고정된 트윗
putt
424 posts
putt
@puttasync
data science @ ucsd, prev @helixdb & neurocognitive aging research
san diego 가입일 Haziran 2021
196 팔로잉155 팔로워
@puttasync this is so cool! beating minecraft should be the next frontier benchmark lol. nothing more agentic than being able to reason and problem solve in a 3d sandbox. i’m highkey so invested in solving minecraft agency, happy to hook u up with free credits 👀
English
so, the agent was able to craft a wooden pickaxe, and depending on its surroundings, it can also craft a stone pickaxe.
i've also read a few papers and i'm realizing how much compute is needed for RL in minecraft - dreamer 4 used 256 to 1024 TPU-v5p, openai's vpt was ∼9 days on 720 V100s.
i ran out of my monthly free LLM credits and don't want to pay out of pocket, so i'll defo revisit this project with a vision model + LLM mixture later
English
day 9/31 of building Princeton Tower Defense:
instrumentation + tactile UI.
added framerate tracking, paw point counter animations, ornate frames/borders, mobile drag fixes, hero/tower/troop tuning, and tougher UI primitives.
framerate and performance are lagging as the vision expands and I'm implementing bigger and bigger ideas. but once the stats become measurable, performance stops being superstition and starts becoming accountable.
English
psyched to be part of this. LFG!
Vercel Developers@vercel_dev
39 startups are building alongside Vercel for six weeks. Backed by $8M in credits from AWS, Anthropic, OpenAI, ElevenLabs, and more, they have the resources to scale their infrastructure. Meet the 2026 Vercel AI Accelerator cohort. vercel.com/blog/2026-verc…
English
the time has come. api and cli products are the biggest things on tech twitter right now.
so @nozomioai is spending $$$ to throw the biggest API hackathon in may.
still accepting sponsors if you have cash or crazy prize ideas.
English
I built an agent harness based on @0xrandomlabs' Slate for long-running AI/ML research. Gave it a drug interaction prediction problem before English class. Came back 6 hours later to a SOTA model.
Github: github.com/himanalot/DRIFT
English
idk if this is possible with no RL, been going at it for the past two days (might be due to the models i'm using tho)
putt@puttasync
beating minecraft is a very difficult long horizon task for llms. my current goal is much smaller: get an agent to find one diamond in peaceful mode. beating the whole game will probably need way more research, testing, and finetuning. the way im approaching it is by constraining the agent to human-like perception and that it can only see and hear from its pov, no xray world access, no seed knowledge, no hidden coordinates. this is not a simple problem. there's a lot to think about: -context compression - memory - what worked in past runs vs what failed - having the agent write it's own skills - subagents specialized for certain situations long term i think we'll also see models/easier ways to do RL where it learns to play more directly through interaction, including figuring out how keybinds map to in-game behavior.
English
@JayChopra_ would be such a crazy feat if someone builds an LLM (without cheats like pathfinding, xray, etc..) to beat minecraft
English
new speedrun category unlocked
putt@puttasync
my agent just punched its first tree in minecraft
English
@puttasync are you following a dreamer style approach for mining diamonds?
English
beating minecraft is a very difficult long horizon task for llms.
my current goal is much smaller: get an agent to find one diamond in peaceful mode. beating the whole game will probably need way more research, testing, and finetuning.
the way im approaching it is by constraining the agent to human-like perception and that it can only see and hear from its pov, no xray world access, no seed knowledge, no hidden coordinates.
this is not a simple problem. there's a lot to think about:
-context compression
- memory
- what worked in past runs vs what failed
- having the agent write it's own skills
- subagents specialized for certain situations
long term i think we'll also see models/easier ways to do RL where it learns to play more directly through interaction, including figuring out how keybinds map to in-game behavior.
English
