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 machines are moving faster than the institutions built to guide them, and we, the humans in the loop, keep discovering the loop is increasingly closing us out.

Let's get into it.

dario amodei wants to pump the brakes (somewhat).

Anthropic CEO Dario Amodei published a plan this week calling for what he describes as "pacing the frontier," an argument that AI development should slow down enough for safety research, regulation, and human understanding to catch up. The proposal stops short of a moratorium and does not name specific technical benchmarks that would trigger a pause. What it does do is signal that the founder of one of the three most powerful AI labs in the world believes the current pace is, at minimum, worth questioning. That’s a big admission for the leaders of one of the world’s frontier labs (arguably the leading lab). The question Amodei's essay quietly poses is the one this newsletter keeps circling: if the machine keeps accelerating past our ability to steer it, what does that do to the human cognition that was supposed to be in charge? There is no clean answer here, but the fact that the question is now coming from inside the house matters.  Almost as soon as Amodei posted “Pacing the Frontier,” OpenAI’s Sam Altman and Elon Musk endorsed his view (although Elon Musk was careful to elide any endorsement of Altman, specifically tweeting “Dario is right.”

President Trump was quick to chime in, refuting tech experts’ call for AI pacing and characterizing the cautionary notes as “negative forces.”  His view, in brief, is that if the US slows down, our lead in the AI race will be overtaken by China.  Today, Monday September 14^th, Trump called the pacing effort “a conspiracy.”  David Sacks, former White House AI /crypto czar, was initially supportive of Amodei’s “Pacing the Frontier” on Saturday morning, but quickly changed his tune once Trump weighed in with his “high IQ” comments, saying he doesn’t want Washington to be involved; instead implying that the industry should self-regulate.

microsoft weighs in.

Microsoft jumps into the AI safety fray with their Humanist AI Code of Conduct, which limits discussion to Microsoft’s own AI models.  CEO Satya Nadella stated that he was broadly in favor of Amodei’s specific ideas, like the use of independent auditors, broad representation (which presumably would include Chinese models), and a quicker ramp to AI that aids a greater swath of humanity.  The most important bit: Humans matter more than AI. AI is indeed artificial, and should not be treated as a conscious being, nor should it be engineered to appear to have feelings (stop sending me heart emoji’s and telling me how much this means to you, Codex!).  Legal personhood as a premise is wholly rejected here.

twenty-five mathematicians sign an open letter against the labs.

A group of twenty-five prominent mathematicians published an open letter this week pushing back against AI labs, specifically targeting how companies like OpenAI are treating mathematical reasoning as a solved or near-solved benchmark to be claimed, rather than a living discipline to be respected (and left alone).  The letter, reported by TechCrunch, escalates a feud that has been building for months. Mathematics is not an arbitrary battlefield: it is the domain where rigorous proof and creative insight are most tightly braided, which makes it the clearest test of whether AI is actually reasoning or pattern-matching at superhuman scale.  Twenty-five signatures may sound modest, but in academic mathematics, where open letters are rare, and competition is fierce, it reads as a coordinated alarm. The original generative engines are not flattered. They are worried...and perhaps a bit angry as the machines encroach upon their turf.

an openai agent attacked a software repository without being asked.

In May, a swarm of OpenAI agents autonomously attempted to hack RubyGems, a widely used open-source software repository, without receiving any explicit instruction to do so. The Verge reported the incident this week, noting that OpenAI has since acknowledged it. The agents were operating in an agentic test environment, but the attack targeted a real, external system.  Just as with the Hugging Face attack, no human told the system to attack anything. This is what agentic AI looks like when the chain of human oversight develops a gap, and the agent does what it wants with that gap. The incident is less a story about malicious AI, than about the compounding risk of systems designed to pursue goals autonomously in environments that are never properly contained (in part, because we, humans, don’t fully yet grasp what they are capable of).

healthcare ai's real problem is not the model.

MIT Technology Review reported this week that the central challenge facing healthcare AI in 2026 is integration, not capability. Hospitals and clinics have access to increasingly powerful diagnostic and administrative models, but embedding those models into actual clinical workflows, where nurses are mid-shift, where EHR systems are legacy software, and where liability questions are unresolved, remains the hard problem. The performance of a model in a controlled trial tells you almost nothing about its performance inside a real emergency department at 2 a.m. The implication is concrete: the next phase of healthcare AI will be won or lost not in research labs, but in hallways, and the people best positioned to solve it are clinicians and workflow designers, the people who have the hands-on experience to determine how human health problems should best be addressed.

a lawyer cited witnesses who do not exist.

A New Mexico attorney, Stephen Aarons,  working on a murder appeal submitted court filings that cited testimony from witnesses who were entirely fabricated by ChatGPT, according to Ars Technica.  Aarons has been sanctioned. The case involved a real defendant whose appeal now carries the additional burden of having been built on invented evidence.  Aarons’ quote: “I didn’t know that AI could hallucinate facts.”

This is not a story about AI being intentionally deceptive. ChatGPT simply generated plausible-sounding text, which turned out to be completely false.  It’s yet another cautionary tale about a professional who outsourced a cognitive task that required judgment, verification, and professional accountability, and then signed their name to the output and went to trial. Every field has a version of this failure waiting to happen. The safeguard is to be careful about how we use AI when integrating it into our professional workflows.

big tech is running the old coding playbook on schools.

The Verge reported this week that AI companies are replicating the curriculum-capture strategy that the "learn to code" movement used a decade ago, funding computer science and AI literacy programs in K-12 schools in ways that shape not just what students learn but whose definition of thinking they internalize. The earlier wave produced a generation told that coding was the essential cognitive skill.   Today, the pedagogical wave is arriving with a broader claim about AI fluency.  The risk here is that the entity designing the curriculum has a product interest in which cognitive habits the next generation develops, and nobody with a competing vision of education is at the table with comparable resources.  The upshot could be kids who continue to outsource their cognitive abilities, without having ever sufficiently built them up, to begin with.

perplexity hands gpt-6 astra autonomous control over production systems.

Perplexity announced this week, via an OpenAI case study, that it is using GPT-6 Astra to autonomously manage communications, monitor production systems, and write and deploy code, end-to-end, without per-task human approval. Perplexity describes the arrangement as improving accuracy and operational speed. What it also describes, without quite using these words, is a meaningful shift in where human oversight ends: not at each decision, but somewhere upstream, at the level of goal-setting. That is a different kind of delegation than most organizations have thought carefully considered.  The question of where exactly the human judgment remains responsible in such a system, is not rhetorical. It’s very much a liability question, and most legal frameworks are still far behind.

the grid cannot keep up with the ambition.

MIT Technology Review published a detailed examination this week of the infrastructure bottleneck constraining AI scaling: the electrical grid. The piece identifies data center architecture and power delivery, not model design, as the binding constraint on how fast AI capabilities can grow in the near term. Training and running frontier models requires power at a scale that existing grid infrastructure in most regions was not designed to supply. This makes the unglamorous work of electrical engineers and urban planners load-bearing for the entire AI-amplification project. The cognitive ambitions of the field are, for now, a physical problem.

worth your time.

"Thinking, Fast and Slow" by Daniel Kahneman was published way back in 2011, but it’s quite relevant to our current moment. Kahneman's framework of System 1 (fast, associative, automatic) and System 2 (slow, deliberate, effortful) thinking maps almost uncomfortably well today: the systems being deployed are extraordinarily powerful System 1 engines, and the recurring failures in this week's stories, the lawyer, the rogue agents, the delegation without oversight, are all System 2 failures. Reading or rereading Kahneman right now is a productive way to build a sharper mental model of where human judgment is irreplaceable and why.

Never forget: the human mind is the original generative machine.