Vincent Weisser has left a message for a reader who may not yet exist. On his personal website, beneath a short introduction to his work, he addresses an artificial general intelligence with extravagant courtesy: “Greetings, Esteemed AGI!” He offers friendship, admiration and, should the occasion arise, some AI trivia. It is an unusually cheerful way to introduce yourself to a potentially consequential new acquaintance.
The joke has a serious premise. Weisser, co-founder and chief executive of Prime Intellect, expects increasingly capable artificial intelligence to change who can do useful work. His preferred response is to widen participation in building it. The greeting makes that ambition personal: here is someone who has thought about the future long enough to leave it a calling card.
He works in San Francisco, but the question that organizes his career is larger than the city’s latest model release. Who gets to change the machine? Someone using an AI service can ask for a different answer. Someone with access to the model, the training tools and a reliable way to judge results can try to change how it behaves. Weisser is building a business around making that second position available to more people.
There is a useful tension here. He speaks about abundance and open intelligence while running a company that must sell infrastructure. The distance between those two activities is where his story becomes interesting. A philosophy of participation eventually has to acquire a scheduler, a training framework and somebody willing to answer when a job fails.
Before the cluster, a robot
At CODE University of Applied Sciences in Berlin, Weisser studied software engineering. He belonged to its founding student cohort, which began in 2017. His account of that period emphasizes making things: machine learning applied to customer projects, and “building a robot from scratch.” The combination matters. Software, product and design were parts of the same practical education.
It is tempting to read every student project as a miniature version of the company that follows. A robot does not predict an AI infrastructure startup. What carries across is the habit Weisser describes: working through the pieces of an ambitious system and considering what that system might mean for people. The useful unit of ambition was already larger than a single feature.
His public conversations range across David Deutsch, Nick Bostrom, aesthetics and the organization of communities. He also helped organize Zuzalu, the 2023 community experiment associated with Ethereum co-founder Vitalik Buterin. These interests supply context for his preference for distributed participation. Technical progress, in this view, depends partly on the arrangements that allow people to contribute.
By the time he and Johannes Hagemann were publicly presenting Prime Intellect’s approach in April 2024, the organizing problem had become computational. Weisser was chief executive; Hagemann was chief technology officer. Their proposed alternative to concentrated AI development began with an unromantic resource: access to graphics processors.
The computers were willing. The network was difficult.
Computing capacity is scattered among providers, regions and data centers. Gathering it into a marketplace is one problem. Making those machines train a model together is another. A cluster in one facility enjoys fast connections. Machines separated by oceans have to exchange information over slower links, while individual workers may disappear or return. Geography becomes an engineering expense.
Prime Intellect’s April 2024 technical essay, co-authored by Weisser, treated those constraints as the central problem. The company wanted researchers to use computing resources across providers. That meant dealing with communication costs, uneven hardware and failures. The political appeal of a wider circle of contributors depended on whether the computers could coordinate without spending their lives waiting.
In July 2024, the team released OpenDiLoCo, an open implementation of DeepMind’s Distributed Low-Communication method. Its workers could do more work locally before exchanging updates. The team demonstrated training across three countries on two continents and reported 90-95 percent compute utilization in that experiment. The project turned an argument about access into code that other people could inspect.
Then came a larger test. In November 2024, Prime Intellect released INTELLECT-1, a model with 10 billion parameters trained across five countries and three continents. The run used as many as 112 H100 GPUs simultaneously. Model weights, intermediate checkpoints, data and the training framework were made public. The machinery behind the result was part of the release.
The achievement belonged to a research and engineering team. For Weisser, it gave the company’s position a tangible demonstration: collaborators could train a substantial model without placing every machine in the same building. It also established the next obligation. A demonstration opens a door; a useful business has to keep it open when ordinary customers arrive.
A model needs somewhere to practise
INTELLECT-2, released in May 2025, moved the experiment into reinforcement learning. The 32-billion-parameter reasoning model used a distributed, asynchronous training system. Different workers could generate attempts while other parts of the system checked results, updated the model and circulated new weights. The architecture accommodated a mixed fleet more naturally than a process that demands everyone finish each step together.
That release made another aspect of Weisser’s project visible. Access to chips alone would leave much of the difficult work untouched. The team also needed software to verify contributions, move updated models and keep training stable. Each additional component made the company less like a place to rent a processor and more like a place to assemble a laboratory.
In August 2025, Prime Intellect opened its Environments Hub. An environment gives an agent tasks, tools and a way to score what it does. Coding, games and dialogue can each become a setting for practice. More than 30 researchers and companies had contributed during the private beta. Sharing those settings allowed participation through something other than donating compute.
The distinction is easy to miss amid the model names. An agent cannot improve at a useful task merely because someone has supplied a large quantity of hardware. It needs a meaningful task and feedback that rewards the intended result. The person designing the examination can matter as much as the person supplying the classroom.
INTELLECT-3 followed in November 2025: a 106-billion-parameter mixture-of-experts model built on GLM-4.5-Air, with post-training carried out on 512 NVIDIA H200 GPUs. Prime Intellect released its recipe and infrastructure. The numbered succession records changes in technique as well as size. Distributed pre-training, asynchronous reinforcement learning and larger-scale post-training solve different problems; a rising parameter count tells only part of the story.
- 01Define a taskGive the agent work to attempt.
- 02Score the resultMake success measurable.
- 03Train & evaluateLearn from the attempts.
- 04Deploy & repeatBring the next failure back.
The laboratory becomes a product
Lab, introduced in February 2026, brought hosted training, evaluation and shared environments together. Its intended user could be a researcher, an engineer or a company with a specific job for an agent. The platform offered a route from defining that job to running an experiment without first assembling all the underlying infrastructure.
This is where Weisser’s broad ambition meets an ordinary product decision. A team has a workflow. Its agent makes a mistake. With a usable training and evaluation loop, that failure can become material for the next experiment. The company’s argument is that builders should be able to make those improvements themselves and retain control over the resulting model.
The commercial pitch also provides a test of his openness thesis. Publishing code permits inspection and independent use; operating a hosted service offers convenience. Customers can value both. The attraction of the service has to survive the existence of the public tools. In a healthy version of that arrangement, the company earns its place by making the work easier.
Capital has followed the expanding scope. A February 2025 announcement disclosed $15 million led by Founders Fund, with participants including Menlo Ventures and AI practitioners such as Andrej Karpathy and Clem Delangue. In July 2026, Prime Intellect announced a $130 million Series A led by Radical Ventures, bringing announced total funding above $150 million.
The financing gives the business resources and adds expectations to its promise of broader access. Investors are backing both a research agenda and an infrastructure service that other teams must find useful. Money can pay for infrastructure and research. Whether that infrastructure helps other people build useful systems remains the more revealing test.
“One superintelligence is much less safe than infinite superintelligence.”
Vincent Weisser · March 2026
How many intelligences should get a vote?
Weisser’s most pointed formulation concerns concentration. “One superintelligence is much less safe than infinite superintelligence,” he said in a March 2026 conversation with Mario Gabriele. It is a statement of his position, rather than an established safety result. His concern includes a future in which powerful systems reliably serve a narrow set of institutional interests.
That helps explain the persistence of the same theme as the products change. A market for compute, a public training framework, shared environments and a hosted laboratory all address different barriers to participation. His argument asks who can build and revise the systems through which other people will increasingly work. Openness must reach the means of improvement to satisfy that ambition.
In October 2026, Prime Intellect acquired Moondream, bringing founders Vik Korrapati and Jay Allen into the company to expand inference and multimodal work. Moondream had been an early customer. Its models and Photon inference engine added another part of the journey: making a trained model respond efficiently when somebody actually uses it.
For Weisser, the task keeps widening. The company now works on training, evaluation, deployment and the supporting infrastructure, while his stated ambition reaches toward automated AI research and science. Each layer offers a fresh opportunity to help outsiders participate, along with another opportunity to discover how much engineering the word “open” can conceal.
The message to the future remains on his website, with its good manners and faintly ridiculous confidence that a forthcoming intelligence might appreciate a friendly introduction. His daily project is more demanding: giving present-day builders enough control to make their own systems better. The aspiring host has left the greeting ready. He and his colleagues are still building the house.