USC Researchers Reveal Only 15% of NIH AI Grants Reach Clinical Use
A USC team used GPT-4o-mini to classify 58,746 NIH grants, finding AI gets a 13.4% funding premium but barely reaches clinics.
20 stories
A USC team used GPT-4o-mini to classify 58,746 NIH grants, finding AI gets a 13.4% funding premium but barely reaches clinics.
OpenAI built Pando, a Git-compatible backend that cut push latency ~80% and now moves nearly a petabyte of Git traffic daily across its monorepo.
Researchers from Prefrontal AI and ETH Zurich introduce a hierarchical clustering trick that keeps attention accurate at 10 million tokens while slashing compute.
Berkeley researchers give humanoid robots a force command channel, letting a Unitree G1 drag, push and kick 20 kg loads that a motion-only tracker cannot budge.
A 56-DOF mobile bimanual humanoid with Vision Pro teleop, built from $8,056 in parts, with every design file and training pipeline released open-source.
Snorkel AI's RIFT framework scores the rubrics behind benchmark scores, finding that half of 147 public benchmark tasks have missing or misaligned criteria.
SGLang's HiCache now extends KV prefix reuse across context parallelism, pipeline parallelism, speculative decoding, and prefill-decode disaggregation serving.
Anthropic opens a paid residency paying up to $500K for senior technical talent to pivot into AI safety research, no ML background required.
Anthropic's Messages API now caches conversations server side, cutting payload sizes from 800KB to 1.8KB and median time to first token by 35% on long sessions.
Exa's Monitors v2 turns scheduled web search into an event stream for agents, cutting prices 7x and only firing webhooks when something genuinely new appears.
humans& open-sourced a lossless GPU compression pipeline that cuts RL sampler pause time to 0.23s, beating intra-cluster RDMA for weight sync.
A new SO(3)-equivariant architecture lifts the kernel and nonlinearity constraints of spherical CNNs, improving wind forecasts and spherical MNIST tasks.
OpenAI's new LifeVal benchmark measures how well models handle messy, personal errands like finding childcare, meal planning, and dealing with government services.
Meta and NVIDIA fused ad retrieval into one GPU kernel, hitting over 1 billion ads per second per server with ranking-level quality.
Timestral's largest model has 209 million parameters and leads three pretrained, zero-shot forecasting benchmarks. The smallest forecasts 96 steps from 2,048 observations in 17 milliseconds on eight CPU threads.
Suno responds to user complaints about cramped Studio controls with a bigger clip settings sidebar, fine pitch tuning, and better grid controls.
ARTEX is an open-source multi-agent system that autonomously runs penetration tests end-to-end using a dual-graph planner-worker architecture.
Sentence Transformers, the SBERT library behind over 16,000 embedding models on the Hub, moves from UKP Lab at TU Darmstadt to Hugging Face stewardship.
Hume's new Audio Decision API collapses a stack of specialized voice classifiers into one call that answers arbitrary questions about any recording.
Parallel open-sources a 300-question benchmark for decision APIs that must consult the live web, where its own API beats Perplexity 1.5x.