My opinions only here. 👨🔬 RS DeepMind 1.9y, Midjourney 1y 🧑🎓 DPhil AIMS 4.5y 🧙♂️ RE DeepMind 1y 📺 SWE Google 3y 🎓 TUM 👤 @nwspk
AI
OpenAI calls the result the first autonomous AI solution to a prominent open problem in mathematics.
AI
Levent Alpöge says their work also resolves the generalized hypothesis, but the available reporting doesn’t establish independent verification.
AI
@osanseviero announced a November 20, 2026 sabbatical after leading GDM DevX, with plans to unplug, travel and build.
AI
Astra’s standard API input and output prices fall 20% versus GPT-6 Astra, but coding results are mixed.
AI
She’ll lead a planned weekly AI newsletter, help shape technology editorials and make podcast appearances, including on “Hard Fork.”
AI
@osanseviero reports running quantized Gemma 3 270m as one PostScript job inside a printer, with 14MB peak RAM use.
AI
According to @andrewwhite01, a scientific-writing model began looking human to Pangram 4.1 during training.
AI
OpenAI says Goodfire will analyze model internals to help monitor training, debug models and improve alignment.
AI
The team reports beating standard distillation at the same training-step and token budget by starting with perturbed inputs.
AI
The method nearly matches on-policy self-distillation accuracy in reported tests, but faster training comes with a two-GPU catch.
AI
kimmonismus praised purported Google “Krypton” leaks but couldn’t verify them; BlackHC separately hoped for a release.
AI
The GPU-based design compresses weight changes before trainers send updates to samplers.
AI
OpenAI attributes the expansion partly to cybersecurity evaluations where its misalignment monitors were not running.
AI
A user says they duplicated a puzzling result with two queries about a minute apart that differed by one digit.
AI
An AI-assisted audit site lists 2,035 alleged problems across 138 papers; Tyler Cowen urged a reassessment.
AI
A developer credits retrieval-augmented n-gram hints for the claimed 83.2% cut in time.
AI
A post says Carbon scored 88.2% on SEC-bench Pro and is rolling out first to cybersecurity specialists.
AI
OpenAI says its claimed refutation of the Exponential Time Hypothesis was formalized in Lean and made public.
AI
A contributor says the merged approach will become Prime-RL’s default algorithm in the next release.
AI
A user quotes a “neolab” friend saying it could resell compute even if it fails to train the best model.