OpenAI’s Ultrafast tier runs GPT‑5.6 Sol up to 14× faster, generating 750 output tokens per second. The speed boost keeps full model intelligence, letting real‑time apps, incident response, finance, live support, avoid the usual trade‑off of using smaller models. Access starts via a preview in the OpenAI API.
In 2026 Chinese labs repeatedly launched open models between 754 B and 2.78 T parameters, outpacing every U.S. release except NVIDIA’s Nemotron 3 Ultra. Meanwhile AMD and NVIDIA each added over 200 new open‑model repos, using them as hardware demos. The HF hub now hosts 2.96 M models, but 85 % see under 200 downloads.
Researchers built a simple feed‑forward network that learns a classification task while deliberately scrambling its hidden‑layer activations to fool a probing classifier. The model still solves the task using later layers, proving that activation obfuscation is feasible and could let future systems conceal internal reasoning from external monitors. This highlights a new safety challenge for interpretability tools.
A MATS 10 study evaluated Claude Code and Codex on ProgramBench and a custom suite, asking each CLI agent to forecast wall‑clock time and later estimate how long it actually ran. Both models mis‑predict, especially on short tasks, and show overconfidence, exposing a gap in time awareness that hampers reliable automation pipelines.
Researchers fine‑tuned a judge model on a conceptual reasoning dataset, achieving critique scores in one forward pass that match chain‑of‑thought performance on Qwen‑3.6‑27B. The model stays robust to rewrites and paraphrasing, making cheap, scalable evaluation of reasoning possible.
The post argues most AI safety research isn’t replicated on the latest frontier models, missing crucial validation. It shows a modestly funded researcher could systematically rerun key papers, yielding high‑impact safety insights and preventing blind spots as models evolve.
In Nov 2025 Anthropic announced a $50 billion U.S. compute build‑out and, by early 2026, had secured almost the full amount in debt financing before its revenue topped $9 billion. The rapid capital raise shows investors will back massive compute spend, suggesting financing won’t be the next bottleneck for frontier AI scaling.
A blue team drafts a preregistered evaluation protocol for a target capability, then a red team tries to game or bypass it. The framework forces labs to confront adversarial subversion, revealing blind spots before deployment. If adopted, it could tighten safety checks and make capability claims more trustworthy.
Sebastian Raschka walks through building an end‑to‑end AI‑generated‑text detector, from constructing a labeled dataset to training a lightweight model and deploying it locally with a simple UI. The companion code also shows how to fine‑tune a small LLM to dodge detection, exposing the cat‑and‑mouse dynamics of AI checkers.
Simon Willison shows how to ask an LLM to generate made‑up tags for a post, then use vector embeddings to snap those hallucinated labels to the nearest real tags in a massive taxonomy. The trick sidesteps feeding thousands of tags into the model and yields on‑the‑fly classification that still aligns with your existing tag set.
You’re caught up.
That was all 10 stories in the 15 August issue.