The AI tool you tried last quarter waited for a prompt, hallucinated a number, then asked if you'd like a summary.

Viktor opened a PR at 2am, rebased it against main, ran your test suite, and posted a note in #eng: "Two flaky tests in payments service, both pre-existing. Recommended merging after fixing them." Then drafted the customer reply for the support ticket the bug created.

That's 619K autonomous actions per day across 20,000+ teams. Not chat replies. Real work shipped to GitHub, Stripe, Linear, Notion, and 3,000+ other tools, from inside Slack and Microsoft Teams.

You don't supervise him any more than you supervise a senior engineer.

SOC 2 certified. Your data never trains models.

"It's what you probably originally thought AI was going to be when you first heard of it in sci-fi movies." Tyler, CEO.

Get Started for Free

This week, researchers got a small, genuine look inside a thinking machine — and what they found was both illuminating and humbling. Meanwhile, elsewhere in the stack, AI was firing people, pretending to be agents, learning to attack itself, and quietly consuming the music of millions of artists who never agreed to train it. The minds we are building are telling us something. We are only beginning to learn how to understand them.

Anthropic's researchers have published findings on what they're calling a new window into Claude's "internal thoughts" during its reasoning process — a technique that surfaces, in human-interpretable form, some of what the model is doing as it works through a problem before it produces an answer. The result, reported in MIT Technology Review, is a genuine milestone in mechanistic interpretability: the first credible attempt to read something like a cognitive trace in a frontier model during live inference, not just after the fact.

The technique matters because the industry has long faced a frustrating asymmetry: we can observe what a model says, and we can observe what inputs produced that output, but the space between stimulus and response has remained largely opaque. What Anthropic claims to have done is find structure inside that space — interpretable patterns in Claude's extended reasoning chain that correspond to identifiable conceptual operations. Think less "we read the model's diary" and more "we found that certain neural activation signatures reliably precede certain types of reasoning moves." It is a subtle but real distinction, and it is where the story gets interesting.

Here is what the findings do not show: that we now understand Claude. Anthropic's own researchers are careful on this point, and it’s worth taking them at their word. Identifying a pattern in a neural network's activations and understanding why that pattern exists, what caused it, and whether it generalizes are three entirely different problems. The history of interpretability research is littered with what the field sometimes calls "ghost features" (correlations that look meaningful until they collapse under scrutiny). These are not criticisms of Anthropic's work; it is simply context that the enthusiasm around the announcement has a tendency to compress.

What the research does establish, more valuably, is a methodology. The approach of probing reasoning-chain activations during inference — rather than examining weights in isolation or analyzing outputs post hoc — opens a different angle of inquiry. It suggests that the "thinking" Claude performs in its extended scratchpad may be more structured than previously believed, which has implications for both safety research and alignment. If models have something resembling internal deliberation, then understanding when that deliberation goes wrong becomes more tractable. And if it doesn't; if the appearance of structured thought is itself a kind of artifact — that is arguably even more important to know.

The deeper provocation here is philosophical, and it is one that cognitive scientists have wrestled with in humans for decades: what is the relationship between a process and the account that process gives of itself? When Claude's reasoning chain says "let me think about this step by step," is that narration a description of cognition or a performance of it? Anthropic's new tools may eventually help answer that question. For now, they have made it considerably more precise; which, in science, is often how progress actually begins. The interpretability problem is not solved. It is now better defined, and that is a different and more durable kind of advance.

A lawsuit filed against Meta alleges that the company's layoff decisions were driven by an AI system — and that the system systematically disadvantaged workers who were on medical leave at the time of the cuts, a pattern the plaintiffs argue no properly supervised human decision-maker would have allowed to stand. The case is still early, but it crystallizes something that employment lawyers, HR practitioners, and anyone with a job should be paying close attention to: the question of who bears legal and moral accountability when an algorithm determines livelihoods is no longer hypothetical, and courts are going to have to develop an answer. Source: Ars Technica

A survey of 101 enterprise organizations found that 71% of the systems companies are currently calling "AI agents" are, by any rigorous definition, chatbots — stateless, single-turn, incapable of taking autonomous action across a workflow. The same report identified that the primary obstacle to deploying genuine agentic AI is not a shortage of orchestration platforms (there are many) but a deployment problem rooted in process ambiguity: companies cannot hand autonomous systems into workflows they have not yet clearly defined themselves, and the sophistication of the tooling is currently running well ahead of the organizational thinking required to use it. Source: VentureBeat

OpenAI has built GPT-Red, an adversarial model trained specifically to probe and stress-test its other models for vulnerabilities — a kind of AI sparring partner whose sole function is to find the seams in its siblings' reasoning before bad actors do. The approach extends a well-established practice in cybersecurity, red-teaming, into the model-versus-model domain, and early indications suggest adversarial LLMs are meaningfully more efficient at surfacing edge-case failures than human testers working alone; what remains an open and genuinely hard question is whether a system trained on the same underlying data distributions as its target can ever fully escape the blind spots they share. Source: MIT Technology Review

MIT researchers have developed a tool that lets non-expert users inspect an AI model's neural network activations before the model produces any output — a before-you-trust-it interface that makes interpretability a user-facing feature rather than a researcher's backstage instrument. The practical implication is specific: a professional evaluating an AI's recommendation in a high-stakes context could see which internal features activated most strongly in generating it, giving them a basis for skepticism or confidence that goes beyond reading the output and hoping for the best. Source: MIT Technology News

The human mind is the original generative engine.