Maryland Researchers Pinpoint Why AI Training Silently Explodes After Weeks
A new study pins long-horizon pretraining failures on BF16 quantization error in the LM head, and fixes it by swapping the optimizer.
20 stories
A new study pins long-horizon pretraining failures on BF16 quantization error in the LM head, and fixes it by swapping the optimizer.
A one-line gradient fix called Soft-MTP strips the noisy residual from multi-token prediction, yielding better perplexity, faster speculative decoding, and no extra cost.
Cartesia's new Managed Agents platform jumps 98 Elo points over its predecessor on blind voice preference tests, with task success climbing to 91.7%.
SYSTRAN's English-only base checkpoint repackages OpenAI Whisper for CTranslate2, delivering roughly 4x faster transcription at the same accuracy on modest hardware.
Anthropic details Claude's work across six biotech partnerships and its own in-house drug programs, with early wins in CAR-T, allergies, and regulatory paperwork.
Vals AI explains how ten Claude agents produced a Lean proof for seven electrons on a sphere, and names which cases should fall next.
Cursor's new Annotate feature lets you point, draw, and type on your running app so the agent can edit the right code.
InclusionAI's updated 560B MoE reasoning model jumps three points on the Intelligence Index while keeping the same price tag and MIT license.
A community fine-tune strips safety refusals from Liquid AI's tri-hybrid LFM3-1.7B and ships it as a 1.1GB GGUF that runs fully offline.
A new research paper derives scaling laws showing that tool and database diversity, not task count, drive tool-use agent generalization.
Arena's Agent Mode now plugs directly into Gmail, GitHub, Linear, Notion, Vercel, Supabase and more, replacing the copy-paste shuffle between tabs.
Google Research proposes a biologically inspired sleep phase that lets LLMs consolidate new facts into weights without forgetting, scaling context to 10M tokens.
Zed's Delta team ships a share command that turns Claude Code sessions into live, multi-developer threads with synced worktrees and uncommitted edits.
Cognition's new coding benchmark tests whether agents respect unstated constraints, with top model Opus 5.5 scoring just 45% on tasks maintainers built over 40+ hours each.
Liquid AI's LFM3 family packs text, vision, audio, omni, and decision models into one 2B backbone built for on-device fine-tuning.
Fish Audio's new streaming speech-to-text claims the top spot on Cekura's leaderboard with 2.21% WER and conversation-aware turn-taking for voice agents.
Reflection AI reveals the raw infrastructure behind Beam's reinforcement learning run and commits to OpenEnv as its future environment standard.
Nous Research bundles always-on Hermes Cloud into every paid tier and swaps monthly credit pools for daily and weekly rate limits.
Google DeepMind's new world model streams photorealistic, playable 3D environments with native audio, live physics, and Gemini-driven reasoning in real time.
DeepVK's new Russian sentence encoder packs 8,192-token context and Matryoshka embeddings into a 149M-parameter model that rivals systems 4x its size.