Z.ai's GLM-5.3 now ships on the API at the same $1.40/$4.40 per-million-token rates as GLM-5.2, with stronger coding and long-horizon agent chops. Open weights remain unannounced, so devs build through the OpenAI-compatible endpoint for now.
Mira Murati's Thinking Machines released Inkling, its first model trained from scratch, a 975B-parameter multimodal MoE (41B active per token) with a 1M-token context window, shipping full weights on Hugging Face under Apache 2.0. An adjustable "thinking effort" setting and the Tinker fine-tuning API are built around letting anyone retrain it on their own data.
Dropping single-LLM-judge reward for two-agent debate recovers 45% of the accuracy gap lost to reward hacking in RLAIF, and holds through many RL steps where the baseline degrades. Tested on a Gemini 2.5 Flash policy against a weaker judge; capping the critic's words is what keeps the game from drifting into own judge-hacking.
A pre-registered study serving Qwen3-235B on one 8GB GPU finds decode is memory-bandwidth-bound (0.44 tok/s), not compute-bound. Auxiliary locality router losses cut cache misses up to 60% but every config blows past a 1% perplexity budget, confirming miss reduction is tightly coupled to quality. Training-free cache-aware rerouting stacks with trained locality for ~80% miss reduction at lower cost.
ClosureBench generates fresh graph-reasoning instances on demand with machine-verified ground truth, so it's contamination-resistant and can measure memorization directly. Across 26 task categories, frontier models degrade sharply from atomic to compositional queries, misreading the graph, then reasoning correctly over the wrong one, while a 4B model fine-tuned to emit programs hits near-frontier accuracy.
FinSkillBench tests whether LLM agents can handle real investment work, point-in-time data retrieval, correct computational inputs, specialized methods, auditable outputs, across 3 domains, 12 subtasks, 2,603 episodes. Curated skill packages lifted mean scores from 0.366 to 0.528 across 9 models, while self-generated skills barely helped despite higher compute. The benchmark, tools, and trajectories are released.
Etched's mission to run the world's inference cleared its first big milestone: the first hardware rack shipped to Jane Street, which led a $700M raise at a $21B valuation after testing the chip. The company is now ramping toward gigawatt-scale deployment with new factories, supply chains, and fleet software.
Anthropic is giving CEO Dario Amodei and co-founders a class of stock with extra voting power ahead of a potential IPO, insulating them from shareholder pressure. It would be the first time founders hold supervoting shares, mirroring dual-class setups at Meta and SpaceX as Anthropic eyes one of the biggest debuts in history.
Cerebras unveiled the CS-4, a rack-scale system it says delivers up to 30x faster AI inference than GPU racks, powered by three WSE-3 Turbo wafers. Initial units are being sampled now, with a modular Nexus architecture that slashes deployment time from days to hours and sharpens the economics challenge to Nvidia.
OpenAI temporarily slowed frontier scaling and paused two weeks of RL training after a security incident and early evidence its Astra model may cross its Critical cybersecurity capability threshold. The largest frontier RL run stays on hold until stronger evidence of aligned behavior is established.
This study shows LLM safety filters are English-centric and fail in non-English languages, so voice assistants and dialogue systems can emit biased, stereotype-reinforcing output for speakers of Hindi, Tamil, Bengali, and more. The authors release INCLUDE, a 2,604-prompt multilingual benchmark, and find open-source models most biased in Bengali while closed models are worst in English.
Miles v0.1 is a production-ready, open full-stack system for post-training AI agents with reinforcement learning after initial training. It runs agentic rollouts in isolated, sandboxed environments via SGLang, scores outcomes, and keeps rollout and training fully asynchronous for scale. Aimed at teams who want frontier-scale RL without building it themselves.
You’re caught up.
That was all 12 stories in the 20 August issue.