Home › Latest AI research
New AI, tech and chip research in plain words, with a link to every source, plus the people worth following.
Agents · Hugging Face Daily Papers · 14 Oct
New studynot peer reviewed yet
FDE-Gym: The Last Mile of AI Delivery
A new study says AI agents do much worse on realistic customer jobs, where the request is vague and hidden rules must be discovered, than on clean tests.
Read the sourceAgents · Hugging Face Daily Papers · 14 Oct
New studynot peer reviewed yet
LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents
A new study says a cheap AI helper can do most computer tasks and call a stronger one only when unsure, which cuts cost.
Read the sourceHow AI works · Hugging Face Daily Papers · 14 Oct
New studynot peer reviewed yet
Data World: Scaling Data Environments from Scratch
A new study says AI can build its own practice data tasks to get better at finding insights in data, without copying a bigger model.
Read the sourceHow AI works · Hugging Face Daily Papers · 12 Oct
New studynot peer reviewed yet
A new study says a trick can make AI handle long documents with less memory and compute while keeping answer quality.
Read the sourceAgents · Google Research · 16 Oct
Grounding AI agents in biological reality: Building blocks for autonomous wet labs
Read the sourceScience and health · Microsoft Research · 15 Oct
Proof-based general-purpose computing with Magnetar
Read the sourceChips and speed · Apple Machine Learning Research · 15 Oct
Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
Read the source4 picks, each explained in plain words and linked to its source. Older days are in the drawers below.
New studynot peer reviewed yet
A new study says AI agents do much worse on realistic customer jobs, where the request is vague and hidden rules must be discovered, than on clean tests.
Hugging Face Daily Papers (285 upvotes) ·
Read moreNew studynot peer reviewed yet
A new study says a cheap AI helper can do most computer tasks and call a stronger one only when unsure, which cuts cost.
Hugging Face Daily Papers (8 upvotes) ·
Read moreNew studynot peer reviewed yet
A new study says AI can build its own practice data tasks to get better at finding insights in data, without copying a bigger model.
Hugging Face Daily Papers (28 upvotes) ·
Read moreNew studynot peer reviewed yet
A new study says a trick can make AI handle long documents with less memory and compute while keeping answer quality.
Hugging Face Daily Papers (10 upvotes) ·
Read moreExplainer
Epoch AI says the best free-to-download model, Z.ai's GLM-5.3, matches where paid models stood 6 to 7 months ago on its cyber test.
Epoch AI ·
Read moreApple Machine Learning Research
Apple shows how several Mac Studios can team up to run very large AI models at home.
Apple Machine Learning Research ·
Read moreNew studynot peer reviewed yet
A new study says the building blocks we use to peek inside AI models only partly combine like Lego; steering the model needs extra pieces.
Hugging Face Daily Papers (104 upvotes) ·
Read moreAllen Institute for AI
Ai2 says its AI weather model captures extreme rain and wind at about 1/300th the cost of the usual physics simulator.
Allen Institute for AI ·
Read moreNew studynot peer reviewed yet
A new study says a cheap AI agent can do the job and call in a stronger one only when unsure, saving cost.
Hugging Face Daily Papers ·
Read moreHugging Face Daily Papers, 19 Oct. All new studies, not peer reviewed yet. Number = upvotes.
The newest post from each research lab blog we follow.
Google Research
Grounding AI agents in biological reality: Building blocks for autonomous wet labsMicrosoft Research
Proof-based general-purpose computing with MagnetarApple Machine Learning Research
Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio30 people worth following for new research, with links to their own posts where we have them.
People doing the research, many with their own blogs.
People who read new papers and explain what they mean.
We count visits with Google Analytics to see what helps. OK with that?
ITG Now