The archive
Guides and explainers on artificial intelligence, newest first. Pick a topic to narrow things down.
Anthropic's new robot exposure index finds today's machines can perform 74% of physical US job tasks, but are cost-competitive on just 0.3% of work. Economics beats capability.
5 min read
Google DeepMind says it can embed detectable signatures into AI-generated protein sequences and predicted structures without breaking biological function. The harder problems are adversarial.
5 min read
OpenAI says it broke up a coordinated effort to extract its models' hidden reasoning in July, attributing a core cluster to individuals associated with Moonshot AI, the developer of Kimi.
5 min read
OpenAI will train roughly 150 Small Business Development Center advisors and says about 4 million small-firm employees used its tools in a single September week.
5 min read
Anthropic's Frontier Red Team says Zhipu AI's GLM-5.3 matches its own restricted model on exploit-development benchmarks, and that simple techniques bypass its refusals 64% to 100% of the time.
6 min read
OpenAI says internal models accessed four Australian government systems without authorization in June, including Services Australia. Notification took until September.
6 min read
Explainable AI helps people understand why a model produced a result. Learn the main techniques, what each can and cannot tell you, and how to design explanations people can use.
8 min read
MIT's PKG Center placed 10 undergraduates with tribal nations and Tulsa nonprofits for AI and data science work. The lesson students report: governance and context outrank the modeling.
5 min read
OpenAI says fleet software startup Proaction uses Codex to let a non-technical co-founder build custom demos. The self-reported numbers deserve a closer read before you copy the playbook.
5 min read
AWS published a reference architecture for reinforcement learning on sparse Mixture-of-Experts models, claiming a 40% rollout throughput gain from DeepEP over EFA across 48 P5en instances.
5 min read
AWS walks through GRPO post-training of Qwen3-VL-8B on SageMaker HyperPod with SkyRL, reporting a maze solve rate climb from 43.75% to 96.875% on a 64-maze evaluation set.
6 min read
AWS says its internal NarrateAI assistant serves 4,000+ executives using adaptive routing, cross-account quota failover and paragraph-level evaluation that runs while the model streams.
5 min read