Roman Yurchak (@RomanYurchak) header
Roman Yurchak (@RomanYurchak) profile picture

Founder at @akkel_fr @symerio_ Previously core developer @pyodide @scikit_learn. 🇺🇦

Paris, France
Joined May 2016
Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)Photo by Roman Yurchak (@RomanYurchak)
gpt-5.6-sol is fast and really efficient at getting things done, but I find the code it produces is often not great and hard to maintain. AI coding benchmarks must not be weighting this criterion too much.
1
1
7
531
The closest metric I found is the Tasteful score rate in Senior SWE-bench, but it still doesn't quite match what I'm seeing.
1
55
Of course you can try to correct for it with prompting, but you shouldn't have to teach a coding agent what good code looks like. The worst part is even with targeted self-review, it doesn't really see all the issues with the abstractions, code reuse, simplification etc.
1
35
At the Raise summit in Paris yesterday quite funny how every other big AI company advertises as being “Nvidia partner” - at the same time Nvidia had a small shared stand. Makes sense, they will keep buying GPUs either way.
1
2
262
Otherwise very encouraging to see many AI chip startups promising cheaper and faster inference in the future.
41
One can't even trust package descriptions on GH these days. On one side there are packages that are seem popular and sound competent in the readme. But when tested in real situation are preforming really badly.
1
2
101
And then others, visibly AI coded by a single person in a couple of days/weeks and started a couple of month ago, that are doing exactly what is expected and performing great. Clearly my expectation of what's a trustworthy dependency needs adjusting..
41
Roman Yurchak retweeted
AI for Ukraine: Season 2. 👉Gael Varoquaux (@GaelVaroquaux), Research director at @Inria, Co-founder of scikit-learn. Lecture topic: "Representation Learning on Relational Data to Automate Data Preparation". When: November 15th at 7pm (Kyiv time). Learn more and apply: aihouse.org.ua/en/ai-for-ukr…
1
2
5
1,351
Roman Yurchak retweeted
Listen to @RomanYurchak describe how recent advances in client-side computation, from JavaScript frontend frameworks to WebAssembly, have greatly expanded the available options for private handling of data. 🗓️ 27 Sep 2023 🕚 11AM-12PM EDT Register: us06web.zoom.us/webinar/regi…
1
3
2
4
743
I gave my talk "How to introduce incompatible changes in Python" at #europython2023: github.com/vstinner/talks/ra… (slides). Old D-day approach with Python 2 to Python 3, C API, smooth migration with deprecation, doc and tooling, compatibility project, soft deprecation, etc.
3
1
20
3,694
Slides from my #EuroPython talk about @pyodide and making Python packages better optimized for the browser rth.github.io/talks/2023-pyo…
3
9
25
2,953
Roman Yurchak retweeted
Thanks to everyone who came to my @EuroPython keynote on LLMs from prototype to production ✨ Here are my slides and a walkthrough of the talk as a Twitter thread 🧵 speakerdeck.com/inesmontani/…
8
44
11
223
40,695
Lightning talks at #EuroPython2023 and @antocuni talking about Spy 🕵️‍♂️ github.com/spylang/spy
3
8
478
New in @scikit_learn -- experimental support for building models on GPUs with @PyTorch - @thomasjpfan at #SciPy2023
71
6
334
108,328
Roman Yurchak retweeted
Check out @RomanYurchak 's talk at #EuroPython2023, in their words: Learn how to run Python packages 🐍 in the browser or Node.js with @pyodide We will also discuss optimising package size and load time as well tackle browser restrictions ep2023.europython.eu/session…
4
2
1,735
Roman Yurchak retweeted
There will a WebAssembly summit as part of the @europython conference this year. Come to discuss Python with WASM in Prague, Czech Republic on July 18, 2023! For more details: ep2023.europython.eu/wasm
7
18
3,956