# The Research Assistants: AI Is Starting to Do the Slow Middle Part of Science

*Ember & Oak Institute, LLC · Oct. 21, 2026*

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.

![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.](https://emberandoakinstitute.com/blog/the-research-assistants/hero.jpg)

For the scientists who built it, a map of how gene changes ripple through a mouse brain was the product of a huge experiment. On Wednesday, OpenAI said thousands of AI agents read through that data in 32 minutes, work the team estimated would take people months to a year ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)).

The agents did not just summarize. Working with the biotech startup PerturbAI and Scripps Research, OpenAI says they turned up new, testable ideas about brain disease, ones that blinded human experts judged worth following up ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)).

It was one of four stories in the past two days with the same shape. NVIDIA's newly released free research software, which drew wide attention on X this week, labels robot training videos and marks the moments a robot fails ([NVIDIA](https://research.nvidia.com/labs/gear/flora/)). NVIDIA and the German company Agile Robots unveiled a small humanoid robot meant to arrive in research labs ready to learn ([NVIDIA](https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data)). And the startup Recursive released a coding model it says was built in under four months with heavy help from its own AI research system ([Recursive](https://www.recursive.com/updates/introducing-recursive-flint)).

Most AI headlines are about the moment of discovery or the finished product. This week's news was about the slow middle part of research: reading the data, labeling it, setting up the equipment and running experiment after experiment. That is where AI is starting to take on real work, while people still decide which ideas are worth testing.

## Thousands of agents, one experiment

Modern biology experiments can switch off thousands of genes and measure what happens in millions of cells. "Modern experiments generate far more data than scientists can thoroughly investigate," OpenAI wrote ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)).

The team's answer was to split the job up. The study's preprint says 1,946 gene disruptions from a whole-brain mouse atlas were each handed to their own agent ([bioRxiv](https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1)). Each agent studied how cells responded, searched the scientific literature and proposed explanations. Their reports were then combined into what the team calls a "reasoning atlas" ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)).

To check the agents' work, the researchers hid the original study's conclusions from them and then compared. Of the 71 findings in the human study, the agents matched 35, refined 13, partly recovered 9 and did not cover 14 ([bioRxiv](https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1)).

![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.](https://emberandoakinstitute.com/blog/the-research-assistants/agents-findings.png)

*The agents re-found or improved on most of the original human team's findings, but missed 14 of 71. The study's authors withheld those conclusions from the agents before the test. Source: [bioRxiv preprint](https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1), Fig. 1C. Chart: Ember & Oak Institute, LLC.*

Splitting the work mattered. When the team gave one agent the whole job instead, it searched less of the literature as the workload grew, citing only 11 references at the largest size compared with more than 700 for the distributed agents. In blinded comparisons, experts preferred the distributed agents' reports 75% of the time ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)).

One lead involves DCTN1, a gene tied to Perry syndrome, a rare disease with Parkinson-like symptoms. The agents noticed that disrupting it switched on cholesterol-making genes, and connected that to earlier research on cholesterol transport ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)). "Sometimes the clues are already there, scattered across different fields — we just haven't connected them," said Tomasz Nowakowski, a neurosurgery researcher at UCSF ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)).

It is still a hypothesis, not a treatment. OpenAI frames the findings as leads that "help prioritize further experiments," and adds: "Scientists remain at the center of discovery" ([OpenAI](https://openai.com/index/scaling-ai-research-for-complex-biological-data)). The paper is a preprint, which means it has not yet been peer reviewed ([bioRxiv](https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1)).

## Teaching robots by grading their homework

Robots have a similar bottleneck. To learn from video, robot AI needs that video labeled: what happened, when, and whether it went well. NVIDIA's new Flora suite does that with three models that cut videos into steps, write a caption for each step and score how well the robot did ([NVIDIA](https://research.nvidia.com/labs/gear/flora/)).

NVIDIA says Flora is the same system it uses to train its own GR00T robot models, and that it labels 200,000 hours of video a day across 512 GPUs ([NVIDIA](https://research.nvidia.com/labs/gear/flora/)). The company says it is roughly 125 times cheaper than using OpenAI's GPT-6 Astra for the job ([NVIDIA](https://research.nvidia.com/labs/gear/flora/)).

The most striking number is about mistakes. On a test of 40 two-armed robot recordings, Flora's grading model caught 53.3% of the moments when the robot acted poorly or failed. None of the other models tested caught more than 11.7% ([NVIDIA paper](https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf)). NVIDIA is upfront about the trade-off: Flora also wrongly flagged about 11% of normal moments as failures ([NVIDIA paper](https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf)).

![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.](https://emberandoakinstitute.com/blog/the-research-assistants/failure-recall.png)

*Catching failures is how a robot learns what not to do. Flora found far more of them than other models in NVIDIA's test, at the cost of some false alarms. Source: [NVIDIA Flora paper](https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf), DenseRewardBench. Chart: Ember & Oak Institute, LLC.*

The models are free to download, but only for noncommercial research, and NVIDIA says Flora's output is labels, not something meant to control a robot directly ([Hugging Face](https://huggingface.co/nvidia/Flora-Rewarder)).

## A lab robot that arrives ready to learn

Hardware takes time too. At NVIDIA's GTC conference in Berlin on Wednesday, Agile Robots launched the Agile ONE SN, a 1.4-meter-tall humanoid built for research labs ([Agile Robots](https://www.agile-robots.com/en/news/detail/agile-robots-launches-agile-one-sn/)). Its body has 31 degrees of freedom, meaning 31 separate ways its joints can move, plus two five-fingered hands with touch sensing, and it comes with NVIDIA's robot software preinstalled ([NVIDIA](https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data)).

The pitch is time saved. Researchers often have to stitch together separate tools for simulation, training and testing, and NVIDIA says the Agile ONE SN puts them on one platform, "cutting setup time from weeks to days" ([NVIDIA](https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data)). It is scheduled to be available in early 2027, with interest from Stanford University and ETH Zurich ([NVIDIA](https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data)). Agile Robots says a common platform also "makes it easier to share and compare research across labs" ([Agile Robots](https://www.agile-robots.com/en/news/detail/agile-robots-launches-agile-one-sn/)).

## An AI that helped build the next AI

The last story turns the idea back on AI itself. Recursive, a San Francisco startup led by AI researcher Richard Socher, released Flint, a coding model designed to run on a personal computer ([Recursive](https://www.recursive.com/updates/introducing-recursive-flint)). The company says a small team of "people and agents" built it from scratch in under four months ([Recursive](https://www.recursive.com/updates/introducing-recursive-flint)).

The agents belong to what Recursive calls its Eureka Machine. The company says it runs the full research loop: "proposing ideas, implementing them, testing them, and using the results to decide what to try next" ([Recursive](https://www.recursive.com/updates/introducing-recursive-flint)). Recursive says the system helped with nearly every part of Flint, from training recipes to the code that makes it run fast on laptops. In one case it shrank part of the model in a way the company describes as "against common practice" and then found a way to recover the lost accuracy ([Recursive](https://www.recursive.com/updates/introducing-recursive-flint)).

Replying to another X user, Socher wrote that the model is "our AI's first early invention. Hence the name flint" ([Richard Socher on X](https://x.com/RichardSocher/status/2112957078239326316)). The speed and benchmark claims are the company's own and have not yet been independently tested. Recursive also notes Flint "may need more guidance" than larger commercial coding models ([Recursive](https://www.recursive.com/updates/introducing-recursive-flint)). Its files are free to download under an open license ([Hugging Face](https://huggingface.co/RecursiveSI/flint)).

## What it adds up to

None of these four stories is a cure, a breakthrough robot or a finished science machine. Together they show something quieter. AI is taking on the patient, repetitive middle of research: reading every result instead of the strongest few, grading thousands of robot attempts, cutting weeks of setup and running experiment after experiment.

The caveats matter as much as the speed. The brain-study agents missed 14 of 71 known findings. Flora raises false alarms. The new robot isn't in labs yet, and Flint's numbers come from the company that made it. In every case, people still choose the questions and check the answers.

That loop will sound familiar to Ember & Oak Institute, LLC learners. Recursive's description of its Eureka Machine, propose an idea, build it, test it and use the results to decide what's next, is nearly the same idea, build, test, improve cycle students use when they direct an AI to make a project. The difference is scale: these systems run that loop thousands of times while a person sleeps.

For Valley families, the useful takeaway is that being good at directing AI increasingly means being good at asking the right question and checking the result, not just getting an answer fast.

**Try it at home:** Pick a pile of something your learner has collected, such as rocks, leaves or game scores. Split it into small groups and have your learner ask an AI helper to suggest one pattern for each group. Then test each suggestion by looking for yourselves, and keep only the ideas that hold up. You will have run a small version of this week's experiment, with a human as the final judge.

## Sources

- [Using thousands of AI agents to uncover new leads in biological data, OpenAI (Oct. 21, 2026)](https://openai.com/index/scaling-ai-research-for-complex-biological-data)
- [Distributed AI agents enable comprehensive discovery of biological insights at scale, bioRxiv preprint (Oct. 2026)](https://www.biorxiv.org/content/10.64898/2026.10.15.759436v1)
- [PerturbAI and OpenAI Unveil First-of-Its-Kind AI Reasoning Agents to Scale Biomedical Discovery, PR Newswire (Oct. 21, 2026)](https://www.prnewswire.com/news-releases/perturbai-and-openai-unveil-first-of-its-kind-ai-reasoning-agents-to-scale-biomedical-discovery-302914929.html)
- [Flora: An Open Model Suite for Scaling Robotics Data Engines, NVIDIA GEAR lab (Oct. 2026)](https://research.nvidia.com/labs/gear/flora/)
- [FLORA paper (PDF), NVIDIA (Oct. 2026)](https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf)
- [nvidia/Flora-Rewarder model card, Hugging Face (released Oct. 15, 2026)](https://huggingface.co/nvidia/Flora-Rewarder)
- [NVIDIA Advances Isaac Platform for General-Purpose Robotics With New Reference Humanoid, NVIDIA Newsroom (Oct. 21, 2026)](https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data)
- [Agile Robots Launches Agile ONE SN, Agile Robots SE (Oct. 21, 2026)](https://www.agile-robots.com/en/news/detail/agile-robots-launches-agile-one-sn/)
- [Introducing Flint, Recursive (Oct. 21, 2026)](https://www.recursive.com/updates/introducing-recursive-flint)
- [Richard Socher launch thread on X (Oct. 21, 2026)](https://x.com/RichardSocher/status/2112944385415426249)
- [Richard Socher on the Eureka Machine research loop, X (Oct. 21, 2026)](https://x.com/RichardSocher/status/2112944397385879899)
- [Richard Socher reply on the name Flint, X (Oct. 21, 2026)](https://x.com/RichardSocher/status/2112957078239326316)
- [RecursiveSI/flint, Hugging Face](https://huggingface.co/RecursiveSI/flint)