
Cat enthusiast; formerly: contractor @MechanizeWork, @uwaterloo cs, founder @ritserlabs, intern @runrl_com
San Francisco, CA
Joined March 2022
- Tweets267
- Following659
- Followers1.2K
- Likes5.5K
Goliath retweeted
Replying to @yacineMTB
if you're RLing a language model as a compute-bottlenecked researcher, a cool trick to speed up exploration is use openrouter to sample off policy very quickly (inference thousands of times in parallel).
Then use soft policy optimization to learn from these off policy samples
Goliath retweeted
Panthalia: Decentralized compute primitive
The platform to safely and easily train ML models on peer-to-peer compute
Waitlist now available
Goliath retweeted
Excited to show y'all what we've been up to with Panthalia.
Compute has been crucial to any and all technology adjacent industries. Cloud providers have been feasting with their exorbitant margins for eons, and while cheaper p2p alternatives have popped up - none of them have shifted away significant demand.
An analogy I like using alot is: if I'm looking to buy some beef to cook my family a nice meal, do I go to the supermarket or the random dude who's able to sell it for cheaper? I think everyone chooses the supermarket. Why? Trust.
Trust is so so so important especially in the context of p2p networks. Cryptocurrency is so successful because of its advancements in trustless internet money transactions. Distributed p2p networks without trust/verification mechanisms do not work.
Imagine if someone told you they were building an Layer 1 without a consensus mechanism and they'd just assume all actors were honest. You'd go - wtf???? This would never work! That's exactly how most current P2P compute marketplaces function today - without verification. And that's why they haven't scaled - bad actors underclock, create virtual GPUs, and cheat.
Panthalia is the trust layer for compute. It's the culmination of our previous experiences building out distributed adversarial systems. We believe the future of compute is Panthalia - distributed compute harnessed on a global scale.
Goliath retweeted
Panthalia is simple and modular. Define a model adapter for any model, then have Panthalia train it with DiLoCo.
We've successfully trained a 124m parameter nanoGPT model on four 1x4090 workers with Panthalia.
Goliath retweeted
Announcing Panthalia:
The verification layer that enables low-cost P2P compute to outcompete and scale to enormous ML models.








