Clauday.

Somebody finally ran the experiment everyone has been arguing about in comment threads. HarnessTax put 21 model-harness pairs through the same gauntlet: seven models across three h…

Take a language model, teach it 100 tasks one after another with no access to the earlier examples, then ask what it remembers. Naive sequential fine-tuning retains 1.2%. That is t…

Enclave ran DeepSeek V4.1 Flash against its hacking benchmark and it went eleven for eleven: code execution on every vulnerable target, and all four patched targets held. Total cos…

Atria Dawn landed on arXiv September 14 as 2609.15818 and went straight to the top of Hugging Face's daily papers with 270 upvotes. Lead author Honglin Guo, and more than 142 co-au…

Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15 at https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemi…

Benchmark Radar is a living database and search engine for AI benchmarks, sitting at 75 upvotes on HuggingFace's daily papers. Paper at https://arxiv.org/abs/2609.11115, submitted …

DataFlex-RL sat at the top of HuggingFace's daily papers board with 96 upvotes, and the reason is that it is a negative result, which almost nobody publishes. Paper at https://arxi…

Dean Valentine built about the simplest alignment probe you can build and the frontier models failed it. Write-up at https://www.lesswrong.com/posts/munJKF7iWMsWJLAH2/astra-and-fab…

A benchmark called Real-SWE went up at https://withspecific.com/real-swe and the headline number is going to annoy a lot of people: the best model on the board resolves 38.8 percen…

Somebody replayed 68,266 real requests from 393 Claude Code sessions plus 23,608 Mooncake requests through a prefix-cache simulator, tried three separate ways to beat plain LRU evi…