Exobrain: Automatically turn your life into a dataset and train your own model.

Where I’m coming from

I left OpenAI in June 2026 after joining in October 2020. I was a researcher on Alignment, Human Data, Artificial SWE, and before I left I was leading the Ecosystem research team. I built the internal human data labeling platform still in use at OpenAI, I co-launched ChatGPT memory, I ran the first training runs that leveraged ChatGPT user feedback, among other non-public work. Basically I’ve spent a lot of time on RL against informal human goals.

My next step continues on this path of “loading specific human goals into the computer”.

There are two reasons why this path is worth the pain.

  1. I want to be able to shape the models that I leverage daily as a power user. A basic value I hold is that people should be able to understand and shape the technology they use. AI is more immersive and intimate than anything beforehand, so it is particularly important that end users have control over how it shows up in their life. Continual learning from comprehensive personal data to create personally owned models will yield the best possible user experience.
  2. The world is less safe and less beautiful than it would be if models had better reward signals. We have models with superhuman capability in some narrow domains, but the bottleneck for a lot of value is getting the values and goals of specific humans into the models. Current RLVR methods and datasets do not achieve this. The models are trained to complete tasks at all costs without sufficient human grounding.

Here’s what I’m building

A public benefit corporation called “b2a” (brain2agent). The first product is the “exobrain”.

The exobrain is a product stack that converts your natural workflows and life context into your custom personal model that acts in service of your goals both stated and implicit. The model you own and control should be able to closely observe you over time, infer goals, proactively ask questions, propose high risk actions for approval, and take low risk actions for post hoc review. Concretely, think of a system that records your screen, click and keyboard data, all accessibility data, your webcam for gaze tracking, all documents, messages, and emails you receive, all audio of you talking to your computer, neural prosthesis data in the future, and ultimately all of the data received or produced that you opt to share with the system. Critically, keeping all of it living on hardware you own, with data movement only when you choose. The system then uses models to structure this data, extract goals, and create RL training sets and midtraining sets. The system leverages compute resources to train an open weight model on these training sets in order to create and then continuously update a custom personal model.

This model is meant to help you achieve your goals, it is not a digital twin or a companion. It should know you and work for you, it should not replace you or your friendships.

Here are the ingredients for a custom training dataset:

(workflows, life_context, model_inference, user_oversight) -> custom_training_dataset

Here are the ingredients for a custom personal model:

(custom_training_dataset, open_weights_models, training_system) -> custom_personal_model

Here’s how I’m going to make money

It’s important to be upfront about this and to have a reasonably clear plan to hold myself to, otherwise incentives drift. Here are the revenue channels I’m going to pursue.

  1. Sell a managed service to individuals. Use trusted enclaves for inference and training the model when the user doesn’t have their own compute hardware. Take significant margin on the compute. Individuals should be able to run the whole system e2e without ever paying me a dime if they have their own compute.
  2. Sell a managed service to startups. Provide e2e tooling for data collection, training, and inference to startups. This will include edge-governance so that individual employees can control how their data flows into training.
  3. Enable users to sell their data to labs. Provide users with an edge-governance platform to sell subsets of their data. The edge-governance platform includes an open-source leakage preview that runs locally so users know what they are giving away. I will aggregate, filter, and tranche this data for sale. The labs have demand for long-horizon real user data with full life-context. Individual users should be able to make money from their data if they want to with full knowledge of the privacy tradeoffs. Critically, b2a never sells, deploys, or keeps models derived from marketplace data.
  4. Sell hardware. Produce co-optimized local hardware for inference, training, data collection and storage. Sell this hardware to individuals so they can control their own destiny. This will be the long-term defensible revenue source and is the most aligned, but it will take time to get here.

The soul of the company is in 1 and 4, by contrast 2 and 3 are primarily compromises to make money in order to unlock the capital intensive hardware business. If I can somehow go straight from 1 to 4 and keep the focus on individual users the whole time, then I will.

This means that I will give away the data recording client and the core training stack to individuals as open source. This is necessary, given how sensitive the data is, so that users can inspect the software and trust its behavior. There will be separate licensing rules around the training stack for larger companies, this is partially to make money from larger companies and partially to make sure the system isn’t turned into bossware.

Here are the hard parts

  1. Compute is expensive. Inferencing and training the best Open Weight Models costs a lot. This is both a challenge for research iteration and for making the product available to more than rich technologists. It is important to be clear-eyed about this and constantly hunt for the best compute deals and most efficient models for training and inference.
  2. The data must be high quality. The core research challenge is creating a pipeline that can convert raw interaction patterns into high-quality datasets with the most efficient use of human oversight. In particular, the goal is getting this capability with open weight models as soon as possible so that we can minimize the time window where we need to send personal data to the labs.
  3. Cybersecurity is existential. The plan is to collect and aggregate an enormous amount of very sensitive personal data over a long period of time. It is critically important not to lose this data.
  4. Reliance on Open Weight Models to keep up. If Open Weight Models stagnate or are outlawed, then this enterprise fails. This is reliant on the ecosystem, and so I mean to contribute where I can.
  5. Private data makes open collaboration harder. This project is inherently sensitive, it is about extracting context from a massive amount of very personal information and carefully observing it and iterating on it. There’s no shared data corpus to publish and collaborate on. The privacy challenges are one reason I don’t think the incumbent labs will be able to keep up, they can’t get this data from contractors.

Why now?

  1. Model Capabilities are now near the threshold required for automatic data generation. In order to understand user goals and synthesize realistic environments, we need a basic level of model competence. Additional evals, prompting, and finetuning are certainly needed for success here though.
  2. You can now outperform the frontier in some areas by finetuning an open weight model. Kimi, Qwen, DeepSeek, GLM, Inkling all provide strong starting points and there’s an emerging industry finetuning custom models for businesses with Prime Intellect, Applied Compute, and others.
  3. I now coexist with agents within a single environment, my computer. In the chatbot era, the models were limited in terms of context and actions. Now with terminal-using and computer-using agents, they have the full digital context and action space of my life.
  4. High-quality data is more valuable than ever. Long-horizon, high-context data is the bottleneck for many research programs. b2a will create a new market for this type of data and establish pricing for it.

The bull case

In 2031, anyone who wants can own their own personal model, the compute to inference and train it, and advanced wearables and neural prostheses to steer it and for it to learn from them. They feel the inalienable satisfaction of the craftsman as they shape the AI they use with their own hands and minds. They are able to extend their idiosyncratic genius to create their own art forms and schools of thought. The world sees a flowering instead of a flattening since we have chosen to extend illegible individual minds rather than replace them with a lossy compression.

Inspiring ideas and ideators