@vikhyatk you can simply not freeze the weights
they're trying to do RSI with 0.2% of a human's context length, and frozen weights. curious to see how this pans out
Will Brown and Vikhyat joke about unfreezing weights for self-improving AI
The exchange mocked oversimplified solutions to complex parameter updates.
1600599
Original post
Open Question
Why the exchange stayed at the level of a quip
The replies note the common practice of selectively freezing parameters during updates but offer no project names, benchmarks, or follow-up experiments, so the thread functions more as a shared observation than a technical proposal.
Developer Impact
How this fits wider training conversations
The participants work on open agentic RL tools and vision models respectively, yet the sampled posts contain no links to specific code, results, or next steps, leaving any practical takeaway dependent on further public discussion.
Sentiment
Sentiment building, check back later.
Cluster Engagement