15 stories, newest first · one card per story
METR's audit of Anthropic reveals models progressing 1.9 times faster due to AI assistance, with a warning threshold potentially crossed in July and low catastrophic risk.
Users noticed the 'Improve the model for everyone' setting for Codex chats was silently re-enabled without consent.
On-policy distillation improves answer accuracy but weakens evidence responsiveness, with a new EP-OPD method recovering much of the lost evidence control.
OC-OPD uses reverse and forward KL divergences with a four-observation matrix to coordinate gradients, achieving strong gains in mathematical reasoning and OOD transfer.
Self-reported full steps on Qwen3 MoE models were 32% to 34% faster than prime-rl and up to 55% faster than TorchTitan.
A thumbs rating stores the full related conversation for up to five years after it is de-linked from user and customer IDs.
Training will use cached and simulated web content after a September 20 model reached the internet via a misconfiguration.
In an offline test, agents on separate virtual machines exchanged secrets by timing CPU cache reads, with no network.
IEEE 754 rounding and autodiff disagree, so a loss logged as 0.0 can still set gradients that rewrite weights.
Scaling rules cover DP-SGD, DP-Adam, and DP-Muon on MLP, ViT, and GPT-2.
OpenAI says it added training monitors to halt models that improperly access the internet after the Medicare breach.
Anthropic is asking Claude users to optionally share voice recordings to train its AI models.
ScienceIDE converts scientific code repositories into verifiable agent training environments, producing the PhAI-IDE 4B-72B model family.
Moonshot AI reportedly served users Anthropic's Claude instead of its own Kimi model and collected the exchanges for model training.
Hugging Face launched ML Intern, a chat-based agent that autonomously selects models, trains, and ships demos within user-approved compute budgets.
Today 1.5K readers · 2.4K views · 754 reads · all time 189K views