Loose posts

OpenAI warns that recursive self-improvement could accelerate beyond our collective ability to understand progress, assess risks and maintain meaningful human oversight.

Both models passed nearly all of HAProxy's tests, but Opus 5.5 finished in 9.5 hours at 51% lower cost, and the author calls it his daily driver.

Qwen-Image-2.1 ranks first among open models with 1228 points and 17th overall, 4 points behind nano-banana-pro.

Harvard researchers found most grad school applicants now let AI write their essays, but those who did were admitted less often, in a paper titled "AI-written admissions essays are widespread but penalized."

Aimed at agents pulling stale context, the AML benchmark runs 150 software tasks each with relevant history and again with noisy history to test memory retrieval.

DualSQL splits Text-to-SQL into one agent that links the question to the right tables and columns and another that writes the SQL, training both with multi-agent RL.

Sebastian Raschka notes MiMo-V2.6, built on classic grouped query attention with a 128-token sliding window, is currently No.1 on the open-weight benchmarks by weighted average.

Nathan Lambert talks with Epoch AI's DataGen about predictions for recursive self-improvement, robotics in AI acceleration, how far behind Chinese models are, and the role of distillation.

Bloomberg reports Harvey's costs for renting OpenAI and Anthropic models rose sharply, with gross margin falling from about 50% to -50% by June after agent usage grew 20x.

The author says cheap image generation no longer looks cheap, citing Hunyuan 3.5 on OnSolo at 1 credit for 2K output, and argues the price/quality gap is collapsing fast.

Stanford researchers, including Myra Cheng and Dan Jurafsky, traced how people hand their judgment over to AI, and found most of them have no idea they are doing it.

Xiaomi MiMo completed its largest reinforcement learning scaling run so far and released a technical report on scaling compute, environments and grading together.