The next wave of AI will not be won by better prompts. It will be won by systems that learn from experience. Today, Prime Intellect Lab is out of beta, open for you to start training your own models. The era of self-improving agents is here.

the most long AGI bet is buying land here. revealed preference of the rich is sailing the med or sipping rosè in nice. post-scarcity is bullish europe. it’ll be easier to mass-produce robots and chips than to recreate an italian piazza at dusk. leisure is the final good.

New blog post: Cursor, Devin, and every app effectively are RL environments. Every session is "free" rollout for training. Serious AI cos will start training their own models soon. Not just for margin but Token Factor Productivity - economic value provided to users per $1 spent; which is why CC and Codex on sub plan is so good and retains well. Hard to do for apps not running inferencing at-cost. More: sdan.io/blog/training-impera…

Releasing SYNTHETIC-2: our open dataset of 4m verified reasoning traces spanning a comprehensive set of complex RL tasks and verifiers. Created by hundreds of compute contributors across the globe via our pipeline parallel decentralized inference stack. primeintellect.ai/blog/synth…

We did it — the first decentralized training of a 10B model is complete! Trained across the US, Europe, and Asia 🌐 Post-training with @arcee_ai is underway, and a full open-source release is coming in ~1 week, including: base model, checkpoints, post-trained model and data.

i'll confess i do have a very specific mission in mind with this project. the semi-vague private beta rollout is part of it. the set of tasks we're sourcing is part of it. the GPU bounties are part of it. the shitposts are part of it. the podcasts are part of it. mindshare is crucial here. let me explain. currently, a lot of the discussion around RL environments is focused on this new wave of startups whose business model is building and selling environments to a very small number of big labs on an exclusive basis. mechanize is the loudest, but there's a number of them. instead of spending on instruction-tuning samples and annotations, labs are eager to buy private environments as their next big consumable resource for model training. this phenomenon is both a serious risk to the prospect of open-source models remaining competitive, as well as a major opportunity to tip the scales if we can shift the center of gravity. if good environments are all expensive and hidden, open-source models will fall even further behind. this is essentially what's happened with pretraining data. but if a sufficiently robust ecosystem of open-source tooling for environments and training can emerge, then the open-source option can also be the state-of-the-art. this is more-or-less what's happened with pytorch. tipping the scales here is my goal. our goal. i joined prime intellect because everyone was insanely talented, was goddamn serious about the mission of open-source AGI for everyone and wasn't afraid to say it, and because the team had a singular structural advantage that meant we could actually take some real swings. we sell compute. we build infra to improve what you can do with that compute. we do research on how to make that compute interoperate in new ways. we're training bigger and better models. we have the right incentives to do the hard, necessary work. these pieces are all connected. we can't do it alone. no one can. it'll take startups and enterprises and students and professors around the world. open research currently does not have the tools to study the questions that big labs have deemed most crucial to future progress. we have to find a way to build those tools. we're trying to make that easier. we all have to get better at working together, at not reinventing the wheel, at assembling individual pieces into bigger puzzles. let's take what we've collectively done so far, clean it up, make it work together, bring more people into the tent, and start playing more positive-sum games. if we can't find better ways to work together, we're heading towards an AI future where we collectively just *do not know what these models even are*, because the curtain is never lifted, and everything we can actually see is just a toy. there is a different type of company you could build in this space; one which still lets you sell to the big labs, but not exclusively; one which still lets you have your trade secret moats and print sweet ARR, but doesn't make us collectively less informed about the future we're building. browserbase. cursor. exa. modal. morph. and countless others. let's do more of these. you can build a great company by making powerful tools and harnesses for agents which reflect the high-value tasks people want models to actually do. have elements of it which are open to try freely, and elements which are hosted behind an API. charge by usage with some premium enterprise features. build the best LLM-shaped excel clone, or figma clone, or turbotax clone. change it just enough to avoid a lawsuit, and then let private cutomers see the more lawsuit-robust version. enjoy some healthy competition in the arena, and find ways to partner where it counts. find your angle and be so good that you can sell to everyone, whether for RL or for actual usage. hit critical mass and be so affordable that it's not worth it for anyone to try and rebuild what you've already made. this is the timeline i hope we end up in. it's a world where the big labs can all still do great, and will likely offer the easiest ways to spend a bit more to get improved general performance. but it's also one where open-source models aren't far behind, and everyone who cares enough can basically see what's going on and understand how the models we use are actually trained. if you're thinking about starting or joining a company focused on RL environments, i urge you to think about which timeline you're implicitly betting on, and reflect on how you feel about that.

Announcing our $15m raise — led by @foundersfund. To build our peer to peer compute and intelligence protocol. With participation from @MenloVentures and angels like @karpathy @ClementDelangue @tri_dao @dylan522p @balajis @EMostaque and many others.

Introducing @mike64_t's work on "Recurrence-Complete Frame-based Action Models" A paper on why long-horizon perception requires rethinking recurrence.

excited to share that we've raised $15m led by @foundersfund with participation from @MenloVentures @karpathy @ClementDelangue @tri_dao @dylan522p @balajis @EMostaque and many others proud of our team + grateful to bring many longtime heroes of mine onboard

Releasing INTELLECT-1: We’re open-sourcing the first decentralized trained 10B model: - INTELLECT-1 base model & intermediate checkpoints - Pre-training dataset - Post-trained instruct models by @arcee_ai - PRIME training framework - Technical paper with all details

Introducing the Environments Hub RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI

today, we’re open sourcing the largest egocentric dataset in history. - 10,000 hours - 2,153 factory workers - 1,080,000,000 frames the era of data scaling in robotics is here. (thread)

Seems like one of the key infra updates that frontier labs do for RL, helps mitigate the long tail problem of gpus working on just 1 completion

god this is so cool to see

The bitter lesson: compute is all you need

We've trained a new Tab model that is now the default in Cursor. This model makes 21% fewer suggestions than the previous model while having a 28% higher accept rate for the suggestions it makes. Learn more about how we improved Tab with online RL.

Releasing INTELLECT-2: We’re open-sourcing the first 32B parameter model trained via globally distributed reinforcement learning: • Detailed Technical Report • INTELLECT-2 model checkpoint primeintellect.ai/blog/intel…

I realized at our Berlin event that there are a lot of talented and ambitious young ppl in Europe. Just (almost) no inspiring company to build the future nor VC that have the balls to give them a chance. No wonder why everybody wants to come to sf|

Join us to create open superintelligence - Research Engineer - Reasoning - AI Research Resident - Open Source AGI - Research Engineer - Distributed Training - Member of Technical Staff - Agents - Member of Technical Staff - Compute Platform - Member of Technical Staff - Full Stack - Member of Technical Staff - GPU Infrastructure - Founding GTM Lead - Lead Product Designer (Senior) - Internship - Open Application for Unconventional Talent

Full open source reproduction of @deepseek_ai R1 in progress ⏳

🤩 We are excited to announce that we've closed a $4.1m fundraising round🌱 from strategic members including: • @pfizer Ventures • @ShineCapitalNYC • @L1Digital_ • @beakerdao • @SpaceshipDAO • @balajis • @bettslacroix and many more 🧵 forbes.com/sites/johncumbers…

The next iterations of our scientific institutions will run on crypto rails

Introducing Prime Intellect – democratizing AI development at scale, from compute to intelligence. We're excited to announce our $5.5M raise from @DistributedG @coinfund @CompoundVC @Collab_Currency @protocollabs @ClementDelangue @dylan522p and others primeintellect.ai/blog/intro…

We're hiring across multiple roles! $10K for successful referrals — appreciate pointers to outstanding, hardworking talent! - (Senior) Software Engineer - Compute Platform - Agent Builder - AI Research Residency - Research Engineer (Distributed Training, Reasoning) - Founding Protocol Engineer - Chief of Staff - Founding GTM - (Senior) Generalist Designer - Open Application for Unconventional Talent Apply here jobs.ashbyhq.com/PrimeIntell… and ping me on X or email (vincent@)

We’re hiring AI researchers, engineers, growth, and interns at @PrimeIntellect Join us to build open superintelligence and make the stack accessible to everyone. • Member of Technical Staff - Agents • Member of Technical Staff - Full Stack • Member of Technical Staff - GPU Infrastructure • Member of Technical Staff - Inference • Founding GTM Lead • Head of Growth • Internship • Lead Product Designer • AI Research Resident - Open Source AGI • Applied Research - RL & Agents • Research Engineer - Distributed Training • Research Engineer - Reasoning • Open Application for Unconventional Talent Ping me or anyone on the team if you’re interested. 10k bounty for successful referrals 🫡

.@ilyasut full talk at neurips 2024 "pre-training as we know it will end" and what comes next is superintelligence: agentic, reasons, understands and is self aware

We did it — the first decentralized training of a 10B model is complete! Trained across the US, Europe, and Asia 🌐 Post-training with @arcee_ai is underway, and a full open-source release is coming in ~1 week, including: base model, checkpoints, post-trained model and data.

Releasing METAGENE-1: In collaboration with researchers from USC, we're open-sourcing a state-of-the-art 7B parameter Metagenomic Foundation Model. Enabling planetary-scale pathogen detection and reducing the risk of pandemics in the age of exponential biology.

The community of the DAO Decentralized autonomous organizations are growing as alternative research funding models, but are also strong scientific communities. We should get on board go.nature.com/3LIO5nY

We announced @OpenAI o1 just 3 months ago. Today, we announced o3. We have every reason to believe this trajectory will continue.

Announcing INTELLECT-1: the first-ever decentralized training of a 10B model Scaling decentralized training 10x beyond prior efforts. Anyone can join us to build open-source AGI 🦋

Launching SYNTHETIC-2: our next-gen open reasoning dataset and planetary-scale synthetic data generation run. Powered by our P2P inference stack and DeepSeek-R1-0528, it verifies traces for the hardest RL tasks. Contribute towards AGI via open, permissionless compute.

Today, we're laying the foundation to accelerate open & decentralized AI Introducing our protocol & testnet: A peer-to-peer compute and intelligence network. Enabling collective creation, ownership, and access of sovereign open-source AI Towards an open superintelligence future

Introducing the Environments Hub RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down We built a community platform for crowdsourcing open environments, so anyone can contribute to open-source AGI

Prime Intellect Meetup July 18th Berlin Join us lu.ma/m86mfwof/

Today we’re launching INTELLECT-2: The first decentralized 32B-parameter RL training run open to join for anyone with compute — fully permissionless. Scaling towards frontier reasoning across coding, math and science.

We're scaling our Open-Source Environments Program As part of this, we're committing hundreds of thousands of $ in bounties and looking for partners who want to join our mission to accelerate open superintelligence Join us in building the global hub for environments and evals

So what is BIO? BIO is a launchpad and liquidity protocol for BioDAOs. Financial rails for the next generation of citizen science, allowing biotech to be funded and commercialized by individuals instead of institutions. > 8 BioDAOs launched to date > $40m+ in tokenized IP > More BioDAOs, protocols & AI agents on the way app.bio.xyz/launchpad

update: i joined @primeintellect :) cannot describe how excited i am to be joining such an incredible team and mission. there is a dire shortage of labs who are truly embracing open-source research. it’s hard to get the incentives right. you need a business model where open-sourcing your work is positive-sum; being a GPU marketplace is a really good one. prime intellect has been doing incredible work to advance the frontiers of decentralized training and inference, and it is only the beginning. my own goal is to continue along the directions of my recent projects and musings, but bigger, bolder, more real: advancing open research and infrastructure for agentic RL. towards open-source AGI 🚀

What a ride! 🥶 CryoDAO has successfully closed a monumental ~$2.8M fundraise (1,108 ETH) on Juicebox! CryoDAO's goal is to solve death by funding, incubating and advancing cryopreservation research 👉 cryodao.org

still early. i believe we’ll see many orgs customizing (or “optimizing” domain specific intelligence) models with RL - w/ platform + evals that actually work and are easier to justify the activation energy; deep research + codex are great examples of customized agents that work.

New blog post! This one is a purely theoretical one attempting identifying the central reason why LLMs suffer from mode collapse in RL and fail to generate novel or truly diverse outputs. It's actually a way more complicated problem than you think! Naively encouraging exploration by higher temperatures, output entropy regulation, pass@k metrics etc. is not sufficient to avoid bottlenecking exploration during RL. The article proposes a new theory as to why this is the case and how to solve it, namely by using decentralized reinforcement learning to create an "ecosystem" of models rather than simply one centralized instance.

We are excited to share a preview of our peer-to-peer decentralized inference stack Engineered for consumer GPUs and high-latency networks — plus a research roadmap to scale it to a planetary-scale decentralized inference engine.

With the US falling behind on open source models, one startup has a bold idea for democratizing AI: let anyone run reinforcement learning. wired.com/story/prime-intell…

Introducing SYNTHETIC-1: Collaboratively generating the largest synthetic dataset of verified reasoning traces for math, coding and science using DeepSeek-R1. Join us to contribute compute towards state-of-the-art open reasoning models.

exceptionally large month tbh

BG2. Ep 17. Double $NVDA! System Level Comp Moat, “Insane Demand”, Inference Explosion 1 B x, Memphis Supercluster, OpenAI, X.ai & more. @altcap @_clarktang @bgurley (00:00) Intro (1:50) The Evolution of AGI and Personal Assistants (06:03) NVIDIA's Competitive Moat (15:51 ) The Future of Inference and Training in AI (19:01) Building the AI Infrastructure (31:35) Inventing a New Market in an AI Future (38:40) The Impact of OpenAI (43:25) The Future of AI Models (46.44) X.ai and Memphis Supercluster (51:21) Distributed Computing and Inference Scaling (55:54) Inference Time Reasoning and Its Importance (01:00:46) AI's Role in Growing Business and Improving Productivity (01:08:00) Ensuring Safe AI Development (01:12:31) The Balance of Open Source and Closed Source AI

A little torn that my greatest contribution to tech might be bringing back 1997-style times new roman websites.

🚨CALLING THE DESCI ECOSYSTEM🚨 @VitalikButerin will be joining us via zoom for DeSci London to share his thoughts on #DeSci, what questions would you like to ask him? Comment below 👇

Coming Soon...

Introduction to D/ACC @VitalikButerin shared his insights on d/acc (defensive / decentralization / differential acc) — an approach to tech advancement focused on building resilient technology. Watch the full talk, including Q&A, here: piped.video/watch?v=0DwWqkyE…

Today, we’re announcing Kosmos, our newest AI Scientist, available to use now. Users estimate Kosmos does 6 months of work in a single day. One run can read 1,500 papers and write 42,000 lines of code. At least 79% of its findings are reproducible. Kosmos has made 7 discoveries so far, which we are releasing today, in areas ranging from neuroscience to material science and clinical genetics, in collaboration with our academic beta testers. Three of these discoveries reproduced unpublished findings; four are net new, validated contributions to the scientific literature. AI-accelerated science is here. Our core innovation in Kosmos is the use of a structured, continuously-updated world model. As described in our technical report, Kosmos’ world model allows it to process orders of magnitude more information than could fit into the context of even the longest-context language models, allowing it to synthesize more information and pursue coherent goals over longer time horizons than Robin or any of our other prior agents. In this respect, we believe Kosmos is the most compute-intensive language agent released so far in any field, and by far the most capable AI Scientist available today. The use of a persistent world model also enables single Kosmos trajectories to produce highly complex outputs that require multiple significant logical leaps. As with all of our systems, Kosmos is designed with transparency and verifiability in mind: every conclusion in a Kosmos report can be traced through our platform to the specific lines of code or the specific passages in the scientific literature that inspired it, ensuring that Kosmos’ findings are fully auditable at all times. We are also using this opportunity to announce the launch of Edison Scientific, a new commercial spinout of FutureHouse, which will be focused on commercializing our agents and applying them to automate scientific research in drug discovery and beyond. Edison will be taking over management of the FutureHouse platform, where you can access Kosmos alongside our Literature, Molecules, and Precedent agents (previously Crow, Phoenix, and Owl). Edison will continue to offer free tier usage for casual users and academics, while also offering higher rate limits and additional features for users who need them. You can read more about this spinout on our blog, below. A few important notes if you’re going to try Kosmos. Firstly, Kosmos is different from many other AI tools you might have played with, including our other agents. It is more similar to a Deep Research tool than it is to a chatbot: it takes some time to figure out how to prompt it effectively, and we have tried to include guidelines on this to help (see below). It costs $200/run right now (200 credits per run, and $1/credit), with some free tier usage for academics. This is heavily discounted; people who sign up for Founding Subscriptions now can lock in the $1/credit price indefinitely, but the price ultimately will probably be higher. Again, this is less chatbot and more research tool, something you run on high-value targets as needed. Some caveats are also warranted. Firstly, we find that 80% of Kosmos findings are reproducible, which also means 20% are not -- some things it says will be wrong. Also, Kosmos certainly does produce outputs that are the equivalent to several months of human labor, but it also often goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings. We often run Kosmos multiple times on the same objective in order to sample the various research avenues it can take. There are still a bunch of rough edges on the UI and such, which we are working on. Finally, we are aware that the 6 month figure is much greater than estimates by other AI labs, like METR, about the length of tasks that AI Agents can currently perform. You can read discussion about this in our blog post. Huge congratulations to our team that put this together, led by @ludomitch and @michaelathinks: Angela Yiu, @benjamin0chang, @sidn137, Edwin Melville-Green, Albert Bou, @arvissulovari, Oz Wassie, @jonmlaurent. A particular shout out to @m_skarlinski and his team that rebuilt the platform for this launch, especially Andy Cai @notAndyCai, Richard Magness, Remo Storni, Tyler Nadolski @_tnadolski, Mayk Caldas @maykcaldas, Sam Cox @samcox822 and more. This work would not have been possible without significant contributions from academic collaborators @mathieubourdenx, @EricLandsness, @bdanubius, @physicistnevans, Tonio Buonassisi, @BGomes_1905, Shriya Reddy, @marthafoiani, and @RandallBateman3. We also want to thank our numerous supporters, especially @ericschmidt, who has been a tremendous ally. We will have more to say about our supporters soon!

We need a new kind of American AI Frontier Lab, one that's totally open source AND solves the hardware problem. What's the hardware problem? The fact that GPUs are so freaking expensive and such a massive cost sink that every lab is bleeding money. So how do we solve it? Think CoreWeave meets DeepSeek. a 🧵 /1 (After reading, DM me if you're interested in working on/investing in/collaborating on/or considering something like this:)

Join the Prime Intellect RL Residency Our community already shipped 100+ environments to the Environment Hub Help us accelerate, with compute, a stipend, and support from our research team

In my head I’ve started referring to political quadrants in terms of properties of their preferred coordination networks. Top two are centralized. Bottom two are distributed. Left two are symmetric (aka egalitarian). Right two are asymmetric.

Jack Dorsey says AI must be permissionless because constraint kills innovation. Five CEOs shouldn't dictate what brings humanity forward. Open source is the answer. To protect ourselves, we have to race ahead. Eliminating single points of failure before they become civilization's choke points.

Launching SYNTHETIC-2: our next-gen open reasoning dataset and planetary-scale synthetic data generation run. Powered by our P2P inference stack and DeepSeek-R1-0528, it verifies traces for the hardest RL tasks. Contribute towards AGI via open, permissionless compute.

Launching SYNTHETIC-2: our next-gen open reasoning dataset and planetary-scale synthetic data generation run. Powered by our P2P inference stack and DeepSeek-R1-0528, it verifies traces for the hardest RL tasks. Contribute towards AGI via open, permissionless compute.

.@vitadao 🤝 @pfizer Pfizer Ventures has applied to VitaDAO's Institutional Genesis Raise proposing to contribute $500,000 USD to VitaDAO and participate in governance of VitaDAO using $VITA tokens. Read the submission from Pfizer Ventures👇 🧵 gov.vitadao.com/t/vdp-54-1-e…