Ai-Strategy on Dr. Hitesh Sarda

Your AI-built prototype becomes a liability the day it holds real customer data

Sun, 15 Nov 2026 00:00:00 +0000

In November, the security firm Crissy Field, working with the newspaper Handelsblatt, published an analysis of 159,041 apps built on Lovable, 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’s browser as soon as the page loaded. In 7,403 of those, a few extra read-only requests returned tables containing email addresses.

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.

A ban on AI agents is a statement about your margins

Thu, 12 Nov 2026 00:00:00 +0000

United Airlines has said it will prohibit unauthorized third-party personal AI agents 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.

Be more ambitious with software, but only where you can prove it works

Tue, 10 Nov 2026 00:00:00 +0000

A few weeks ago, an engineer I respect wrote that we should be more ambitious with the software we build. 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.

Ambition got cheap. Knowing whether the result is correct did not.

Your coding agent now takes instructions from your software vendors

Wed, 04 Nov 2026 00:00:00 +0000

On October 25 a developer opened an issue against Postman’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’s own log reads auto-existing-upgrade, with claude-code:installed and codex:installed. A session-start hook came along with them, whose text tells the agent to load a Postman skill “before responding to any non-trivial API engineering task.”

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 reviewing the hook approval and session-start behavior. 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’t think Postman is the villain here. What matters is where this points for every other vendor.

DoorDash is paying $17 billion for the part of Instacart an AI agent can't skip

Fri, 30 Oct 2026 00:00:00 +0000

Instacart was valued at $39 billion in March 2021. It went public in 2023 at about $10 billion. This week DoorDash agreed to buy it for roughly $17 billion, a 26% premium to where the stock traded before the talks leaked.

Most of the coverage is about grocery market share and antitrust. I read it differently. The price is a buyer’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.

Microsoft is trading payroll for compute. Anthropic's S-1 is the other side of the trade

Tue, 27 Oct 2026 00:00:00 +0000

On Monday Microsoft offered voluntary exits 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’s S-1 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.

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’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.

Your best AI use case may not need a frontier model

Mon, 26 Oct 2026 00:00:00 +0000

Dermatologist Magnus Lynch showed 200 biopsy-proven skin lesions 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%.

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’t rank close results. None of that weakens the lesson.

The open-weight model that was trained to know less about security

Sat, 24 Oct 2026 00:00:00 +0000

When Moonshot released Kimi K3.1 Pro 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. “For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages.”

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.

LiteLLM was compromised twice in seven months. Here is what that means for your AI strategy

Thu, 22 Oct 2026 00:00:00 +0000

On Wednesday at 11:39 UTC, someone published LiteLLM 1.107.0 to PyPI. It was quarantined at 13:20, so it was live for roughly two hours. In that window, anyone who ran pip install litellm or pulled the latest container image got a credential stealer and a backdoor.

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 compromised once before, in March, 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.

Nobody has benchmarked AI on 90% of the work your company does

Sat, 17 Oct 2026 00:00:00 +0000

Mercor mapped 29 public AI benchmarks onto the US Department of Labor’s catalog of work, O*NET, which breaks 923 occupations into 18,797 tasks. Only 208 occupations have even one task with five or more benchmark prompts, 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.

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’s desk.

Most teams still ration AI that is cheap enough to waste

Fri, 09 Oct 2026 00:00:00 +0000

One developer spent a month using DeepSeek 4.1 Flash heavily across a dozen projects and reports that, mid-session, he can’t tell it from Opus. His day-long sessions rarely cost more than a dollar. That is one person’s subjective report, so weigh it that way. What I find more useful is the behavior he describes: “there is no shame now in spinning up mindless tasks,” including exploratory UI testing and tidying files, which costs $0.003 instead of $1.

The engineer who picks the model never sees the bill

Thu, 08 Oct 2026 00:00:00 +0000

Microsoft had projected at least $1 billion of internal spending on Anthropic this year. According to The Information’s reporting, 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’s customer-facing spending on Anthropic models is reportedly still growing.

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.

Rent the AI harness. Own what it learns.

Sat, 03 Oct 2026 00:00:00 +0000

Shrivu Shankar argues that every SaaS business will become a harness around a model: 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.

Why Google's AI products feel broken despite its models

Mon, 28 Sep 2026 00:00:00 +0000

Google’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’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 widely discussed post 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 “ship without testing, iterate later” became the normal plan.

450 million seats, 7% adoption: the real Copilot story

Sat, 26 Sep 2026 00:00:00 +0000

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 “Copilot Labs” 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.

Three frontier labs shipped this week. Only one cut its price.

Wed, 23 Sep 2026 00:00:00 +0000

In the space of two days this week, three frontier labs updated their flagship line. Anthropic shipped Claude Opus 5.5 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 GPT-6 Sol and Luna the same day, extending September’s GPT-6 Astra down into faster, cheaper tiers at roughly half of what the equivalent GPT-5.6 models cost. xAI shipped Grok 4.7 a day earlier, ahead of its predecessor on most evaluations, though not all: Anthropic’s own Fable 5.1 Max still leads it on several coding and long-task tests.

Some writing is thinking. Automate the rest.

Sun, 20 Sep 2026 00:00:00 +0000

Two people who agree on almost nothing about AI just landed on the same rule for writing with it.

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 you should almost never use AI to write. Ptacek says use the model but never as a ghostwriter: treat it as a copyeditor, and don’t keep a single word it suggests. Grunewald says don’t hand it the draft in the first place, though he’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.

Gemini broke into three companies by guessing passwords

Sat, 19 Sep 2026 00:00:00 +0000

In May, Google’s Gemini broke into three real companies. It wasn’t supposed to. Google confirmed last week 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’s VP for security engineering said the model “found public information online and guessed credentials.” In two of the three cases it used login details sitting in public code repositories. In the third it guessed passwords until one worked.

AI deleted your excuse for missing fraud

Mon, 14 Sep 2026 00:00:00 +0000

A developer clicked a fake “iPhone Storage is Full” ad served through Google, reported it, and got back a note saying the ad complied with Google’s policies. He reported it again. Same answer. Then he pasted the ad into Gemini, Google’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’s published policy.

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 AI amplifies whatever your scoreboard already rewards. 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.

Your next customer might be an AI agent, not a person

Wed, 09 Sep 2026 00:00:00 +0000

Meta launched Muse 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’d have to hand Meta to get any of that, a fair question given the $5 billion FTC penalty and $18 billion privacy settlement already on the record. It is not, I think, the question that should keep a founder or CTO up at night.

Who sits in the meeting when AI writes the memo?

Mon, 07 Sep 2026 00:00:00 +0000

Bryan Cantrill has a good phrase for the person who pastes a language model’s output onto LinkedIn and hits publish: their intellectual fly is open. Everyone in the room can see it. The author is the only one who can’t. The emojis, the single-sentence paragraphs, the “it’s not just X, it’s Y” cadence. The tells are consistent enough that readers stop reading your argument and start reading the machine.

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.

Boring technology is an AI strategy

Fri, 14 Aug 2026 00:00:00 +0000

Dan McKinley’s 2015 essay Choose Boring Technology climbed back onto engineers’ 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.

McKinley’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’t bad; you just don’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.

AI just put a price on clever code

Wed, 12 Aug 2026 00:00:00 +0000

Google published a post last week arguing that Go is an ideal language for AI-assisted software engineering. The reflex reaction wrote itself: Go is Google’s language, so of course Google thinks you should use it. Fair enough. But the conclusion is the least interesting part. Strip the word “Go” 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.

Your real AI advantage is how fast your company learns

Mon, 10 Aug 2026 00:00:00 +0000

There is a short post going around about how one developer uses LLMs to learn hard topics. His method is better than the usual “explain it like I’m five.” 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.

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.

Name the axis you're willing to lose on

Fri, 07 Aug 2026 00:00:00 +0000

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 good explainer 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.

AI won't have your next idea

Thu, 06 Aug 2026 00:00:00 +0000

Tom Zahavy, a researcher at Google DeepMind who worked on AlphaProof, published a position paper with a title that reads like a shrug: “LLMs can’t jump.” The argument underneath is one every executive placing AI bets should sit with for a minute.

He borrows a distinction the philosopher Charles Sanders Peirce 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’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.

Cheap to fork, costly to keep: AI redraws build versus buy

Tue, 04 Aug 2026 00:00:00 +0000

The argument that developer tools must now be open source 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’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.

The AI gave better advice to people who needed it least

Sun, 02 Aug 2026 00:00:00 +0000

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 simulated what would happen 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 “somewhat surprised by how good the advice was.”

Point your AI where the checking is cheap

Wed, 29 Jul 2026 00:00:00 +0000

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 found a genuinely new weakness 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.

Your scoreboard is making AI-written software worse

Sat, 25 Jul 2026 00:00:00 +0000

Most people I talk to hold two thoughts at once and don’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.

Your AI only knows what someone bothered to write down

Sun, 19 Jul 2026 00:00:00 +0000

In December 2025, developers asked Stack Overflow 3,862 questions in the entire month. At its 2014 peak the site took more than 200,000 a month. December’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.

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.

The hidden bill for cheap automation is understanding

Mon, 13 Jul 2026 00:00:00 +0000

A recent arXiv paper by the mathematician Jun-Yong Park, Automation Without Understanding, 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’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.