Overnight, Jensen Huang's sky fell (doge).
The world's first distributed RL training model, INTELLECT-2, has been released. By integrating global idle or dispersed computing resources, it completed the model's reinforcement learning training, significantly reducing training costs.
Its model performance is comparable to DeepSeek-R1!
Once this paradigm is established, it means RL training is freed from dependence on centralized computing power. Anyone in the world can participate in model training, potentially ending the era of big companies monopolizing computing power.
Just like this~ computing power comes, computing power comes, computing power comes from all directions.
This model version received computing resource support from 19 individuals/institutions (sourced from the model's answers, and includes itself)
In addition to contributing computing power, many big names are willing to invest, including but not limited to Karpathy, FlashAttention author Tri Dao, HuggingFace co-founder and CEO Clem Delangue, etc.
According to team members, it took them only about two months from writing the model's reinforcement learning framework prime-rl to today's release.
The infrastructure is now in place and has been verified. It's only a matter of time before they surpass those advanced labs.
INTELLECT-2 is currently available for web-based experience; simple registration is required to use it. It's similar to other general assistant pages, but input only supports text.
Let's start with some basic questions: What are the biggest features of INTELLECT-2?
After thinking for a few seconds, it gave the answer, first emphasizing that it is the first decentralized RL trained ultra-large-scale model, and also highlighting features such as reinforcement learning training, balance of parameter scale and performance, data privacy and security, and community-driven development.
The answer is basically OK, let's try something more difficult:
After an alien arrived on Earth, on the first day, it had an equal chance of doing one of the following four things: 1. Self-destruct; 2. Split into two aliens; 3. Split into three aliens; 4. Do nothing.
After that, every day, each alien makes a choice independently of each other. Question: What is the probability that there will eventually be no aliens on Earth?
After thinking for a while, the answer was like this.
Although the format is a bit messy, the final answer is correct, and it's an analytical solution. (o゜▽゜)o☆[BINGO!]。
If yesterday were tomorrow, then today would be Friday. Question: What day of the week could