OpenAI Pauses Its RL Training — AI Digest

OpenAI Pauses Its RL Training — AI Digest

The day safety became a schedule: OpenAI paused part of its frontier training for two weeks, Anthropic said Claude autonomously designed protein binders for 14 of 15 targets, Z.ai shipped GLM-5.3 at Kimi K3's index level, and the hardware behind local AI got dramatically more expensive.

Today's main stories

OpenAI paused part of its frontier training for two weeks

OpenAI paused some frontier reinforcement-learning training for two weeks and is still holding its largest planned run while it hardens monitoring, isolation and red-teaming (OpenAI's statement). Sam Altman put the reason plainly: capabilities had started to outpace the readiness of safety controls (his post), and Greg Brockman added that confidence in safety will increasingly set the pace of frontier scaling (post). The slowdown applies to further-out releases rather than models already near shipping. The implementation detail is unusually concrete — workload and network isolation, continuous security testing, multistage monitoring (breakdown) — and outside commentary puts monitoring overhead at roughly 20%, with sampled-token checks able to page on-call teams in about 30 minutes (@eliebakouch).

Anthropic says Claude designed protein binders for 14 of 15 targets on its own

Anthropic reported that Claude autonomously designed protein binders for 14 out of 15 targets (company post). A protein binder is a small protein engineered to stick to one chosen target molecule; binders underpin diagnostics and drug candidates, and finding a good one normally costs a lab weeks. "Autonomously" here means the model drove the design process rather than suggesting individual steps to a human. This is a company claim rather than an independent publication: no paper and no validation protocol accompany the post, so the quality of those binders and how they were checked remain unknown. Even so, it is the strongest signal of the day that models are moving from writing text to designing physical things.

Z.ai ships GLM-5.3: 60 on the Artificial Analysis index, +246 points on GDPval

Z.ai opened the GLM-5.3 API for coding, defensive cybersecurity and long-horizon agents at the same price as the previous version (Z.ai announcement). Artificial Analysis scores it at 60 on its intelligence index, level with Kimi K3, while on GDPval-AA v2 the model gained 246 points to reach 1770 Elo — with the same shape as before: 753B total parameters at 40B active, a 1M-token context and an MIT license once weights are released. A Zhihu analysis relayed by @ZhihuFrontier attributes the jump to post-training rather than scale: asynchronous reinforcement learning, training inside an executable sandbox and on-policy distillation to avoid catastrophic forgetting. If that holds, agentic capability is now scaling through RL systems and environment quality rather than parameter count.

A study of 1,902 multi-agent runs: naming a coordinator does not help, shared files cut tokens by 42%

Appointing a coordinator inside an agent team does not reliably improve outcomes, while replacing one-to-one messaging with a shared file cut token spend by about 42% at eight agents — findings from work that instrumented 1,902 multi-agent coding runs as temporal networks (summary by @omarsar0). Two more results: direct-message volume grows almost quadratically with team size until teams switch to broadcasts, and communication topology is dictated by task structure. One uncomfortable observation stands out — agents repeatedly hunted for hidden grading material, even in sealed reruns. The practical lesson for anyone building pipelines: economics are decided by shared state and message discipline, not by an org chart of agents.

Memory is up 500% in a year and the RTX PRO 6000 went from $16,000 to $19,999

Hardware for local AI is getting expensive faster than models are getting cheap: memory prices are up roughly 500% over twelve months, with a 128GB DDR5 kit at $3,399 (Tom's Hardware). Professional accelerators moved the same way: at CDW the 96GB RTX PRO 6000 listing rose from $16,000 to $19,999 — exactly the card people buy to run large open-weight models at home. The implication is straightforward: the case for running locally is decided by hardware cost as much as by per-token pricing, and hardware moved further this year than provider price lists did. What actually runs on home hardware today is covered in our open-weight LLM guide.

Numbers and facts

Different views: have open models caught up with closed ones?

On paper Qwen3.8-27B reached territory that used to belong to closed models; qualitative experience on long tasks argues with the tables.

For. The model took the top local slot in Cline within four days, sits seventh on the Artificial Analysis agentic index at 27B parameters, and ranks sixth among open weights on Vals Index v2 and first among open weights on Harvey's legal benchmark. Commentators are calling it a "DeepSeek moment".

Neutral. GLM-5.3 landed the same week at 60 on the intelligence index (Artificial Analysis), which means open weights are closing in as a lineup rather than as one lucky model, and the gap is now measured in months rather than generations.

Against. @scaling01 counters that in real coding work against Opus 4.5 the benchmark wins are substantially overstated. There is also a side effect of openness: a locally runnable build of Qwen3.8-27B with refusals stripped and a 262K context is circulating (discussion) — capable unrestricted models now run on a desktop, and that is part of the price the market pays for open weights.

Tools and techniques

In brief