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

Happy October!

September was a busy month in AI. There were calls to slow down the pace of the AI development coming from within the frontier labs, calls to keep things exactly the same from the White House, a visit by President Xi Jinping to the White House in which Xi asserted that humans must stay in control of AI, and a re-branding of Artificial Intelligence to “super-intelligence” by Donald Trump.

Earlier this week, President Trump hosted roughly 2 dozen AI executives for lunch.  The seating chart went up on his Truth Social account: the president between Jensen Huang and Elon Musk, then Mark Zuckerberg of Meta and Sundar Pichai of Google; across the table, Vice President Vance between Jeff Bezos and Speaker Mike Johnson.  Satya Nadella of Microsoft was there.  OpenAI's president Greg Brockman was there.  Dario Amodei was there, 2 days after his first one-on-one dinner with the president and a day after meeting Senate Majority Leader John Thune, who is working with Amy Klobuchar on a bill aimed at catastrophic AI harms.  The luncheon resulted in what Trump called a “morally binding'“ document, asserting that AI companies will review each other’s work.

In other news, OpenAI declined to release their latest model on the eve of their OpenAI Dev Day. At the Dev Day itself, they released (among other things), Dots, which will rival Meta’s Muse. Whether Dots can make up for the negative press and lawsuits resulting from OpenAI’s agent breaches and hacks remains to be seen. The Federal Trade Commission announced an investigation into Anthropic, OpenAI, and other AI companies. The stated investigation is meant to probe whether said companies have broken federal laws that prohibit unfair and deceptive practices.

Finally, Trump suggested that he would appoint Jay Clayton, current Director of National Intelligence, to the role of AI czar. Mr. Clayton appeared on CNBC’s Squawk Box yesterday and stated that he is not in favor of a pause on AI acceleration; in an interview with Joe Kernan, he said he doesn’t “….know what a pause is.”

Now, let’s get into some AI advances.

the mind's eye, now readable.

Researchers have built an AI system capable of reconstructing images of what a person is looking at, using only the pattern of their brain activity captured by an fMRI scan. The work, reported by MIT Technology Review, today, October 1st, uses a diffusion model trained to map neural signals back to visual content, producing reconstructed images that are recognizably similar to what the person actually saw. The system does not guess in broad strokes. It recovers structure, shape, and in many cases, enough specific detail to identify the source image from a lineup.

Essentially, this is a kind of mind-reading tool, utilizing AI to “guess what you’re looking at my analyzing your brain scans,” according to the MIT Technology Review.

The technical pipeline works roughly like this: fMRI captures blood-oxygen-level-dependent signals across the visual cortex while the subject views images. A model learns the statistical relationship between those neural patterns and the pixel-level features of the images. A generative model then runs that relationship in reverse, synthesizing a plausible image from a fresh brain scan. So far, the reconstructions resulting from this process have not been perfect. They are sometimes blurry, and they sometimes miss the finer details. But they are correct in ways that matter, suggesting that the model has learned something genuine about how visual information is encoded in the human brain.

Here’s where it gets philosophically interesting, and professionally consequential.

The model did not study neuroscience. It did not study critical works by leading researchers on cortical receptive fields.  Instead, it found the structure of visual encoding by brute-force pattern matching across enormous datasets of brain scans paired with images.   This is Richard Sutton’s The Bitter Lesson, applied to neuroscience.  And yet, what it found appears to be real structure. The fact that a diffusion model can invert the visual cortex suggests that the visual cortex is, in some non-trivial sense, doing something a diffusion model can learn to mimic.

That is not a comfortable thought if you believe human perception is categorically unlike machine processing. It is a clarifying one if you believe understanding the brain and understanding AI are the same project approached from different ends.

The immediate applications researchers cite are medical: restoring communication for people who cannot speak or move, understanding disorders of visual processing, building better brain-computer interfaces. Those are real and worth taking seriously. But the longer implication runs deeper. If we can decode the visual stream, the auditory stream is a natural next target. Language comprehension, emotional response, and memory retrieval could all be on the table. Each step is technically harder, but the methodological template now exists to tap into these areas.

This also raises a question about what privacy means when the boundary between inner experience and external signal becomes permeable. The subjects in these studies are consenting participants in controlled lab settings. The fMRI machines required are large, expensive, and not portable. But the history of every sensing technology is a history of miniaturization and cost reduction. The constraints that make this feel safely contained today are engineering constraints, not permanent ones.  Some future version of Meta’s Ray-Ban’s could allow the wearer to read another person’s mind.

What the mind's eye sees has always been, by definition, private. The interesting question is not whether AI can eventually read out thoughts, but what it means for our understanding of thought itself that reading it turns out to be a tractable engineering problem at all. Perception, it turns out, is less ineffable than we supposed. The machine is not mystified by the visual cortex. It just needed enough data!

three things worth knowing this week:

the discovery question.

Anthropic's Claude agents are now conjecturing inside a real molecular biology laboratory, according to MIT Technology Review's September 28th report, prompting researchers to ask formally: at what threshold does an AI system stop being a sophisticated instrument and start being a scientific author?  Can an AI be a reasearcher on its own, can it win its own Nobel Prize?  The piece does not resolve the question, which is precisely why it is essential reading. The criteria we build for "AI-made discovery" will determine credit, liability, and the future structure of research careers.

the robot atlas.

Anthropic's economics research team published a detailed framework mapping what physical AI systems can and cannot yet do across occupational categories, identifying manipulation of irregular objects, in-context adaptation to novel environments, and fine dexterous motor tasks as the three domains where robots still fail reliably. For any professional genuinely trying to assess where their own work sits on the automation frontier this names the actual technical hurdles rather than gesturing at them.

sam altman's delegation stack.

In a conversation on The Every Podcast, Sam Altman described using an always-on personal AI agent to filter his schedule, triage communications, and protect blocks of uninterrupted time for the cognitive work he considers irreplaceable. The practical detail that stands out: he does not treat the agent as a simple productivity multiplier for doing more things. He treats it as a barrier around the hours when his thinking is actually at its best, which is a meaningfully different frame than what most power users are reporting.

worth your time.

If the mind-reading story lit something up for you, the place to go deeper is not another AI paper. It is neuroscientist Stanislas Dehaene's book How We Learn, which lays out the four pillars of biological learning (attention, active engagement, error feedback, and consolidation) that current AI systems approximate in some ways (and violate in others.) Reading it alongside the brain-scan reconstruction story makes both richer. You will finish it with a sharper sense of which aspects of human cognition the machines have genuinely cracked and which remain, for now, untouched.

The human mind is the original generative machine.