Wisconsin Researchers Build a Scaling Law That Makes Synthetic Rewrites Worth the Compute
A new scaling law predicts the optimal mix of synthetic rewrites in pretraining, and shows a 22B model can outperform its own rewriter at a tenth the compute.
20 stories
A new scaling law predicts the optimal mix of synthetic rewrites in pretraining, and shows a 22B model can outperform its own rewriter at a tenth the compute.
A new theoretical paper proves Anthropic's Jacobian lens isolates slow, context-stable features that grow across layers, reframing what its 'workspace' really measures.
New research shows hidden model preferences slip through distillation even when teacher and student share no base model, data, or architecture.
Physicists built a context-sensitive grammar where only nearby symbols interact, and still found a phase transition, hinting that language itself drives the order.
Thinking Machines Lab's 276B-parameter MoE matches its 975B flagship across reasoning, coding, and multimodal tasks while using 12B active parameters per token.
Anthropic ships its first versioned Claude API release since 2023, dropping temperature controls, legacy fields, and the old X-Api-Key header.
Nous Research's open-source Hermes Agent anchored Jensen Huang's GTC Berlin finale, orchestrating a full sim-to-real training pipeline for a Microduck robot on an RTX Spark laptop.
An independent developer vibe-coded a from-scratch reproduction of the SuperMimic humanoid control paper in roughly an hour and hit the paper's numbers in 1.5 GPU-hours.
OpenAI is placing its frontier models and Codex agent workflows inside Google Cloud's Model Garden, making its biggest search rival a sales channel.
Google completes its open 32-bit Coral NPU RISC-V design, ships a v1.0 MLIR compiler, and lands commercial Synaptics silicon hitting 1 TOPS under 10 mW.
Microsoft Research and GitHub Next release the Proof-Driven Development Kit in research preview, pairing LLM agents with Verus to formally verify Rust code against specs.
Anthropic opens a waitlist for Claude Duet, a proactive agent built on Opus 5.5 that monitors your apps and messages you only when something needs you.
Perplexity and NVIDIA introduce JOLT, a self-distillation post-training method that matches GRPO accuracy on MATH using 6.5x fewer training tokens.
AMD and Meta's new FlyDSL backend plugs into PyTorch on MI350-series GPUs, delivering up to 3.01x end-to-end speedup for vLLM inference workloads.
Anthropic's code vulnerability scanner now runs continuously on connected repos, tracks fixes automatically, and plugs into dev workflows via API and MCP.
HeyGen's new Turbo text-to-speech model takes the top blind-test ranking at $4 per million characters, undercutting its own predecessor by 7.5x.
A new continuous language model called LAW ditches discrete token sampling between steps, propagating expected embeddings instead and outperforming diffusion baselines on reasoning tasks.
A year-old research lab led by the creators of ThunderKittens just raised $1B at a $6B valuation to rebuild how GPUs run AI workloads.
Keenable's search API paired with GPT-5.6 Sol takes the top spot on the Vals Web Search Index at 49.1%, leading finance research.
Microsoft researchers used human-directed AI agents to prove a 31-year-old geometry conjecture at the heart of randomized sampling algorithms.