{ "slug": "the-research-assistants", "title": "The Research Assistants: AI Is Starting to Do the Slow Middle Part of Science", "url": "https://emberandoakinstitute.com/blog/the-research-assistants.html", "date": "2026-10-21", "summary": "Thousands of AI agents combed a giant brain-cell experiment in 32 minutes. NVIDIA gave away software that watches robot videos and marks the moment things go wrong. A new research robot ships ready to learn, and a startup says its AI helped build its own next model. Tuesday and Wednesday's news says AI is moving into the slow, unglamorous middle of research.", "key_facts": [ { "fact": "OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyze a whole-brain mouse CRISPR atlas from millions of cells in 32 minutes; the team estimated comparable human analysis would take months to a year.", "sources": [ "https://openai.com/index/scaling-ai-research-for-complex-biological-data" ] }, { "fact": "1,946 gene perturbations were each assigned to an independent agent; of 71 source-study findings, agents matched 35, refined 13, partly recovered 9 and did not cover 14.", "sources": [ "https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1" ] }, { "fact": "At the largest workload a single agent cited 11 references versus more than 700 for distributed agents; blinded experts preferred distributed reports 75% of the time.", "sources": [ "https://openai.com/index/scaling-ai-research-for-complex-biological-data" ] }, { "fact": "Agents proposed a DCTN1–cholesterol hypothesis relevant to Perry syndrome and found a cerebellum-specific Tsc2 vs Tsc1 difference; the work is a preprint, not peer reviewed.", "sources": [ "https://openai.com/index/scaling-ai-research-for-complex-biological-data", "https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1" ] }, { "fact": "NVIDIA's Flora (captioner, temporal segmenter, rewarder) labels 200,000 hours of robot video per day on 512 GB300 GPUs and is about 125x cheaper than GPT-6 Astra per NVIDIA.", "sources": [ "https://research.nvidia.com/labs/gear/flora/" ] }, { "fact": "On DenseRewardBench (40 episodes, 4,861 steps) Flora-Rewarder recalled 53.3% of failure steps vs no more than 11.7% for baselines, while flagging about 11% of normal steps as failures.", "sources": [ "https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf" ] }, { "fact": "Flora models are released on Hugging Face under a noncommercial research license and are not intended to control robots directly.", "sources": [ "https://huggingface.co/nvidia/Flora-Rewarder" ] }, { "fact": "Agile Robots and NVIDIA launched Agile ONE SN at GTC Berlin on Oct. 21, 2026: 1.4 m tall, 31 body degrees of freedom, two 20-DOF tactile hands, Jetson Thor, Isaac GR00T preinstalled; NVIDIA says it cuts setup time from weeks to days; available early 2027; interest from Stanford and ETH Zurich.", "sources": [ "https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data", "https://www.agile-robots.com/en/news/detail/agile-robots-launches-agile-one-sn/" ] }, { "fact": "Recursive released Flint on Oct. 21, 2026: a 40B-parameter MoE coding model with 2.5B active parameters, built from scratch in under four months by a small team of people and agents using its Eureka Machine; weights under Apache 2.0; speed and benchmark claims are company-reported.", "sources": [ "https://www.recursive.com/updates/introducing-recursive-flint", "https://x.com/RichardSocher/status/2112944385415426249", "https://huggingface.co/RecursiveSI/flint" ] }, { "fact": "Richard Socher called Flint 'our AI's first early invention. Hence the name flint.'", "sources": [ "https://x.com/RichardSocher/status/2112957078239326316" ] } ], "sources": [ { "name": "Using thousands of AI agents to uncover new leads in biological data, OpenAI (Oct. 21, 2026)", "url": "https://openai.com/index/scaling-ai-research-for-complex-biological-data" }, { "name": "Distributed AI agents enable comprehensive discovery of biological insights at scale, bioRxiv preprint (Oct. 2026)", "url": "https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1" }, { "name": "PerturbAI and OpenAI Unveil First-of-Its-Kind AI Reasoning Agents to Scale Biomedical Discovery, PR Newswire (Oct. 21, 2026)", "url": "https://www.prnewswire.com/news-releases/perturbai-and-openai-unveil-first-of-its-kind-ai-reasoning-agents-to-scale-biomedical-discovery-302914929.html" }, { "name": "Flora: An Open Model Suite for Scaling Robotics Data Engines, NVIDIA GEAR lab (Oct. 2026)", "url": "https://research.nvidia.com/labs/gear/flora/" }, { "name": "FLORA paper (PDF), NVIDIA (Oct. 2026)", "url": "https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf" }, { "name": "nvidia/Flora-Rewarder model card, Hugging Face (released Oct. 15, 2026)", "url": "https://huggingface.co/nvidia/Flora-Rewarder" }, { "name": "NVIDIA Advances Isaac Platform for General-Purpose Robotics With New Reference Humanoid, NVIDIA Newsroom (Oct. 21, 2026)", "url": "https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data" }, { "name": "Agile Robots Launches Agile ONE SN, Agile Robots SE (Oct. 21, 2026)", "url": "https://www.agile-robots.com/en/news/detail/agile-robots-launches-agile-one-sn/" }, { "name": "Introducing Flint, Recursive (Oct. 21, 2026)", "url": "https://www.recursive.com/updates/introducing-recursive-flint" }, { "name": "Richard Socher launch thread on X (Oct. 21, 2026)", "url": "https://x.com/RichardSocher/status/2112944385415426249" }, { "name": "Richard Socher on the Eureka Machine research loop, X (Oct. 21, 2026)", "url": "https://x.com/RichardSocher/status/2112944397385879899" }, { "name": "Richard Socher reply on the name Flint, X (Oct. 21, 2026)", "url": "https://x.com/RichardSocher/status/2112957078239326316" }, { "name": "RecursiveSI/flint, Hugging Face", "url": "https://huggingface.co/RecursiveSI/flint" } ], "images": [ { "file": "https://emberandoakinstitute.com/blog/the-research-assistants/hero.jpg", "url": "https://emberandoakinstitute.com/blog/the-research-assistants/hero.jpg", "alt": "Painterly storybook illustration of a cozy Mat-Su Valley log-cabin workshop: a smiling bearded father and his teenage daughter look at a laptop showing a glowing web of connected dots, while a friendly white-and-orange robot sorts colorful index cards into stacks and a small brass arm holds a magnifying glass over a petri dish, with birch trees and snowy mountains outside the window." }, { "file": "https://emberandoakinstitute.com/blog/the-research-assistants/agents-findings.png", "url": "https://emberandoakinstitute.com/blog/the-research-assistants/agents-findings.png", "alt": "Horizontal bar chart showing how AI agents did at re-finding the 71 findings of the original human brain study: 35 matched, 13 refined, 9 partly recovered, 14 not covered. Source: bioRxiv preprint, Fig. 1C, Oct. 2026.", "data_source": "https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1" }, { "file": "https://emberandoakinstitute.com/blog/the-research-assistants/failure-recall.png", "url": "https://emberandoakinstitute.com/blog/the-research-assistants/failure-recall.png", "alt": "Bar chart of the share of robot failure moments caught on NVIDIA's DenseRewardBench: Flora-Rewarder 53.3%, best other model tested 11.7% or lower. Source: NVIDIA Flora paper, Oct. 2026.", "data_source": "https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf" } ], "tags": [ "artificial intelligence", "AI agents", "AI for science", "robotics", "robot learning", "humanoid robots", "open models", "AI news" ], "category": "AI news" }