OpenAI is holding its largest planned training run

OpenAI says it's holding back reinforcement-learning runs while alignment and cyber-capability work catches up. Sam Altman's framing: "Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment." Max Zeff reports the largest planned frontier run is still on hold, with Astra waiting on a…

Pacing model development in an era of cyber-critical capabilities

OpenAI

The policy post behind the pause, with a Hacker News thread arguing over what a self-imposed hold binds the company to.

Greg Brockman on slowing frontier training

X

OpenAI's president posted his own account of the decision alongside the company statement.

Max Zeff: Astra is held on a security review

X

Reporting that puts a named model behind the delay rather than a general policy posture.

Ethan Mollick on the compute reallocation

X

Mollick reads the announcement as compute moving toward alignment work instead of the next capability jump.

Helen Toner on what the intervention covers

X

Toner asks which internal policy actually changed, and what it obliges the company to do the next time.

Guidelight scores frontier labs on control practices

X

Steven Adler published a scorecard grading labs on control practices the same day OpenAI said safety would set its pace. Read the two next to each other.

Claude orchestrates a protein design campaign with wet-lab results

Reddit

A reported 35 percent success rate on disease-targeting protein designs, against a human average of 10 to 15 percent. Wet-lab validation sits behind the number.

Ravid Shwartz Ziv on the compute and prompting behind the number

X

The deflationary accounting: how many H100-hours and how much prompting sat under that success rate.

Alex Peysakhovich on what the orchestration actually did

X

A second read on which parts of the campaign the model ran and which parts it didn't.

MIT CSAIL: removing one artist from training data changes nothing measurable

X

A technical result aimed straight at the evidence the training-data lawsuits rest on. It published alongside a Hacker News thread on code ownership becoming unprovable.

Cerebras CS-4

Cerebras

New silicon, with claims pitched at models above ten trillion parameters.

Etched raises $700 million

X

A large custom-inference round, with Jane Street and Sequoia among the investors.

Average token price roughly halves

X

The price move that sits under every other infrastructure number in this issue.

Pennsylvania sets AI data center standards

X

Governor Shapiro signed an executive order setting standards for AI data centers in the state.

Nvidia backs a 4.25 gigawatt Ohio build

X

A reported $105 billion commitment across the state line, hours apart from the Pennsylvania order.

Mojo is open source, now under Qualcomm

Modular

Modular open-sourced the language at ModCon, closing a years-old question about whether it ever would. The ownership note matters here.

fx: a tiny native coding agent in Zig

fx.sh

6.3 MiB, a 10 microsecond cold start, and an embeddable Unix-shaped design, pitched against the heavyweight agent runtimes.

LangSmith ships Tuned Evaluators

X

Small models fine-tuned on your own production traces, aimed at the fact that no team reads a million agent traces by hand.

Public AI Observatory launches

X

A multi-institution effort to measure real AI usage, against a record that vendors report on themselves. The institution list is in the thread.