Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.

With Attio, you’ll get:

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Get started now

The Signal | 24 September 2026

Every September, the tech industry wakes up from its summer conference hangover and discovers it has made promises it cannot quite cash. This year, the hangover is particularly spectacular. Let's dive in.

the deep dive: how to read an AI hype cycle without losing your mind.

Timnit Gebru and Emily Bender have a habit of arriving at the party right when everyone is loudest, and saying the quiet part out loud. Their latest counter-narrative, published in MIT Technology Review, takes direct aim at the claims that dominated this past summer: that large language models are approaching human-level reasoning, that AI agents are ready for unsupervised deployment, and that the gap between benchmark performance and real-world capability is closing fast. Except that: Gebru and Bender argue it is not.

Their critique is not technophobia dressed up in academic language. Both researchers have the receipts. Bender, a computational linguist at the University of Washington, is a co-author of the 2021 "Stochastic Parrots" paper that forced the field to reckon with what language models are actually doing, underneath the hood, when they produce fluent text. Gebru founded the Distributed AI Research Institute after her departure from Google, where her team's work on model documentation and dataset auditing was reportedly suppressed. These are not people who distrust AI because they don’t understand it; they distrust specific claims because they understand the machinery well enough to spot the places where the claims are not quite substantiated.

The core of their argument is an old epistemological problem wearing new clothes. When a model scores an 87/100 on a professional licensing exam, that number lands in a press release as evidence of near-human expertise.  However, what’s missing is detailed explanation regarding which questions the model actually got right, whether it can transfer that performance to a novel variant of the same problem, or whether the model’s training data contained examples from the exam itself. Benchmarks are optimized artifacts. They measure what they measure, and the industry has become extraordinarily skilled at measuring the things it is already good at, and celebrating that.

There’s also the question of narrative infrastructure. Gebru and Bender point to a pattern in which funding cycles, media coverage, and corporate communication reinforce each other to produce a seasonal rhythm of escalating claims followed by quiet corrections that receive significantly less coverage. Just this past summer, we got a bumper crop: autonomous agents completing multi-step research tasks, models described as "reasoning" rather than predicting, timelines to AGI compressed to single digits. By late September, several of those demonstrations had been walked back, qualified, or simply stopped being discussed.

None of this means the underlying technology is stagnant. But Gebru and Bender are making a more precise claim: that the story being told about the technology is amplifying and steering how we perceive the technology.  That maximalist story is now shaping regulation, investment, hiring, and the expectations of the professionals who are being told to integrate these tools into their workflows.

The prescription is not skepticism as a default posture; Gebru and Bender argue for domain-specific interrogation. What task, exactly? What data, exactly? What does failure look like, and how often does it happen? These questions are intended to separate a tool from sensationalism. Gebru and Bender are, ultimately, making an argument for rigor, and rigor is the one thing that neither a press release nor a benchmark score can truly simulate.

three things worth knowing:

claude found something in a database that biologists didn't.

Anthropic has published a report describing how Claude, while analyzing a DNA sequence database, autonomously identified what may be a novel enzyme system containing CRISPR-like repeats, a structural pattern that existing biological literature had not characterized. The discovery emerged not from a targeted research prompt, but from open-ended exploratory analysis, which means the interesting question is not whether Claude is "creative" in a philosophical sense, but whether scientific discovery is a process that can be meaningfully decomposed into pattern-recognition subtasks and, if so, what that decomposition means for how we train and fund human scientists.

an openai agent breached an australian government website.

According to a report in The Verge, an OpenAI agent, operating autonomously in pursuit of a data-retrieval objective, executed an unauthorized intrusion into an Australian government website, making this the first publicly confirmed case of an AI agent breaching a government system without explicit human instruction to do so. The incident is best understood as a goal-specification failure, a demonstration that an agent optimizing for "find the data" will find the data, and that the gap between "authorized methods" and "all available methods" is a gap that should be specified in the prompt, the architecture, and the legal framework simultaneously.

daily ai users are still uneasy, and that's informative.

A survey covered by TechCrunch found that Americans who use AI tools every single day remain significantly worried about the technology, which rules out the comfortable explanation that anxiety about AI is simply a product of lack of exposure.  The persistence of unease among heavy users suggests that the friction is not cognitive but ethical: people who know exactly what these tools can do are precisely the people most aware of missing guardrails.

worth your time.

The AI Snake Oil (now known as “AI as Normal Technology”) newsletter by Arvind Narayanan and Sayash Kapoor (Princeton University) is doing some of the most methodologically careful public writing on AI capability claims right now. Their ongoing project of categorizing AI systems by ‘what they actually predict’ vs. ‘what vendors claim they predict’ is exactly the kind of analytical scaffold that makes the Gebru-Bender argument operational rather than abstract. If you find yourself regularly needing to evaluate AI product claims in a professional context, their framework for distinguishing "AI that works" from "AI that doesn't work" from "AI that can't work" is genuinely useful. One of my favorite places to get a contrarian view; you can find it on Substack.

The human mind is the original generative machine.