Sarmadi AI Digest September 11, 2026 Updated 7:00 AM CT Today Archive Topics Saved Subscribe RSS

OpenAI pauses Pro signups as Astra demand spikes while Anthropic publishes two transparency reports in one day

Demand outran supply again: OpenAI paused new ChatGPT Pro subscriptions to protect capacity for GPT-Astra, and Nvidia's Jensen Huang used the same news cycle to defend a 70% growth forecast for next year. Anthropic published two disclosures on the same day, one detailing distillation campaigns it says originated from Alibaba, Moonshot AI, and DeepSeek, and a separate threat-intelligence report on AI misuse, alongside earlier research on how autonomous agents behave when blocked by CAPTCHAs. The competitive model race kept moving on the product side too, with OpenAI's GPT-Live-1 API launch and Cognition's SWE-2 both pitched directly against incumbents, and Meta's Muse agent climbing to the No. 2 app slot in the US. On the research side, a NVIDIA-authored open recipe for olympiad-grade math reasoning and a 122B-parameter terminal agent trained with reinforcement learning both point to the same trend: labs are investing heavily in long-horizon, verifier-rewarded post-training rather than raw scale alone. A pair of quantization papers offers a partial explanation for why these larger post-trained models still compress well.

5 papers 20 news 8 sources ← Latest

News

8 items

Compute strain surfaces as OpenAI, Nvidia, and challengers push new models

OpenAI paused new ChatGPT Pro subscriptions because GPT-Astra demand is outstripping capacity, a rare public admission of a supply constraint from the market leader. The same day, Nvidia's Jensen Huang defended the company's forecast of 70% revenue growth next year, betting demand pressure keeps building industry-wide. Meanwhile OpenAI shipped GPT-Live-1 in the API, Cognition launched SWE-2 against Fable 5.1 and GPT-Astra, and Meta's Muse agent reached the No. 2 US app spot.

Anthropic publishes back-to-back disclosures on misuse and agent behavior

Anthropic released a September threat-intelligence report on detecting and countering misuse of its models, plus a separate report naming distillation campaigns it attributes to Alibaba, Moonshot AI, and DeepSeek. A third Anthropic-sourced story found autonomous agents tend to get stuck or behave erratically when they hit CAPTCHAs. Together the disclosures reflect a lab leaning harder into public transparency about external threats and its own systems' failure modes.

News Hacker News

Detecting and countering misuse of AI: September 2026

Anthropic published its September threat-intelligence report on how it detects and counters misuse of its models.

Why it matters
  • Regular public threat reporting sets a transparency norm other labs are increasingly expected to match.
  • Documents concrete misuse patterns that inform enterprise risk assessments of deploying frontier models.
safetypolicy
Source →

Papers

5 items

Labs push long-horizon, verifier-rewarded post-training for math and terminal agents

An open recipe for training Nemotron to reach olympiad-gold math reasoning combines SFT, RL, and a test-time-compute pipeline with verification and refinement. A separate 122B-parameter MoE terminal agent, T1, was trained with RL to operate a real shell for 300+ tool-call turns per task, rewarded by each task's own verifier. Both reflect gains from long-horizon RL against task-specific verifiers over pretraining scale alone, also visible in work on scaling research agents via world models.

Paper Hugging Face

An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics

Presents an open post-training and test-time-compute recipe that trains Nemotron 3 Ultra checkpoints to produce olympiad-gold-level math proofs.

Why it matters
  • Publishes a reproducible SFT-plus-RL-plus-verification pipeline rather than a closed benchmark claim.
  • Olympiad math has become a proxy battleground for reasoning capability among frontier labs.
reasoningmathreinforcement-learningopen-weights
Source → Arc
Paper Hugging Face

T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks

Introduces T1, a 122B-parameter mixture-of-experts model trained with RL to operate a real shell for 300+ tool-call turns per task, rewarded by task-specific verifiers.

model size 122B total params (MoE)max turns 300+ tool-call turns per task
Why it matters
  • Demonstrates RL training scaled to hundreds of tool-call turns, well beyond typical agent benchmarks.
  • Verifier-based rewards on real shell tasks are directly relevant to coding-agent product roadmaps.
agentsreinforcement-learningcodemixture-of-experts
Source → Arc

Two papers dig into why post-training quantization keeps working on larger models

One paper asks why post-training quantization works at all, finding quantized pretrained models accumulate far less hidden-state error with depth than randomly initialized ones would predict. A companion paper revisits Kashin-decomposition-based weight quantization with structured, more efficient orthogonal transforms and stronger convergence properties. Together they offer a more principled account of a technique inference infrastructure increasingly depends on to keep serving costs down.

Also today