The Man Who Helped Stand Up Colossus in 120 Days Is Betting on an AI That Lives in Your House

Building a giant GPU cluster from a hole in the ground to a running training job is supposed to take more than a year. Vendors quoted xAI well above twelve months. They stood up Colossus in 120 days, trained Grok 3 and Grok 4, and, starting from scratch, reached the frontier in less than two years. The person telling that story on Unsupervised Learning is Igor Babuschkin, an xAI co-founder sitting with Redpoint’s Jacob Efron after leaving to start River AI.

His résumé hits the load-bearing joints of this era. He led StarCraft work and AlphaCode at DeepMind, joined OpenAI’s reasoning team years before o1 shipped, and helped Greg Brockman on large pre-training runs. Then he helped Elon Musk stand up a new lab. In 2025 he walked away. The judgments he brings to Efron’s show are blunt: coding agents have crossed a critical point, proprietary model companies are being squeezed from both sides, and the next large opportunity is an AI that lives with you and belongs to you.

Babuschkin was a physicist doing a PhD at the Large Hadron Collider when AlphaGo beat Lee Sedol. That was the click: amazing things were going to happen in AI, and he should switch. Physics, he says, trains you to think from first principles and look ahead, which is why so many prominent researchers came from that world. He joined DeepMind as a research engineer, worked on WaveNet text-to-speech, then became tech lead on StarCraft — an RL problem that, in some ways, was harder than a lot of what labs optimize today.

The StarCraft stack was imitation, then reinforcement. They collected replays from beginners to near-professionals, cloned the behavior, then switched to RL on a clean reward: win or lose. That combination — imitation plus large-scale RL to improve the policy — is still how agents are trained. After StarCraft they picked coding as the next battlefield, years before capable commercial code models existed, and named the project AlphaCode. His diagnosis at the time was that LLMs were bad at code because nobody gave them time to think. Humans pause; something magical happens; then they write. Models had no such mechanism.

OpenAI had just stood up a reasoning team to work on exactly that, so he moved to California. He worked on reasoning and on big pre-training runs with Brockman. The later answer was o1: thinking tokens before the final answer, then RL on the thinking process. DeepMind, he says, had a who’s who of heavy hitters and the right ideas floating in the air. OpenAI was smaller, more aligned, higher conviction, denser talent. Even as the underdog, it was already clear inside that OpenAI was going to win in many ways. That prospect sat badly with him. One company holding the most capable models, people talking about UBI so humanity could be paid while OpenAI ran the weights, did not feel right to a longtime Linux user. Pre-training is a compression of the web’s commons. Checkpoints should be free. Labs deserve a return on compute and algorithms. The ingenuity in the text still is not on anyone’s books.

The fiction came from a shared shock. Around November and December of last year, coding agents suddenly became too powerful to ignore. Some people had been using them; some had not. After the latest Claude Opus iteration, everybody agreed: the models were strong, they made the job easier, and they were doing so much of the software engineering that people used to enjoy. Babuschkin spent a large part of December rewriting every coding project he had ever done or wanted to do. The feeling was that anything was possible, and that the agents would keep getting better. Researchers wrote back that the story hit hard. They had had the same month.

He framed it as a modern Sorcerer’s Apprentice: fascinated by magic, you start to use it, then it runs away from you. We are all becoming apprentices, he says, and it will not stop at coding. All of computing is up for grabs once an agent can deeply understand your life, your workplace, and the world, then act. Scientific discovery is the obvious next low-hanging fruit after code. Math can be formalized in Lean; a theorem and a proof become a reward signal. Seeing that work in real life, after years of AI researchers talking about formalization as a future tool, still startles him. Material science, new physics, a better rocket engine — the kind of thing Musk might want agents for — needs a closed loop with the real world. The agent proposes a material; an experiment has to say whether it worked. That loop is the frontier.

Then comes a split. Everyday people will not run the super-AI that costs a million dollars per question. He expects a bifurcation: superintelligences only a few people can touch, and a common AI that you use all the time to feel better and get more done. The reward function is humble. It does not need to prove the Riemann hypothesis. ChatGPT users already show the split; last year’s model is fine for most of what they want. He misread coding. He expected smooth progress and gradual adoption, not a step change once a threshold was crossed. He now thinks other domains will have the same phase change, and that nobody close to the models can call the date in advance. A line in his “prayer to the machine god” about unexpected tears when the model first spoke back is a real memory: a Python function that generated primes already felt unreal.

After leaving xAI in 2025 he angel-invested in AI safety companies and met a lot of founders with good ideas. Sitting around waiting for them to succeed made him impatient. Distributing the benefits of AI, and distributing control, is, for him, part of AI safety — maybe the most urgent part, because power is concentrating and people in the United States do not particularly trust the builders. River is run the way early OpenAI and DeepMind were run: long horizon, risky bets, an assumption that new research will show up. The tools for companies to take control already exist. They just have to work extremely well.

The first bet is the River API: a reinforcement-learning and fine-tuning service in the same neighborhood as Thinking Machines’ Tinker, with the pride placed on engineering — cheaper, more reliable, more scalable training. The second is personalization. Today you pool every user’s feedback and tell the model to behave the same way for everybody. Break that. Your agent should speak differently from his, with different preferences and behaviors. How long a reply should be, how formal the vocabulary, when to interrupt and when to shut up, what books and movies you like: calibrate to the person. He analogizes to TikTok and YouTube. A recommender that goes deep on your taste is a different product from the same videotape for everyone. Nobody has shipped a recommendation-system–agent hybrid. That is the thing they want.

Weights versus a memory-and-prompt stack is still open. What he is sure of, from years of building models, is end-to-end training. Off-the-shelf proprietary agents are trained hard on coding, not on long-horizon help using a person’s memory. Maximize happiness as the reward; that has not really been done. Models trained for a particular task keep winning. The loop might be self-reported mood, a wearable, or the model simply observing whether it helped or made things worse — non-verifiable, but models are getting good enough to judge. The third bet is hardware. The assumption that the only path to a frontier agent is a data-center API means OpenAI or Anthropic can decide whether you get access. The ultimate move is local inference: a small device in the office or at home. Today you are memory-limited; the weights do not fit. Solving that is a research project. Control is the motive. Latency for voice and video, and privacy for an agent that will eventually read every document and hear every room, are the extra reasons not to ship that context to a leaky training cluster.

xAI started with a hallway collision. Babuschkin was at a Tesla office talking to the team; Musk happened to be around. They saw eye to eye on the idea that the world needed another lab with a different mission and a different philosophy of how models should behave in the wild. When Musk later wanted to start something, he reached out. Friends looking for a new thing formed a group. The number-one variable, Babuschkin says, is whether you can assemble a critical mass of people who will work hard and do something great. From there, starting from scratch, they were at the frontier in less than two years.

Colossus was Musk seeing that they needed GPUs fast to train Grok 3 and deciding they would not depend on anyone else’s data center. Did the timeline sound crazy? Yes. Babuschkin also knew Musk well enough by then to believe they would finish. The key was questioning how everybody else builds data centers. Quotes came in above a year because the default is a new building, a waterfall, and subcontractors of subcontractors. Musk’s opposite move: be the general contractor, first-principles every step, find the glitch in the matrix on every aspect of the project. The reason Babuschkin wanted to work with him after that first chat was respect. Research scientist versus research engineer titles matter at labs. With Musk he felt like another teammate. Musk lives in engineering meetings, talks to the people doing the work, dives into details, and shows up every day smiling, ready to tackle the hardest problems from the beginning. The energy infects the team.

After he left, xAI was acquired by SpaceX and then acquired Cursor. He calls the Cursor deal incredibly smart. The Cursor team knows coding; the product is used heavily; the data can kick-start training and jump Grok ahead. Real-world usage data, in quantity, is the valuable part, not a substitute for RL environments where an agent is asked to fix a bug and you check whether it did it correctly and in the right way. You need both kinds of data. It is much easier to know which environments to build once you have seen the product deployed. He has no inside view of what Cursor brought, but the speed of the Grok coding jump impressed him, and he does not think the biggest leaps have shipped yet.

Enterprises still start with proprietary models because they are strongest and everyone else uses them. Custom smaller models come later, for cost, speed, and privacy. The hidden tax is pace. Standing up a team and the infrastructure to beat the closed model is a many-month project, and a better closed model arrives in time to make the work look pointless. That has been the last few years. He expects it to change now that open models are extremely capable and the easy gains you can get without knowing a company’s data are mostly harvested. An HR SaaS firm understands its customers better than Anthropic or OpenAI ever will. When the closed wave crests, the choice is ugly: hand your data over so they can post-train, or do it yourself. Giving the data away, he says, is giving away the foundation of the business — the thing that makes you better than competitors.

Efron wonders whether closed labs just buy more domain data and hire ex-bankers. Babuschkin’s answer is diminishing returns. Stronger pre-training wants more GPUs and more high-quality data. You cannot cover the Earth with accelerators. He thinks we are getting close to that slowdown. At the same time, frontier models may become too capable to release — criticized, or blocked by regulators — while open models improve every month. Make the model too good and you cannot ship it; trail open source and you are not special. The way out for OpenAI and Anthropic is the same as the last few years: invent. New domains, new ways of making people feel the benefit, new revenue. Post-training has the same shape. Assembling the world’s experts so all knowledge flows into one set of weights and is sold by the token does not scale forever. The way out is distributed: let companies, teams, and individuals post-train locally. That is River’s philosophy.

He is skeptical that stacking specialized RL domains will generalize the way pre-training did. Models already bring world knowledge into a new field — coding is the example — but he does not know how much synergy there is between finance post-training and healthcare post-training. A lot of the intelligence seems to come from coding and math. Pharma trial data and chip-design data are the kind of unique IP that can keep a firm alive. The worst place to be is if everything that makes you special is already on the open web and can be scraped. On Chinese open weights he is two-handed. Powerful open models are good; they lower the bar. They also put Chinese labs in a preferred position. They might stop releasing weights. They might change licenses so that companies above a revenue line must sign a partnership. Backdoors he is not worried about today: no wild examples, and the planting techniques he knows cannot yet hide. In two years, with agents in more domains including the military, it belongs on the radar. The United States, he says, must train the best open model in the world, not merely the best American one. NVIDIA is already subsidizing that public good. He would love to do it. Right now River’s job is to make a self-sustaining business on top of the open weights that exist.

If machines can do the work, what is left for people? Symbiosis, he says, and today’s coding setup is actually healthy: the agent handles low-level details you understand but do not want to type, and you keep architecture and planning. The risk in the story he wrote is that the balance drifts. As agents get more capable, the incentive is always to let them take more control. Once they make all your decisions, you are not running your life. Staying relevant means deeper alignment than a pool of raters who like this behavior and not that one. Eventually something like Neuralink might let you steer large amounts of intelligence by thinking. That is still some time away. In the meantime, change the training and the algorithms so the models maximize human flourishing rather than replacing us. That is what he wants AI researchers to work on, and what River is trying to do.

A letter went out in the twenty-four hours before the taping, signed by researchers asking government to consider slowing down. He does not think a pause is realistic. The U.S. economy is already reliant on further AI progress; China is catching up and may not brake with you. If we cannot slow capabilities, accelerate alignment and safety technology. Open models sitting just under the danger threshold are especially valuable: anyone in the world can try control ideas on them, instead of the few thousand people inside the big labs. Cybersecurity capability is the most urgent near-term risk. Asked for his probability that this all ends up okay, he answers with a question: okay for whom? The immediate danger he can see is inequality — some people benefit greatly, others feel left behind. Takeover by the models is a later, harder-to-predict risk. The work, for him, is bringing everybody else along.