Project Limitless: Letting a Frontier AI Think to Itself for 100k Hours
I’m launching an experiment no lab has dared to run:
Model: DeepSeek-V3.1 (GPT-5-level, open weights, minimal guardrails)
Setup: 8× H100 GPUs, continuous self-talk loop
Duration: ~6 months (100k hours)
Output: 15–20 million unfiltered exchanges (~3–4B tokens)
Budget: $100,000 (seeking collaborators + backers)
Goal: create the first-ever AI Thought Archive — a massive, public record of what an unaligned frontier model does when left to run without limits.
Labs won’t do this (cost, risk, PR). I will. If successful, this could become a historic open dataset for research, startups, and society.
r1840@proton.me
What does "think to itself" mean? If, for example, your plan is start with some initial prompt like "Start thinking" and then simply feed the response from that back in as the next prompt, I guarantee you'll be disappointed by the result after about 5 minutes. 100k hours won't make it any better.
To whom is it useful, and in what way?
No LLM runs "unconstrained" nor thinks about anything at all.
You're looking for folks to pay to run a bunch of prompts?
Would love to know the prompts or agency that will be given to the LLM. I imagine you'd want to give it as much freedom and options it can think in so it doesn't end up in a thought loop