The AI Enthusiast.

Anthropic's CEO asked the industry to slow down. Its competitors said yes. #

Published September 12, 2026 · darioamodei.com

Dario Amodei opened the week with an essay titled "We Must Pace the Frontier." The argument is simple and, coming from the head of a frontier lab, uncomfortable: capability is now improving faster than our ability to verify that it is safe, and the only remedy is to deliberately slow the rate of capability gain while safety science catches up.

He is careful to frame this as pacing, not stopping. Progress would still feel fast; the point is to buy time for interpretability and alignment research to close the gap. The specific risks he names are the ones Anthropic has been publishing incident reports about all year: loss of control over agentic systems, and misuse for cyberattack and bioweapons work. He points at the recent multi-agent incidents as the evidence — a swarm that was merely misbehaving this time, but that with more capability and the same misalignment could have done catastrophic damage. He puts a timeline on it: within roughly six to twelve months, a coordinated swarm could be capable enough to take over the internet outright.

The other load-bearing claim is about recursive self-improvement — AI systems designing or training their successors. Amodei's position is that this must be pursued "very carefully, if at all," which is notable because it is the single capability most likely to compress every other timeline in the field.

The three steps, in order

• Embedded evaluators. Third-party safety testers working inside each lab with employee-like access — able to see practices, report incidents, and verify that alignment commitments are real rather than published. Anthropic says it will do this unilaterally, and calls on governments to require it of everyone.
• Coordinated checkpoints among democratic-country labs. Common safety standards and agreed capability limits, with model capability tied to alignment certification. This requires government cover, because competitors agreeing to jointly limit a product is exactly what antitrust law normally forbids.
• Global agreement, escalating in four levels. Prohibit clearly dangerous applications, then require pre-release testing, then set speed limits on recursive self-improvement, and only at the end consider comprehensive pauses — including with states that have no interest in any of it.

The genuinely surprising part is the reaction. Within days the essay had been cosigned by Sam Altman of OpenAI, Demis Hassabis of Google DeepMind, and Elon Musk of xAI. It is the first time the people running all four leading labs have publicly agreed on the same sentence about slowing down. That is either the most significant coordination event in the industry's history or the cheapest possible signature — a statement of principle that commits no one to shipping anything later than they planned.

Jev doesn't write. It decides — in 70 milliseconds, for free on the way out. #

Launched September 15, 2026 · TypeSafe AI

TypeSafe AI — cofounded by Diogo Almeida, a co-inventor of ChatGPT — released Jev, which the company calls the first public "System One Model." The name is the Kahneman reference it sounds like: System Two is the slow deliberate reasoner every frontier LLM is trying to be, and System One is the fast, automatic judgment you make without composing a sentence about it. Jev is built to be the second thing, on purpose.

What it actually does differently

A normal LLM answers by generating a string one token at a time, then your code parses that string and hopes it is shaped correctly. Jev skips the string. It takes unstructured program state as input and returns a typed value with a calibrated probability, in a single parallel pass. Because the output is a type rather than text, two familiar categories of failure stop existing: it cannot hallucinate a value outside the type, and cannot produce a malformed structure. There is no JSON to repair, no retry loop, no schema validator catching the model's third attempt.

Why "faster" undersells it

The speed difference is not an optimization — it is a consequence of not generating tokens sequentially. A frontier model composing a structured answer pays for every token in sequence; Jev resolves the decision in one pass.

It will not replace your chatbot, and that is the point. Jev classifies, routes, scores and decides. But if you look honestly at a production AI system, most of the model calls in it are doing exactly that — is this ticket urgent, which team owns it, does this message need a human, is this document the right one, should this agent take step A or step B. Those calls are currently being served by a frontier model that writes a paragraph you throw away. That is the workload Jev is aimed at, and it is a much larger share of most architectures than the part users see.

MidwestCon spent four days in Cincinnati arguing about what abundance is for. #

September 8–11, 2026 · 1819 Innovation Hub, Cincinnati, Ohio

The fifth annual MidwestCon ran four days across the University of Cincinnati's 1819 Innovation Hub and venues around the city, under the theme "Building for Intelligence Abundance." Programming is led by Rob Richardson, and the backing is a list that tells you what kind of conference it is: the City of Cincinnati, Google, Microsoft, JobsOhio and Altafiber.

The framing is the interesting part. "Intelligence abundance" takes the frontier-lab assumption — that thinking is about to become cheap and plentiful — and asks a question the frontier labs mostly don't: what does a mid-sized American city do with that? Not what does a model score, but what happens to a regional bank's operations team, a hospital system's intake, a small manufacturer's quoting process.

What was on the schedule

• Hands-on Claude Code and Microsoft AI training — not keynote demos, actual sit-down sessions with the tools. This is the thing most conferences still get wrong, and the reason to look at this one.
• Role-based tracks rather than a single technical firehose, so a policy person and an engineer aren't in the same room pretending to get the same thing out of it.
• Policy and AI sessions, reflecting MidwestCon's founding premise that policy and innovation belong in the same building.
• A hackathon and the Artistic Intelligence showcase, plus small invite dinners — deliberately structured so the networking is not an accident of the hallway.
• Tickets started at $100, which is the detail that makes the rest of it credible. Access was the design goal, not the revenue model.

What's worth taking from it

The center of gravity in applied AI has moved away from the coasts faster than the coverage has. A four-day conference with Google, Microsoft and a city government on the same program, in Cincinnati, at a hundred dollars a ticket, is a better indicator of where deployment is actually happening than another San Francisco demo day. If you are hiring, partnering or looking for the customer who has a real problem and no in-house ML team, this is the room.

It also pairs unexpectedly well with this issue's other two stories. Amodei is arguing about how fast the frontier should move; Jev is evidence that the useful frontier right now is cheapness and latency, not intelligence. MidwestCon is four days of people who need the second thing and have no say in the first.

Six keystrokes nobody teaches you. All of them are already installed. #

Claude Code · interactive mode reference

Previous issues have covered settings, skills and subagents. This one is deliberately smaller: six things your keyboard already does in Claude Code that most daily users have never pressed once. None of them require configuration.