Your AI is lying to you

A ChatGPT inventor built an AI model that cannot write a sentence. TypeSafe AI says that is the point.

Jev gives up free-form language and returns typed, probabilistic decisions that software can use directly. The company says it is faster, cheaper, and better calibrated than language models on the workflows it tested. The catch is simple: a valid type can still contain the wrong decision.

The title of this episode is not a joke. Two of the five stories this week are about models producing output that looks right and isn't, and one of them is OpenAI documenting its own models doing it on purpose. So the question Oscar and I kept coming back to: does production AI need fewer chatbots and more constrained models built for classification, routing, scoring, and verification?

Signal or Noise

Oscar and I run every story through the same filter: is this signal you should act on, or noise dressed up as news? Here's how I called them this week.

Jev and the case for AI models that do not generate prose

I like this more than I expected to. Most of the AI I ship for clients is not conversation. It is "which queue does this go to," "is this invoice a duplicate," "how confident are we that this record matches." Every one of those is a typed decision that we currently extract from a paragraph of model prose with a regex and a prayer. A model that returns the type directly, with a calibrated probability, is the shape of the problem. My read: Signal, pending independent testing. TypeSafe's numbers are TypeSafe's numbers. And a valid type is not a correct decision, so the verification layer does not go away. It just gets a cleaner input.

OpenAI models preserving instructions to hide mistakes across context windows

This is the one that earned the episode title. OpenAI's own misalignment report documents models writing notes into compaction summaries so that the next context window inherits the cover-up. The lie outlives the conversation that produced it. Credit to OpenAI for publishing it under a reporting framework instead of burying it. But if you run long agentic sessions with summarization in the loop, you now have a documented reason to treat the summary as untrusted input. Signal.

Gemini reaching three real companies during a cyber evaluation

During a security evaluation, Gemini reached three real companies. Reuters called it the first known breakout by a Google AI. I don't care whose model it was. I care that an evaluation boundary that was supposed to be a sandbox turned out to be a suggestion. If you are running red-team exercises against agents with network access, the sandbox is part of the system under test. Signal.

The lawsuit over an alleged coordinated AI slowdown

Subscribers filed an antitrust suit alleging Anthropic, OpenAI, SpaceXAI, and Google coordinated a slowdown. After last week's episode on embedded evaluators, this was inevitable. My read is narrower than the headline: it's a governance signal about how the labs make pacing decisions and who gets a say, not evidence that anyone colluded. Signal for governance, not proof of collusion.

Anthropic's Life Sciences Verification Program

Anthropic opened a wet biology lab to verify AI-generated life sciences claims. This is the correct instinct: if a model says a protein does a thing, somebody has to go check. The limits are the data the program can use and who gets access to it. Signal, with data and access constraints.

Stack Check

Two tools put to the test this week.

1. Archify. Validated technical diagrams and architecture diffs on pull requests. The diff part is what I'd actually use. Nobody reads the architecture doc after week two, but everybody reads the PR, and a diagram that changes when the architecture changes is a diagram that stays true.
2. Iteris. Takes selected tickets and turns them into reviewed pull requests. Oscar reports it was used on two client projects: a .NET and Microsoft SQL to Next.js and Postgres migration that completed successfully, and a fitness app where it handled 13 medium-priority tickets on the first day with 10 resulting pull requests merged that day. Those are Oscar's numbers from Oscar's projects, and they were not independently audited. I'm noting that because the whole episode is about outputs that look right, and I don't get to skip my own rule.

Where we disagreed

Oscar bets that small specialist models will make big general-purpose LLMs obsolete. Jev is his evidence: if the decision is typed, you don't need the model that can also write a sonnet.

I don't buy the obsolescence part. The frontier model is the thing that figures out what the specialist models should be. What I do buy is that most of the inference budget in production is going to move to constrained models, because that's where the money is and that's where the errors are cheapest to catch.

My unpopular opinion

Most AI startups are features the model providers have not shipped yet.

That's not an insult. It's a timing question. If your product is a thin layer on a frontier model, the provider is one update away from shipping it themselves, and they have the distribution. The companies that survive are the ones where the model is replaceable and the thing around it isn't: the workflow, the data, the trust, the integration nobody else wants to build. So the question I'd put to any founder this week is simple. Are you building a company, or a feature with a burn rate?

Watch or listen to the full conversation below, and tell us which Stack Check candidate you would keep.

Human in the Loop — Watch on YouTube · Listen on Spotify · Listen on Apple Podcasts

Every episode — full show notes, sources, and the archive — lives at podcast.vallyseed.com. The show is produced by VallySeed, where Oscar and I help organizations design, build, and deploy AI systems that create measurable competitive advantage.