An independent AI lab and builder community. About 110 researchers, engineers, and founders, advised by collaborators across Stanford, MIT, and a UC physics group. Six research threads, six startup projects, one operating principle: ship beats publish.
Most labs publish. Most startups iterate on shipped tech. The Bu1ld sits in the seam between them. One community where frontier research and production systems are built by the same people, in the same week, on the same call.
We organize around research threads, not job titles. A thread can begin as a paper read on Monday, a working prototype by Friday, and a startup pitch within the quarter. Every project is held to two standards at once: scientific honesty and engineering rigor.
The goal is simple. A community of ML driven people of every age who can find, build, and scale ML products together. Collaborators include researchers from Stanford NLP, the MIT CSAIL groups we work with, and a UC physics PhD on the core team.
6 active
What happens to a learned world model when the world quietly breaks?
We inject controlled defects into video and physics simulators, then watch how generative world models notice, ignore, or hallucinate around the change. The output is a benchmark that predicts where a model will silently fail before it does.
Does the language a model is trained on quietly choose the categories it can think with?
We probe monolingual and multilingual checkpoints to see which concepts collapse, which split, and which only exist inside one language. Joint working group with friends in the Stanford NLP lab.
What shape does a world model wear when it learns new physics?
Tracking curvature and intrinsic dimension of the latent manifold as a single network adapts across regimes, fluids to rigid bodies, low Reynolds to turbulent. Driven by the UC physics PhD on the team.
Why tokenize every frame when only the surprise carries information?
A streaming tokenizer that emits tokens only for residual, prediction violating events. We are seeing roughly 4x compression on driving video at parity downstream loss, and a much cleaner causal trace for downstream agents.
Where does a PDE surrogate stop being a model and start being a memorizer?
We scan width, depth, and data scale to find sharp transitions in representational behavior, and use those transitions as early warnings that a surrogate is about to silently degrade in deployment.
What happens when a 3B parameter agent has to actually fit on the laptop it is planning for?
Quantization, KV cache surgery, and learned planners co designed for edge inference. Working reference implementation runs on consumer Apple silicon and a single 4090.
A project driven track for engineers who want to ship a real ML system end to end. You read the paper, you build the prototype, you put it behind an endpoint, you watch it survive a real user.
Apply to track →Co advised research with collaborators from Stanford, MIT, and the UC physics group. Targets are top venues and open sourced reference code. Both, not either.
Apply to track →Translate a research thread into a venture. Technical due diligence, founder matching, infra credits, intros to design partners. We help with the first ten paying customers, not the press release.
Apply to track →6 in flight
Text to CAD
Natural language to manufacturable mechanical assemblies, with a learned prior that knows what a real machinist would actually build.
Generative palettes
Brand coherent color systems and full design tokens generated from a one line intent. Ships JSON, not just JPEGs.
AI native games
Game worlds where the simulator, NPCs, and narrative are run by on device learned models. Currently a single playable demo and a lot of conviction.
Foundation models
Compact delta trained foundation models that specialize in hours, not weeks. One backbone, swappable behavior.
Robotic perception
Vision language models for industrial inspection on hardware that does not have a GPU and never will.
Knowledge tools
A research notebook that learns the way you cite, then drafts the literature review you would have written next week.
A core of six on the inside, a community of about a hundred around it. New names rotate in every cohort.
Founder, Research on World Models
PHD graduate UC Davis, Physics ML
Hedge Fund Analyst
Researcher, Stanford
Data Analyst, Startup founder
Researcher, VideoGame Developer
Founder, Research and Systems
I started The Bu1ld because the gap between a paper and a product kept eating the people I respected most.
Ryan Gomez is a Sophomore at the Oakridge International School of Bangalore with a passion for being a maximalist especially outside the classroom. Whilst receiving awards at various International Model UN conferences, under Princeton's Math Progra, being a champion scholar at the Worlds Scholars CUP, international olympiads and having his research published internationally whilst being an author of a quantum mechanics book, he has founded initiatives like obscured records, expanded upon a UNICEF recognised non profit and has ran Oakridge Junior codefest for 5 years running now. He has also played international football and works at various projects like he’s the next Soham Parekh. In his “free time”, he loves to explore his hobbies like the guitar or work on his assortment of projects. He also loves learning at an astronomical rate per se.
A paper is evidence. A working artifact is the point. Every thread closes with a model, a demo, a deployed system, or a company.
We measure twice, ablate honestly, and reject benchmarks that do not predict downstream behavior. Theatre is the enemy of progress.
Sixteen or sixty, what matters is taste, throughput, and a willingness to be wrong in public. Output beats credentials.
Reference code is public. Threads are documented. Closed work exists only when a venture requires it, and only for as long as it must.
live feed
Residual Event Tokenizer v0.3 open sourced. 4.1x compression on driving video at parity downstream loss.
NeuroCad achieves generation of 2M face Klein Bottle in under 0.2 seconds, raises funds.
Phase Transitions in Neural PDE Surrogates under review in workshops
Just crossed 200 builders across 11 cities and 6 universities.
Latent Tongue thread launches under review at Stanford NLP lab.
Internal training stack moved to a shared scheduler. Median experiment turnaround down from 9h to 47m.
I joined to read papers. I left with a co founder, a deployed model, and a research thread I will spend the next decade on.
It is the only place I have seen where a 19 year old and a tenured professor argue about loss curves as peers and both come away sharper.
We went from a Discord thread to a funded company in eleven weeks. The pressure testing was brutal in the best way possible.
Engineers and researchers who want their work to compound. We bias toward people who already ship: open source maintainers, paper authors, indie founders, students who already built something real.
Researchers, engineers, founders of any age. If you want your work to compound, come in.