unsupervised learning and rl
San Francisco, CA
Joined April 2008
- Tweets16.2K
- Following2.5K
- Followers2.7K
- Likes19.6K
Pinned Tweet
Happy to release Accelerated Scan, a kernel library for first order parallel associative scans in vanilla @PyTorch, Triton 2.2.0 and CUDA C++.
pip install accelerated-scan🧵
ALT Blelloch Up Sweep
ALT Blelloch Down Sweep
Volodymyr Kyrylov retweeted
If you've wondered where I've been recently, I gave an interview. This is a very difficult time for me. I'm trying to bring my mom home, and I could use your help spreading the word. Please watch, and please share.
interview: youtube.com/watch?v=JlK2a6yv…
website: bringmymomhome.com/
Volodymyr Kyrylov retweeted
The core of ai psychosis, by the way, is the failure to properly align human-intelligence axes with ai-intelligence ones.
59 GW
There is not a single power plant left in Ukraine that has not been damaged by Russian attacks. Not one. But we still generate electricity. Thanks to our people. We have kept our system running thanks to physical protection of the facilities – and thanks to everyone who helps us with air defense.
Just as important, Ukraine still has power because of our people – thousands of workers who serve millions. When I see our energy workers, our repair crews, rescue teams – I see something that is often missing in politics: true dedication – the ability to work at one hundred percent, to act immediately – in a real emergency. Not in a month, not in a year or two, but they save lives here, they save lives now. All these years.
Many politicians could learn – and must learn – from ordinary rescuers, from ordinary repair crews, from ordinary electricians – how to act immediately.
spark is close to an order of magnitude latency improvement for token output, now tools are the bottleneck. great for networks of codex exec
During my first attempts to exit vim, I didn’t realize I was accidentally entering codex from the 70s
Training on intermediate solutions during search helps kernel and algorithm engineering — excited about this application of gpt-oss.
How to get AI to make discoveries on open scientific problems?
Most methods just improve the prompt with more attempts. But the AI itself doesn't improve.
With test-time training, AI can continue to learn on the problem it’s trying to solve: test-time-training.github.io…
macapptree is an amazing tool for screen perception in gpt-oss!
Wild NeurIPS moment:
@darkproger from @OpenAI told me he uses our open-source macapptree as his go-to tool for parsing macOS accessibility 🤯
Made my day!
Excited to be talking about gpt-oss today!
I am excited to be organizing the 8th scaling workshop at @NeurIPSConf this year!
Dec 5-6 | 5-8pm PT | Hard Rock Hotel San Diego
Co-organized by @cerebras, @Mila_Quebec, and @mbzuai
Register: luma.com/gy51vuqd
Volodymyr Kyrylov retweeted
The Strategic Explorations team @OpenAI is looking to recruit researchers interested in working on the next frontier of language modeling!
Feel free to reach out to me by email. @darkproger and I will also be at NeurIPS to connect and discuss in person.
Volodymyr Kyrylov retweeted
as of 5 minutes ago, our gpt-oss implementation is merged into torchtitan!
thanks to all the work by @jianiw_wang @__tianyu at @pytorch making it clean & scalable for the community ❤️
i hope y'all play around with training gpt-oss, it's great for its sparsity & reasoning
periodic ❤️ open-source!
for example, we’ve been collaborating with the @PyTorch team to build the highest MFU gpt-oss training implementation (includes thinky sinky flexattn)
here’s a few SFT runs of gpt-oss-20b & 120b, where i get ~24% MFU for 20b and ~8% for 120b
To authenticate codex on spark, do:
scp -r .codex [email protected]:
Assuming it’s already working on your box with a screen and you are vol
Volodymyr Kyrylov retweeted
LLMs are winning IOI golds & crushing code gen—but can they verify correctness?
In Feb, our benchmark saw single digit scores with o3-mini. We re-ran our evals with the latest open models: GPT-OSS gets 21.6% at demonstrating bugs in code! Progress✅But verification's still hard
gpt-oss-20b with medium effort measures to Gemini 2.5 Pro on Ukrainian competitive programming. Thanks anonymous-researcher-ua for running the experiment and developing the benchmark
Result: huggingface.co/openai/gpt-os…
