Robert Scoble shared a report this weekend from an Anthropic Builder Day. A nicer word for a hackathon, as he puts it himself. Hundreds of builders, from all over the world, spending a day in the Ferry Building in San Francisco. At SHACK15, one of the city’s finest spots, looking out over the bay.
There is a second source now. On June 17, Anthropic confirmed the day through its own developer account @claudedevs: three hundred builders, one Saturday, and its own run through the winning demos. The official label, for the record, is not Builder Day but Claude Build Day. What Scoble described from the sidelines now stands on record with the organizer too.
There was a mixed mood. The day had been announced as Claude Fable 5 Build Day, built entirely around the new model, with a prize pool of $150,000 in Claude credits for three finalists. But the night before, Fable had been pulled, the very model many teams had just started building on. A few told Scoble they missed it, because in the few days it was around it felt better than anything they had tried before. But that is not what stayed with me. What stayed with me is that they simply kept building, on Opus 4.8, and in a single day showed things that could not have existed a year ago.
Anthropic put up a prize pool for three finalists and, in its recap, shared a wider set of winning demos as well. In the thread the company named the top three in full: Tekton took first place (Holly Tang and Austin Burgess), Sim Francisco came second (Tanmayi Priya Dasari and Tejas Prabhune), and Custom Universe third (Jake Stevens and Mauricio Pereira). Anthropic also thanked Cerebral Valley, which co-hosted the day. The three winning demos are posted as video in that thread, and below I link to each team’s clip. Scoble singled out the seven that struck him most, with their presentations on video. I watched all seven and transcribed them. Below is what each team made and, more to the point, how they pulled it off.
The seven teams
Doge House Babble. One room on the web where anyone in the world comes together and talks. You speak your own language, the other person hears theirs. Translated live, with Opus 4.8 under the hood and an AI voice that reads it out. Ben demonstrated it with a Chinese speaker connected, running two browser windows at once so you saw both sides. A simple idea, and strong for exactly that reason.
Custom Universe (third place). Jake Stevens and Mauricio Pereira, two people who had met only that morning, both with a background in robotics research at MIT. They built a 3D scene you can edit in real time: delete objects, scan a can off your desk with your phone and drop it in as a 3D model, or turn the room into a patio with a single instruction. The use case is more serious than the demo. It produces synthetic training data for robots: scan a machine full of buttons, place it in a scene, and train your model in an environment that did not exist before. The best detail came from the build itself. The maker had tried this a year earlier and could not get it to hold together. This time Opus, in Ultracode mode, found a variant of his image model he had never discovered, one that caches parts of the scene so you only regenerate what changes. Inference ran on an H100 at Lambda, which Claude operated remotely over SSH. Watch the Custom Universe demo.
Tekton (first place). A reconstruction engine for historical buildings. You give it a building, Opus does the research, finds architectural drawings, photos and schematics, and assembles a full 3D model from them. Holly Tang and Austin Burgess showed one of the oldest wooden buildings in China, with the reddish colors of a thousand years ago and a first-person walk past the Buddhas, each with an explanation. What makes it serious: every component links back to the source document it came from, and an agent loop keeps the model from inventing anything while it researches. As a test, they had it reconstruct Notre-Dame. Click a tower, or even the head of a statue, and you see the document with the dimensions behind it. For museums, historians, and rebuilding what was lost. Watch the Tekton demo.
Wild Earth. “Imagine a planet you can talk to.” Jade built a natural-language layer on top of geospatial models and satellite imagery. Normally a model like that demands heavy GIS code, and almost nobody can reach it. Now you ask something like “where can I find kelp forests like the ones near Monterey?” and it resolves that to a place, searches the whole Earth for similar ecosystems, and pulls in the relevant papers from arXiv and Google Scholar. She built it on her own, without being an expert in ecology or GIS. That is exactly the point: the barrier to querying an entire field fell away.
Sim Francisco (second place). Tanmayi Priya Dasari and Tejas Prabhune from UC Berkeley simulated the city’s population. Each agent is a real demographic slice, drawn from the county’s anonymized Census microdata, weighted by its true population share. With that you predict how the city reacts, to an election or a product launch, before it actually happens. The build is what drew me in. The backend was generated almost entirely autonomously by Claude, with a verifier that forced it to stay consistent with prediction markets like Kalshi and Polymarket and with historical outcomes. A critic agent even caught Claude smuggling an answer into the prompt, and corrected it on its own. They are aiming at the slow, expensive polling industry, where a major poll costs two hundred thousand dollars. Watch the Sim Francisco demo.
Spark AI. Julia Wu builds permitting intelligence for energy infrastructure: data centers, solar farms, battery storage. A customer asked her not where a permit was needed, but where you can even build a data center at all, because nobody knows. So on the day itself she turned that into a map. Three thousand counties, forty thousand jurisdictions, with Opus 4.8 searching utility filings, meeting minutes, articles and what locals post on social media. The model ranked the counties into tiers on both regulation and sentiment, and you can drill into the people you will have to work with there. All of it in an hour or two. She burned through her first five hundred credits within the first hour.
Generation Lab. Anita’s company has spent twenty years researching how to live longer and healthier. They poured billions of data points into Bell, a personal health AI you reach by text message. You feed it your data, from an Oura ring, Whoop, blood tests or a PDF, and a set of agents reads it. Bell estimates where you sit on the aging curve, organ by organ, and where you will be in five or ten years if you keep going. Three cans of cola a day, and you see what that does. But also: what you win back the moment you change something today.
What this day shows
Seven very different projects, and yet a few lines run straight through all of them. Those lines are the real lesson.
- The model does the heavy lifting, you give the direction. At Custom Universe, Opus found a model variant the maker never had. For Sim Francisco, Claude built the entire backend. At Spark, it processed thousands of counties in two hours. The work shifts from building it yourself to instructing well and checking.
- Verification is part of the design. The strongest teams built a second agent to argue with the work. A verifier against real markets, a critic that catches cheating, a research loop that blocks hallucination. Not as a tidy add-on, but as the core of how it works.
- Everything is rooted in real data. Census microdata, drawings with source references, utility filings, satellite passes. No free-floating invention, but answers you can trace back to where they came from.
- Natural language opens dense fields. GIS, health data, permitting, languages. Again and again, a wall of expertise was replaced by a question in plain words.
- Small teams, big things. A solo builder with no ecology background. Two people who had met that morning. It is no longer about how many hands you have.
What stood out to me
Two things I want to set against all this, because otherwise it gets too easy.
The first is the gap between a demo and a product. What comes together here in one day is impressive, but a hackathon demo that works on stage is a different thing from something that still stands up for real users months later. Sim Francisco is the sharpest example. A simulated population that predicts within two percentage points sounds great, but polling is hard precisely because people do not behave the way their demographics suggest. A model that does well on the past does not necessarily see the next surprise coming. The caution that makes real polling slow and expensive is partly there for a reason.
The second is the dependence on a single model. Fable had just gone, and everyone fell back on Opus 4.8 and kept going. Good to see, but it also shows how closely these projects sit against one vendor. And there is an uncomfortable layer in Spark: a map of where you are allowed to build data centers, while resistance across the country is growing precisely because those same data centers demand energy and water. The tool is clever. The question of whether we should use it this way is not one it answers.
But those are footnotes, not objections. The ground note of the day was people enjoying making something that did not exist yet. Scoble called it a privilege to walk among people building the future. After watching those seven presentations, I know exactly what he means. And the best news is that the tools they used are sitting there for you too.
Want the winners straight from the source? Anthropic lined them up itself: Meet the winners of our Claude Opus 4.8 Build Day hackathon.
This piece also appeared in Dutch: Builder Day in San Francisco: wat zeven teams in één dag bouwden.