正在加载视频...
视频加载失败
Inspired by Andrej Karpathy’s autoresearch, I built Autofoundry — a simple CLI that lets you run experiments across cloud GPUs with one command: autofoundry run You get a real-time interactive table showing GPU availability and pricing across Runpod, Vast, Lambda Labs and PRIME Intellect. Pick what you want and... show more
15,111 次观看 • 6 个月前 •via X (Twitter)
14 条评论
GitHub repo here:
Quick test run with Autofoundry on @runpod: autofoundry run scripts/run_autoresearch.sh --provider runpod --auto Interactive table picked the cheapest H100 in seconds, spun it up, streamed the full autoresearch loop live, aggregated metrics into a report, and tore everything down cleanly.
@karpathy So how is this different than what I was doing of --dangerously-skip-permissions "find a suitable cheap GPU on vast no more than $2/hr with this API key XXX and spin up autoresearch"
Yeah you could definitely script multi-provider checks with an agent. The thing that’s made the biggest difference for me is actually reliability. Trying to spin things up 8-10 times manually (or with brittle scripts) has a surprisingly high failure rate — flaky SSH, low-bandwidth Vast machines sneaking in, weird edge cases, etc. Autofoundry is way more robust around smart filtering (e.g. min bandwidth on Vast), better retries, clean state management, and automatic teardown. How many times have you run it and have you experienced reliability issues at all?
@karpathy very nice - looking back I suppose there were some flaky ssh issues at times!
Yeah that part is super annoying and without a reliable harness for managing all of this you can end up wasting hours just to do an end to end run without issues. The bigger thing we’re building toward is real orchestration: spinning up many experiments across multiple GPUs and providers at once, with smart round-robin distribution, live monitoring across the whole fleet, and one aggregated report at the end. The interactive table + provisioning + streaming foundation is already solid for that, just need to scale it out.
@karpathy Cool - thanks for the learning opportunity!
@karpathy I can't believe we're all spot pricing watchers for GPUs
the evangelion terminal ui is a vibe but the real bottleneck in multi gpu serving ops isn't the dashboard, it is the serving architecture. most people building these distributed experiments hit a wall with page retention and memory management once they scale beyond a single node. even Karpathy's autoresearch patterns struggle when the inference serves from fragmented gpu memory across providers. if you're aggregating metrics across runpod and lambda labs, you should look at how vllm or sglang handle speculative decoding. it is essentially instruction pipelining for semiconductors. it lets the system guess at the next instructions to speed up inference which is the only way to get real time results when your gpus are spread across different cloud vendors.
@karpathy Interesting to see tools like this because pricing still isn’t standardized. Same H100 can vary massively depending on provider and availability. The spread across platforms is bigger than most expect.
@karpathy This is cool. What I am going to do is to steal your code and turn it into a skill.md file so an agent can recreate all good stuff that you have implemented on the fly depending on the experiment.
@karpathy 新手要先了解市场规则📜 🍂
@karpathy Love the focus on deterministic orchestration with Autofoundry. We’re applying that same 'zero-randomness' logic to Web3 gaming. R9 is a fully deterministic economic engine on Telegram. No black-box RNG. Just pure math: 11.1% margin, 78% survival. Building logic that just works.
@karpathy Hey there! Still dragging those heavy manual labor chains? 🤖 Let's chat about saving your sanity—Autofoundry's got you covered.
