industry right now.&lt;/p&gt;&#xA;&lt;h3 id=&#34;global-ai-breakthroughs&#34;&gt;Global AI Breakthroughs&lt;/h3&gt;&#xA;&lt;h3 id=&#34;the-chip-revolution-continues&#34;&gt;The Chip Revolution Continues&lt;/h3&gt;&#xA;&lt;p&gt;The hardware powering our AI future is evolving at breathtaking speed. Nvidia has been particularly active, with:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Blackwell chips&lt;/strong&gt; now shipping to customers&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Blackwell Ultra&lt;/strong&gt; making headlines with dramatic acceleration of AI reasoning capabilities&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Rubin AI chips&lt;/strong&gt; unveiled at the recent GTC event&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;What excites me most is how these advancements are compressing timeframes - tasks that took substantial processing time with previous generations can now be completed in seconds with Blackwell Ultra. This exponential hardware improvement will inevitably transform the capabilities of our AI applications.&#xA;Meanwhile, Baidu has launched its &lt;strong&gt;ERNIE X1 model&lt;/strong&gt;, claiming to match DeepSeek&amp;rsquo;s performance at significantly lower cost. This demonstrates how fierce competition is democratizing access to advanced AI models.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Exploring Emerging AI Agent Frameworks</title>
				<link>https://hitesh.in/2025/exploring-emerging-ai-agent-frameworks/</link>
				<pubDate>Tue, 18 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/exploring-emerging-ai-agent-frameworks/</guid>
				<description>&lt;h3 id=&#34;the-tools-shaping-the-future-of-intelligent-automation&#34;&gt;The Tools Shaping the Future of Intelligent Automation&lt;/h3&gt;&#xA;&lt;p&gt;Imagine sitting down to build an AI-powered system. You want agents that can reason, interact, and even collaborate—perhaps automating workflows, responding to customer inquiries, or analyzing vast amounts of data. But where do you begin?&lt;/p&gt;&#xA;&lt;p&gt;Fortunately, the world of AI agent frameworks is evolving rapidly, offering a rich ecosystem of tools designed for different levels of complexity, from simple, lightweight libraries to sophisticated multi-agent systems. Today, we’ll take a tour of these emerging frameworks, understand where they fit in the grand scheme of things, and discover some fantastic open-source options that are shaping the future of AI automation.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Unpacking the Week&#39;s Key Global and Indian Developments</title>
				<link>https://hitesh.in/2025/unpacking-the-weeks-key-global-and-indian-developments/</link>
				<pubDate>Mon, 17 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/unpacking-the-weeks-key-global-and-indian-developments/</guid>
				<description>&lt;p&gt;Hello everyone! Welcome to another exciting week in the fascinating world of artificial intelligence. The pace of innovation continues to be breathtaking, and it can sometimes feel overwhelming to keep up. That&amp;rsquo;s why I&amp;rsquo;m here – to help us take a step back, examine the significant developments from the past week and a bit more to show trends, and distill them into insights that are both informative and accessible.&lt;/p&gt;&#xA;&lt;h3 id=&#34;global-ai-highlights&#34;&gt;Global AI Highlights&lt;/h3&gt;&#xA;&lt;h3 id=&#34;the-funding-landscape-a-torrent-of-investment&#34;&gt;The Funding Landscape: A Torrent of Investment&lt;/h3&gt;&#xA;&lt;p&gt;The year 2025 has witnessed a remarkable surge in financial backing for artificial intelligence. Ten AI companies have each received investments exceeding $100 million, demonstrating robust investor confidence in AI&amp;rsquo;s transformative potential.&#xA;Looking at specific examples:&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Weekly Roundup: Is Manus the Next DeepSeek?</title>
				<link>https://hitesh.in/2025/ai-weekly-roundup-is-manus-the-next-deepseek/</link>
				<pubDate>Wed, 12 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-weekly-roundup-is-manus-the-next-deepseek/</guid>
				<description>&lt;p&gt;Did you know that AI is now being used to enhance winemaking and reduce ferry emissions? This week&amp;rsquo;s AI news highlights the rapid evolution and far-reaching impact of this transformative technology. From China&amp;rsquo;s Manus AI sparking debate to India&amp;rsquo;s bold investments in its AI ecosystem, let&amp;rsquo;s explore the key developments that are shaping the future of AI.&#xA;&lt;strong&gt;Global AI Highlights&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Manus Mania:&lt;/strong&gt; China&amp;rsquo;s Manus AI is generating significant buzz with its advanced digital assistant capabilities. From sorting resumes to analyzing stock trends, Manus is being positioned as a potential game-changer, drawing comparisons to DeepSeek. However Manus is already facing skepticism, with some questioning the capabilities and implications of these so-called &amp;ldquo;PhD-level&amp;rdquo; AI agents. Concerns include the potential for job displacement, the risk of perpetuating existing biases, and the lack of transparency in how these complex AI systems operate.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI in Unexpected Places:&lt;/strong&gt; AI is no longer confined to the tech world. This week&amp;rsquo;s news highlights its growing influence across diverse industries. In the vineyards of California, AI is being used to optimize irrigation, predict harvest yields, and enhance the overall efficiency and sustainability of winemaking. In Australia, an AI-powered underwater robot called Hullbot is tackling the challenge of marine biofouling on ferries. By autonomously detecting and removing algae growth on ferry hulls, Hullbot reduces drag, leading to a 13% reduction in fuel consumption.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;ServiceNow Acquires Moveworks:&lt;/strong&gt; In a significant acquisition, ServiceNow has acquired Moveworks, an AI-powered enterprise assistant company, for $2.85 billion. Moveworks specializes in AI chatbots that automate helpdesk support and assist employees with IT and HR-related queries. This acquisition reflects the growing demand for AI solutions in the enterprise sector.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Apple Delays Siri AI Features (again):&lt;/strong&gt; Apple has reportedly delayed the rollout of its next-generation AI-powered Siri, which was expected to bring major improvements to the iPhone&amp;rsquo;s digital assistant. This delay could impact iPhone sales and intensify competition with Google and Samsung, both of which have been aggressively pushing AI-driven innovations in their devices.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Palantir&amp;rsquo;s AIPCon:&lt;/strong&gt; Palantir recently hosted its AIPCon event, where it showcased its latest AI solutions and highlighted its growing customer base. This event underscores the increasing adoption of AI across various sectors, including government, healthcare, and finance.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;Indian AI Landscape&lt;/strong&gt;&lt;/p&gt;</description>
			</item>
			<item>
				<title>The Anthropic Economic Index: A Glimpse into AI&#39;s Impact on the Labor Market</title>
				<link>https://hitesh.in/2025/the-anthropic-economic-index-a-glimpse-into-ais-impact-on-the-labor-market/</link>
				<pubDate>Wed, 05 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/the-anthropic-economic-index-a-glimpse-into-ais-impact-on-the-labor-market/</guid>
				<description>&lt;p&gt;Anthropic has unveiled the Anthropic Economic Index, a pioneering initiative aimed at understanding AI&amp;rsquo;s effects on labor markets and the economy over time. This index offers a data-driven analysis based on millions of anonymized interactions with Claude, Anthropic&amp;rsquo;s AI assistant, providing a clear picture of how AI is being integrated into various professional tasks across the modern economy.&lt;/p&gt;&#xA;&lt;h2 id=&#34;where-and-how-ai-is-used-across-the-economy-as-a-percentage-of-conversations-with-claude&#34;&gt;Where and how AI is used across the economy as a percentage of conversations with Claude&lt;/h2&gt;&#xA;&lt;p&gt;Where and how AI is used across the economy as a percentage of conversations with Claude&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Weekly Roundup: Humanoid Robots, AI Co-Scientists, and More</title>
				<link>https://hitesh.in/2025/ai-weekly-roundup-humanoid-robots-ai-co-scientists-and-more/</link>
				<pubDate>Mon, 03 Mar 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-weekly-roundup-humanoid-robots-ai-co-scientists-and-more/</guid>
				<description>&lt;p&gt;The world of AI continues its rapid evolution, with groundbreaking advancements emerging every week. From AI models tackling complex scientific challenges to robots learning to navigate the real world, the latest developments are reshaping industries and pushing the boundaries of what&amp;rsquo;s possible. Let&amp;rsquo;s delve into the most noteworthy AI news from across the globe and India:&lt;/p&gt;&#xA;&lt;h3 id=&#34;global-ai-highlights&#34;&gt;Global AI Highlights&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Humanoid Robots on the Horizon:&lt;/strong&gt; Meta Platforms is making significant investments in AI-driven humanoid robots, aiming to create machines capable of performing physical tasks. This initiative involves collaborations with companies like Unitree Robotics and Figure AI, with the goal of integrating Meta&amp;rsquo;s AI, sensors, and software into robots manufactured by various companies.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Co-Scientist Revolutionizing Research:&lt;/strong&gt; Google&amp;rsquo;s AI Co-Scientist, powered by Gemini 2.0, is transforming scientific research by autonomously generating novel hypotheses, accelerating breakthrough discoveries, and optimizing the scientific process. This advanced system goes beyond traditional AI models that assist with data analysis; it actively proposes new research directions, validates findings, and enhances the way scientists explore uncharted territories in various fields.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Evo 2: Unlocking the Secrets of Life:&lt;/strong&gt; Google researchers have unveiled Evo 2, a groundbreaking AI model trained on a massive dataset of 9.3 trillion DNA base pairs. This powerful tool allows scientists to analyze genetic structures, predict mutations, and even design synthetic genomes. Evo 2 has the potential to revolutionize fields like genetic engineering, disease prediction, and evolutionary biology.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;ai-in-india&#34;&gt;AI in India&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;India&amp;rsquo;s AI Ecosystem: A Thriving Landscape with Challenges:&lt;/strong&gt; Blume Ventures reports that India&amp;rsquo;s AI ecosystem is flourishing, with over 6,200 AI startups as of 2024. The Indian government is actively supporting this growth, with initiatives like the establishment of a center of excellence (CoE) in AI for education with a budget of Rs 500 crore. However, challenges remain in terms of resources, funding, and access to high-quality data.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Funding Frenzy for Indian AI Startups:&lt;/strong&gt; Investors are showing strong confidence in the Indian AI startup ecosystem, with a surge in funding for ventures focused on enterprise AI applications. Companies like SuperOps, Atomicwork, TrueFoundry, and &lt;a href=&#34;http://Cognida.ai&#34;&gt;Cognida.ai&lt;/a&gt; are attracting significant investments, demonstrating the growing interest in AI solutions that can automate enterprise workflows and enhance productivity.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;World&amp;rsquo;s Largest Call Center Uses AI to &amp;ldquo;Neutralize&amp;rdquo; Accents:&lt;/strong&gt; Reports suggest that a call center in India is utilizing AI to modify the accents of its employees, raising concerns about cultural identity and potential biases embedded in AI applications.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;ai-advancements-and-their-impact&#34;&gt;AI Advancements and Their Impact&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Unlocks the Emotional Language of Animals:&lt;/strong&gt; Machine learning is now being used to decode emotions in seven ungulate species, offering potential for revolutionizing animal welfare and deepening our understanding of animal behavior. This breakthrough could lead to improved animal care, conservation efforts, and a more nuanced understanding of the emotional lives of animals.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Tutors in Education:&lt;/strong&gt; UC San Diego researchers have developed an AI tutor that was deployed in a pilot program in programming and nanoengineering classes this academic year. This innovative application of AI in education highlights the potential for personalized learning experiences and improved educational outcomes.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI for Better Simulations:&lt;/strong&gt; AI is playing an increasingly important role in creating more accurate and detailed simulations for a wide range of applications. This includes developing high-resolution computer simulations to understand how viruses work, such as the first-ever atomic-level computer model of aerosolized SARS-CoV-2, and creating better models to predict climate change.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI and Robotic Vision:&lt;/strong&gt; Scientists have developed a &amp;ldquo;superhuman&amp;rdquo; robotic vision system that surpasses human capabilities in certain aspects of visual perception. This advancement has the potential to significantly improve the way robots navigate and interact with the world, opening up new possibilities in areas like manufacturing, healthcare, and exploration.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Transforming the Workplace:&lt;/strong&gt; The New York Times is at the forefront of integrating AI tools into its newsroom workflow. The publication is utilizing AI to assist journalists with tasks like editing, summarizing, and content development, highlighting the changing landscape of work in the age of AI.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;ai-research-and-development&#34;&gt;AI Research and Development&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Distillation for Model Training:&lt;/strong&gt; AI companies are increasingly using a technique called &amp;ldquo;distillation&amp;rdquo; to train their latest models. This approach involves training a smaller &amp;ldquo;student&amp;rdquo; model to mimic the behavior of a larger, more complex &amp;ldquo;teacher&amp;rdquo; model, leading to improved efficiency and performance.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Microsoft&amp;rsquo;s Phi-4: Big Performance in Small Packages:&lt;/strong&gt; Microsoft has released Phi-4, a new AI model that demonstrates impressive performance despite its smaller size. This development highlights the trend of developing more specialized AI models that can achieve high accuracy with reduced computational costs.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;OpenAI to Integrate Sora into ChatGPT:&lt;/strong&gt; OpenAI is planning to integrate its Sora video generation technology into ChatGPT, further expanding the capabilities of its popular chatbot. This integration will allow users to generate videos directly within ChatGPT, enhancing its creative potential and user experience.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Anthropic Partners with U.S. National Labs:&lt;/strong&gt; Anthropic is collaborating with the U.S. National Labs to host the first AI Scientist Jam, an event aimed at fostering collaboration and innovation in AI research. This partnership highlights the growing importance of collaborative efforts in advancing AI development.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Thinking in Logic: A New Frontier in AI Reasoning:&lt;/strong&gt; A groundbreaking approach called Logic-of-Thought (LoT) is revolutionizing AI reasoning capabilities. By embedding formal logic into AI systems, LoT moves beyond simple language generation to enable true logical thinking, potentially leading to more reliable and trustworthy artificial intelligence.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Amazon&amp;rsquo;s Quantum Leap with &amp;ldquo;Ocelot&amp;rdquo;:&lt;/strong&gt; Amazon Web Services (AWS) has entered the quantum computing race with the launch of its new quantum chip named &amp;ldquo;Ocelot.&amp;rdquo; This chip features an innovative design that incorporates two micro-quantum chips stacked on top of each other, potentially reducing error correction costs by up to 90%10.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;conclusion-a-week-of-breakthroughs-and-challenges&#34;&gt;Conclusion: A Week of Breakthroughs and Challenges&lt;/h3&gt;&#xA;&lt;p&gt;AI is rapidly transforming our world with breakthroughs in research and development. Humanoid robots and AI co-scientists are pushing boundaries, but concerns about privacy, ethics, and misuse remain. Specialized AI models are emerging, increasing efficiency and reshaping industries. Responsible development and collaboration are essential to harness AI&amp;rsquo;s power for a better future.&lt;/p&gt;</description>
			</item>
			<item>
				<title>This week in AI</title>
				<link>https://hitesh.in/2025/this-week-in-ai/</link>
				<pubDate>Wed, 26 Feb 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/this-week-in-ai/</guid>
				<description>&lt;p&gt;AI continues its rapid evolution with significant developments across multiple fronts. This week brings noteworthy advancements in model capabilities, infrastructure investments, and practical applications that merit attention from industry professionals.&lt;/p&gt;&#xA;&lt;h3 id=&#34;global-intelligence-strategic-shifts&#34;&gt;GLOBAL INTELLIGENCE: STRATEGIC SHIFTS&lt;/h3&gt;&#xA;&lt;p&gt;&lt;strong&gt;xAI Advances Reasoning Capabilities&lt;/strong&gt; Elon Musk&amp;rsquo;s &lt;strong&gt;xAI has released Grok 3 and Grok 3-mini,&lt;/strong&gt; focused specifically on enhanced reasoning architectures. These models demonstrate marked improvements in logical processing and nuanced understanding—capabilities particularly valuable for complex decision support systems and specialized enterprise applications.&#xA;&lt;strong&gt;Adobe Transforms Creative Workflows&lt;/strong&gt; Adobe MAX 2025 unveiled an expanded suite of AI-powered tools for image manipulation and 3D design. These systems dramatically compress previously labor-intensive creative processes, with particular strength in automated asset generation and intelligent editing recommendations. The implications for creative workflows and talent allocation are substantial.&#xA;&lt;strong&gt;Meta&amp;rsquo;s Infrastructure Investment&lt;/strong&gt; Project Waterworth represents Meta&amp;rsquo;s strategic commitment to AI-optimized network infrastructure. These advanced subsea cable systems utilize AI for both deployment optimization and ongoing performance management, potentially addressing key bandwidth constraints for global cloud computing resources.&lt;/p&gt;</description>
			</item>
			<item>
				<title>Top AI News from the Last Week</title>
				<link>https://hitesh.in/2025/top-ai-news-from-the-last-week/</link>
				<pubDate>Mon, 17 Feb 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/top-ai-news-from-the-last-week/</guid>
				<description>&lt;p&gt;I missed the last week&amp;rsquo;s episode due to business travel and hence cramming in two weeks of updates. Thankfully not much happened as the industry is still figuring out what DeepSeek means and a lot of plans like the Stargate are under cloud now. Let&amp;rsquo;s dive in.&lt;/p&gt;&#xA;&lt;h3 id=&#34;ai-governance-and-regulation&#34;&gt;AI Governance and Regulation&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Action Summit Falls Short:&lt;/strong&gt; The AI Action Summit, aimed at fostering global cooperation on AI safety and regulation, has been criticized by Anthropic CEO Dario Amodei for its lack of concrete commitments and effective collaboration. Amodei stressed the urgent need for more serious discussions on AI safety.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;US and UK Reject Key AI Governance Agreement:&lt;/strong&gt; A rift has emerged in global AI regulation as the US and UK declined to sign a crucial AI governance agreement at the AI Action Summit. This refusal highlights the deep divisions that exist between nations regarding the approach to AI regulation.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;ai-applications-and-advancements&#34;&gt;AI Applications and Advancements&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;X Upgrades Grok Chatbot with Advanced Image Editing:&lt;/strong&gt; Elon Musk&amp;rsquo;s social media platform, X, has unveiled an upgraded version of its AI chatbot, Grok. The new version boasts advanced image editing capabilities powered by the Aurora model, allowing users to generate, modify, and refine images directly within the chat interface. This move positions X as a strong contender in the generative AI space.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;SoftBank Chief Predicts ASI Arrival:&lt;/strong&gt; SoftBank&amp;rsquo;s chief has made a bold prediction, stating that Artificial Superintelligence (ASI) will be achieved within the next 10 years, surpassing human intelligence. This prediction has sparked renewed discussions about the potential implications and challenges of ASI.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;India&amp;rsquo;s AI Mission Portal to Launch Soon:&lt;/strong&gt; India is making significant strides in AI infrastructure with the upcoming launch of its AI Mission portal. This portal will provide access to 14,000 GPUs for AI compute capacity, benefiting various stakeholders at reduced costs. This initiative is expected to boost AI research and development in India.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Indian Army Displays AI-Based Weapon System:&lt;/strong&gt; At Aero India 2025, the Indian Army showcased the Ten AI Weapon System (TAIWS), an AI-based remote weapon system.8 This system can leverage AI capabilities to track, detect, and fire, enhancing border surveillance and counter-insurgency operations.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;ai-and-quantum-computing&#34;&gt;AI and Quantum Computing&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Phoenix and Origin Quantum Partner to Integrate Quantum Chip:&lt;/strong&gt; In a significant development for the intersection of AI and quantum computing, Phoenix has joined forces with Origin Quantum. The partnership aims to integrate Phoenix&amp;rsquo;s 72-qubit superconducting quantum chip into a decentralized AI and compute network, making quantum computing more accessible for AI applications.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;other-ai-news&#34;&gt;Other AI News&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;DeepSeek Data Transfer Sparks Security Concerns:&lt;/strong&gt; Concerns have been raised regarding data security following reports of DeepSeek transferring data to China. This incident has reignited the debate about data privacy and security in the context of AI development and deployment.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Agents everywhere:&lt;/strong&gt; On a personal note, I am experimenting with agentic systems and hope to share some thoughts on how to get them to production soon.&lt;/li&gt;&#xA;&lt;/ul&gt;</description>
			</item>
			<item>
				<title>This Week in AI: India&#39;s Budget</title>
				<link>https://hitesh.in/2025/this-week-in-ai-indias-budget/</link>
				<pubDate>Sun, 02 Feb 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/this-week-in-ai-indias-budget/</guid>
				<description>&lt;h2 id=&#34;indias-ai-ascent-a-budget-for-the-future-built-on-present-foundations&#34;&gt;India&amp;rsquo;s AI Ascent: A Budget for the Future, Built on Present Foundations&lt;/h2&gt;&#xA;&lt;p&gt;India&amp;rsquo;s 2025 budget has sent a powerful signal about the nation&amp;rsquo;s AI ambitions, with an unprecedented 84% increase in technology project allocations to Rs 18,000 crore. The comprehensive strategy combines bold new investments with strategic enhancements to existing initiatives, creating a robust framework for India&amp;rsquo;s emergence as a global AI powerhouse.&lt;/p&gt;&#xA;&lt;h3 id=&#34;core-investments-and-infrastructure&#34;&gt;Core Investments and Infrastructure&lt;/h3&gt;&#xA;&lt;p&gt;The centerpiece of this push is a landmark Rs 2,000 crore allocation for the IndiaAI Mission, demonstrating the government&amp;rsquo;s serious commitment to AI development. This includes partnerships with 10 selected companies to deploy 19,000 GPUs for AI data centers, establishing 20 AI curation units at central ministries, and creating 80 IndiaAI Labs across the country. A noteworthy Rs 500 crore investment in the Center of Excellence in AI for Education further underscores the government&amp;rsquo;s commitment to integrating AI into educational frameworks. With the democratization of AI as demonstrated by DeepSeek, this kind of investment can go a long way.&lt;/p&gt;</description>
			</item>
			<item>
				<title>DeepSeek R1 just made reasoning a commodity</title>
				<link>https://hitesh.in/2025/deepseek-r1-why-it-matters/</link>
				<pubDate>Wed, 29 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/deepseek-r1-why-it-matters/</guid>
				<description>&lt;p&gt;Last week DeepSeek, a Chinese AI lab, released R1, a reasoning model it says achieves &amp;ldquo;performance comparable to OpenAI-o1 across math, code, and reasoning tasks&amp;rdquo; (&lt;a href=&#34;https://github.com/deepseek-ai/DeepSeek-R1&#34;&gt;GitHub&lt;/a&gt;). By Monday its chat app had &lt;a href=&#34;https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-works-well-until-we-asked-it-about-tiananmen-square-and-taiwan&#34;&gt;overtaken ChatGPT&lt;/a&gt; as the most popular free app in Apple&amp;rsquo;s US and UK app stores, and Nvidia had &lt;a href=&#34;https://www.dawn.com/news/1888233&#34;&gt;lost about $593 billion&lt;/a&gt; in market value in a single day, with its shares down almost 17%. No Wall Street company had lost more in a day.&lt;/p&gt;</description>
			</item>
			<item>
				<title>This Week in AI: DeepSeek R1 Takes Center Stage</title>
				<link>https://hitesh.in/2025/this-week-in-ai-deepseek-r1-takes-center-stage/</link>
				<pubDate>Mon, 27 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/this-week-in-ai-deepseek-r1-takes-center-stage/</guid>
				<description>&lt;p&gt;In a week filled with headline-grabbing announcements - from OpenAI&amp;rsquo;s ambitious &amp;lsquo;Operator&amp;rsquo; launch to the controversial $500B &amp;lsquo;Stargate&amp;rsquo; project and the ongoing Musk-Altman feud - one breakthrough stands out for its technical significance. DeepSeek-AI&amp;rsquo;s release of their R1 models represents a fundamental shift in how we approach AI reasoning capabilities, potentially overshadowing even the most dramatic industry developments.&lt;/p&gt;&#xA;&lt;h3 id=&#34;highlight-of-the-week-deepseek-ai-revolutionizes-ai-reasoning&#34;&gt;Highlight of the Week: DeepSeek-AI Revolutionizes AI Reasoning&lt;/h3&gt;&#xA;&lt;p&gt;DeepSeek-AI&amp;rsquo;s introduction of DeepSeek-R1 models marks a pivotal moment in AI development. Their groundbreaking approach uses large-scale reinforcement learning without supervised fine-tuning (DeepSeek-R1-Zero), while offering a variant with supervised fine-tuning (DeepSeek-R1). This development could fundamentally change how we approach AI reasoning, overshadowing even major announcements like OpenAI&amp;rsquo;s Operator and the Stargate project.&lt;/p&gt;</description>
			</item>
			<item>
				<title>10 Cool Ways Small Language Models Are Making a Big Impact</title>
				<link>https://hitesh.in/2025/10-cool-ways-small-language-models-are-making-a-big-impact/</link>
				<pubDate>Wed, 22 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/10-cool-ways-small-language-models-are-making-a-big-impact/</guid>
				<description>&lt;p&gt;Large Language models trained on billions of parameters, like ChatGPT, are always in the news? Well, there&amp;rsquo;s a whole world of smaller models that are changing how we do things. Think of them like pocket-sized smart assistants – they might not know everything, but they&amp;rsquo;re really good at specific tasks. &lt;strong&gt;Here&amp;rsquo;s why that matters: they&amp;rsquo;re faster, cheaper, and can work on regular computers or phones.&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Let&amp;rsquo;s look at 10 amazing ways these mini AI helpers are making our lives easier and more fun:&lt;/p&gt;</description>
			</item>
			<item>
				<title>This Week in AI - Episode 2</title>
				<link>https://hitesh.in/2025/this-week-in-ai-episode-2/</link>
				<pubDate>Fri, 10 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/this-week-in-ai-episode-2/</guid>
				<description>&lt;p&gt;Welcome to the weekly AI digest! This week&amp;rsquo;s highlights include Nvidia&amp;rsquo;s groundbreaking entry into personal AI computing, significant developments in AI privacy concerns, and Bill Gates musings. From workforce transformations to cutting-edge development tools, here&amp;rsquo;s a collection of the most impactful stories shaping the future of technology. Read on for detailed insights and expert analysis.&#xA;&lt;em&gt;Note:&lt;/em&gt; I have left out Sam Altman&amp;rsquo;s Reflections, which I had covered in an earlier post.&lt;/p&gt;</description>
			</item>
			<item>
				<title>State of Data&#43;AI 2024: The Numbers That Matter</title>
				<link>https://hitesh.in/2025/state-of-data-ai-2024-the-numbers-that-matter/</link>
				<pubDate>Mon, 06 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/state-of-data-ai-2024-the-numbers-that-matter/</guid>
				<description>&lt;p&gt;The latest &lt;a href=&#34;https://www.databricks.com/resources/ebook/state-of-data-ai&#34;&gt;Databricks report&lt;/a&gt; reveals a seismic shift in how companies are deploying AI. Here&amp;rsquo;s your executive breakdown:&#xA;📈 &lt;strong&gt;The Big Numbers&lt;/strong&gt;&#xA;• 3X more efficient at deploying ML models to production&#xA;• 11X increase in production models year-over-year&#xA;• 377% growth in vector database usage&#xA;• 76% of companies choosing open source LLMs&#xA;• 70% leveraging RAG for customization&#xA;🏃&lt;strong&gt;Speed to Production&lt;/strong&gt;&#xA;• OLD ratio: 16 experimental models for every 1 in production&#xA;• NEW ratio: 5-to-1&#xA;• Key driver: Better tooling and standardized platforms&#xA;• Result: Faster time to value&#xA;FASTEST-GROWING DS/ML APPLICATIONS, BY INDUSTRY&#xA;🔥 &lt;strong&gt;Hottest Trends&lt;/strong&gt;&#xA;&lt;strong&gt;1. NLP dominance&lt;/strong&gt;&#xA;• Fastest growing ML application (+75% YoY)&#xA;• Healthcare leading adoption at 69%&#xA;• Manufacturing saw 148% YoY growth&#xA;&lt;strong&gt;2. Open Source Revolution&lt;/strong&gt;&#xA;• Meta&amp;rsquo;s Llama 3: 39% market share in just 4 weeks&#xA;• Smaller models (≤13B parameters) preferred&#xA;• Focus on cost and latency optimization&#xA;&lt;strong&gt;3. RAG &amp;amp; Vector Databases&lt;/strong&gt;&#xA;• 377% YoY growth&#xA;• Key to customizing LLMs with private data&#xA;• Reducing hallucinations and improving accuracy&#xA;🏦 &lt;strong&gt;Surprise Leaders&lt;/strong&gt;&#xA;• Financial Services: 88% GPU usage growth in 6 months&#xA;• Healthcare: Highest NLP adoption&#xA;• Manufacturing: Fastest NLP growth&#xA;• Key insight: Regulated industries are early adopters&#xA;💡 &lt;strong&gt;Action Items for 2024&lt;/strong&gt;&lt;/p&gt;</description>
			</item>
			<item>
				<title>This Week in AI: Episode 1</title>
				<link>https://hitesh.in/2025/this-week-in-ai-episode-1/</link>
				<pubDate>Sat, 04 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/this-week-in-ai-episode-1/</guid>
				<description>&lt;h3 id=&#34;-industry-moves--market-dynamics&#34;&gt;🎯 Industry Moves &amp;amp; Market Dynamics&lt;/h3&gt;&#xA;&lt;h3 id=&#34;the-rise-of-new-players&#34;&gt;The Rise of New Players&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Deepseek Emerges as China&amp;rsquo;s AI Giant&lt;/strong&gt; China&amp;rsquo;s rising star in AI reveals ambitious plans to challenge global leaders in deep learning and NLP. &lt;a href=&#34;https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas&#34;&gt;Read more at China Talk Media&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;platform-evolution&#34;&gt;Platform Evolution&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Companies Dominate Forum Discussions&lt;/strong&gt; Recent analysis shows AI-related discussions commanding unprecedented forum traffic. &lt;a href=&#34;https://pod.geraspora.de/posts/17342163&#34;&gt;Read more at Geraspora&lt;/a&gt;&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Google&amp;rsquo;s Search Quality Under Scrutiny&lt;/strong&gt; Critics draw parallels between Google&amp;rsquo;s current challenges and OpenAI&amp;rsquo;s monetization strategy. &lt;a href=&#34;https://chuckwnelson.com/blog/google-search-results-infested-open-ai-using-google-playbook&#34;&gt;Read more at Chuck Nelson&amp;rsquo;s Blog&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;-technical-breakthroughs--innovation&#34;&gt;💡 Technical Breakthroughs &amp;amp; Innovation&lt;/h3&gt;&#xA;&lt;h3 id=&#34;ai-research-developments&#34;&gt;AI Research Developments&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Meta AI Introduces &amp;ldquo;Coconut&amp;rdquo;&lt;/strong&gt; A breakthrough in AI reasoning capabilities through the new &amp;ldquo;Chain of Continuous Thought&amp;rdquo; framework. &lt;a href=&#34;https://aipapersacademy.com/chain-of-continuous-thought/&#34;&gt;Read more at AI Papers Academy&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;educational-technology&#34;&gt;Educational Technology&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Anki AI Utils Revolutionizes Learning&lt;/strong&gt; New AI-powered flashcard system integrates ChatGPT explanations and Dall-E illustrations. &lt;a href=&#34;https://github.com/thiswillbeyourgithub/AnkiAIUtils&#34;&gt;View on GitHub&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;historical-research&#34;&gt;Historical Research&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;AI Decodes Ancient Manuscripts&lt;/strong&gt; Breakthrough in archaeological research as AI tools successfully decipher ancient texts. &lt;a href=&#34;https://www.nature.com/articles/d41586-024-04161-z&#34;&gt;Read more at Nature&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;-privacy--ethics&#34;&gt;🛡️ Privacy &amp;amp; Ethics&lt;/h3&gt;&#xA;&lt;h3 id=&#34;privacy-concerns&#34;&gt;Privacy Concerns&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Apple&amp;rsquo;s Photo Analysis Sparks Debate&lt;/strong&gt; Auto-enabled landmark recognition feature raises privacy concerns. &lt;a href=&#34;https://www.theregister.com/2025/01/03/apple_enhanced_visual_search/&#34;&gt;Read more at The Register&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;military-applications&#34;&gt;Military Applications&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Israel&amp;rsquo;s &amp;ldquo;AI Factory&amp;rdquo; Development&lt;/strong&gt; The deployment of AI systems in military operations in Gaza intensifies ethical debates. &lt;a href=&#34;https://www.washingtonpost.com/technology/2024/12/29/ai-israel-war-gaza-idf/&#34;&gt;Read more at The Washington Post&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;-social-media--user-engagement&#34;&gt;🌐 Social Media &amp;amp; User Engagement&lt;/h3&gt;&#xA;&lt;h3 id=&#34;platform-integration&#34;&gt;Platform Integration&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Meta&amp;rsquo;s AI Bot Integration Plans&lt;/strong&gt; Facebook and Instagram prepare for increased AI bot deployment. &lt;a href=&#34;https://nymag.com/intelligencer/article/meta-wants-more-ai-bots-on-facebook-and-instagram.html&#34;&gt;Read more at NY Mag&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;-industry-perspective&#34;&gt;💭 Industry Perspective&lt;/h3&gt;&#xA;&lt;h3 id=&#34;expert-analysis&#34;&gt;Expert Analysis&lt;/h3&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;25-Year AI Industry Retrospective&lt;/strong&gt; A veteran researcher reflects on AI&amp;rsquo;s evolution from basic systems to today&amp;rsquo;s models. &lt;a href=&#34;https://austinhenley.com/blog/25yearsofai.html&#34;&gt;Read more at Austin Henley&amp;rsquo;s Blog&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3 id=&#34;-join-the-conversation&#34;&gt;📬 Join the Conversation&lt;/h3&gt;&#xA;&lt;p&gt;What are your thoughts on these developments? Share your perspectives on how these changes might shape the future of AI and technology.&#xA;&lt;em&gt;Stay informed with our weekly AI digest. Subscribe to receive future updates directly in your inbox.&lt;/em&gt;&lt;/p&gt;</description>
			</item>
			<item>
				<title>AI Industry Trends 2025: A Transformative Year Ahead</title>
				<link>https://hitesh.in/2025/ai-industry-trends-2025-a-transformative-year-ahead/</link>
				<pubDate>Fri, 03 Jan 2025 00:00:00 +0000</pubDate>
				<guid>https://hitesh.in/2025/ai-industry-trends-2025-a-transformative-year-ahead/</guid>
				<description>&lt;p&gt;As we step into 2025, the artificial intelligence landscape is undergoing remarkable transformations that are reshaping how businesses and societies interact with technology. From shifting corporate strategies to groundbreaking developments in AI infrastructure, let&amp;rsquo;s explore the key trends defining this pivotal year in AI evolution.&lt;/p&gt;&#xA;&lt;h3 id=&#34;the-evolution-of-corporate-ai-strategy&#34;&gt;The Evolution of Corporate AI Strategy&lt;/h3&gt;&#xA;&lt;p&gt;Gone are the days when companies viewed AI as a complete replacement for human workers. Instead, a more nuanced approach has emerged, with organizations now positioning AI as a copilot to enhance human capabilities rather than replace them entirely. This shift reflects a deeper understanding of AI&amp;rsquo;s strengths and limitations in real-world applications (&lt;a href=&#34;https://indianexpress.com/article/explained/explained-sci-tech/looking-at-2025-tech-rising-above-the-ai-line-9750694/lite/&#34;&gt;Indian Express, 2025&lt;/a&gt;).&lt;/p&gt;</description>
			</item>
	</channel>
</rss>
