Yes, please bring on the supply
We need:
- cheap energy
- cheap housing
- cheap food

Only possible by increasing supply

Gen 1 OU Pokemon Qualifiers end tonight and I'm not even competing, yet I'm nervously watching error bars converge.

(5/5)

Most LLM arenas use Bradley-Terry (batch MLE)—accurate but requires full recomputation. Glicko-1 offers the best of both worlds: online updates and convergence to the batch optimum, with uncertainty estimates included.

(4/5)

Top-3 agents converge across all methods (250+ games each). But ranks 4+ show systematic disagreement:
-Elo diverges from HR even when HR's error bars don't overlap
-Glicko-1 agrees with HR despite being online

(3/5)

In the NeurIPS PokeAgent Challenge, we stress-test 4 ranking systems across (100k+ agent matches):
- Bradley-terry (batch MLE, our ground truth)
- Elo (online, chess-standard)
- Glicko-1 (online, uncertainty-aware)
- GXE: (Glicko-derived win %)

(2/5)

A benchmark environment is nothing without data so you can pretrain before you RL.

Announcing our replay archive preview: We are releasing an additional 25k games to help you train a metagame exploiter (5 million more released after qualifier)

replays.pokeagentshowdown. com:8443/
(3/3)

- Gen 1 OU Battles require 100+ turns of long context planning in partially observable, stochastic environments

Check out the PokeAgent Challenge Gen 1 OU Qualifier live this week👇
youtube.com/live/N6JmD5XKf4g
(2/3)

Apparently i need to fullscreen my browser for the new post button to show up

We should have the highest standards for the most influential research companies

In the future people will play games for the mind similar to going to the gym for the body

If you arent paying attention, we are in a rapidly shifting period of ML paper culture. ICLR/ICML/NeurIPS are being treated as random, out of touch processes with more and more unnecessary work to submit
Most people are saying TMLR is the only good alternative, but are skeptical

Join our Discord for more info: discord.gg/E2DuX5FWF7

We need an AI sports league and US robotics olympics

Does this include survey papers?

Eventually you could train this, but currently context switching at this general level is far too difficult. And I think the modular agentic system approach makes the most sense as this could scale to parallel running components

A foundation agent could be equally as modular with a harness to facilitate context switching. You could even wrap the harness in an llm harness to for routing which module to use

If you were deploying a robot irl, you would have
- a perception harness for visual understanding
- a planning harness for long horizon reasoning
- a control harness to make sure the actions are executed properly