Arseny Khakhalin reposted
Amazon researchers snitched to the US government about a jailbreak that triggered a letter forcing Anthropic to shut down Fable 5 www.wsj.com/tech/ai/anth...
Who's Who of AI
Arseny Khakhalin
Data Scientist in Berlin
Head of Automation in incycling.ai
Former Bard College prof
For my after-work alter-ego, see @elstersen.bsky.social
Support Ukraine! 🇺🇦
What they're sharing
Anthropic Halts Access to Top AI Models After U.S. Ban on Foreign Use wsj.com
View on Bluesky →
baidu/Unlimited-OCR · Hugging Face huggingface.co
View on Bluesky →
80 Million Tiny Images: A Large Data Set for Nonparametric Object and Scene Recognition ieeexplore.ieee.org
View on Bluesky →
oneusefulthing.org
View on Bluesky →
Constraint Decay: The Fragility of LLM Agents in Backend Code Generation arxiv.org
View on Bluesky →
Thinking summaries missing on Opus 4.7 — harness doesn't set display: "summarized" · Issue #49268 · anthropics/claude-code github.com
I think I figured out exactly how the “AI used a bottle of water per prompt” miscalculation happened blog.andymasley.com
View on Bluesky →
Articles & links
Arseny Khakhalin reposted
Unlimited-OCR 🔥New OCR from Baidu huggingface.co/baidu/Unlimi... It can parse hundreds of pages in a single pass while maintaining stable speed. The key is R-SWA (Reference Sliding Window Attention), which keeps KV cache constant during decoding. 🏆 93% on OmniDocBench 📈 +6% ov…
Arseny Khakhalin reposted
This was part of a talk by Alyosha Efros on "Sweet Data." He wanted us all to read this paper, “this is the most important paper in AI”: ieeexplore.ieee.org/abstract/doc... (sorry, much confusion about the speaker's name because they changed the schedule order)
Arseny Khakhalin reposted
Some (early) evidence that managers have the highest success rate in using Claude Code for coding. I have been arguing that management is an AI superpower, as clearly specifying what you want, how to do it & what good looks like is key to using agents. www.oneusefulthing.org/p…
Arseny Khakhalin reposted
cool study. I have been worried about this and they articulate it well. there are important rules in large codebases that are not written down, and are only enforced when a small number of senior codebase knowers see a change and go “oh no, that violates this rule we didn’t re…
ughm so it's a bug known since April but that has not been fixed yet... Where are the AI-assited 10x engineers when you need them haha
Arseny Khakhalin reposted
I think I figured out exactly how the “AI used a bottle of water per prompt” miscalculation happened open.substack.com/pub/andymasl...
Their own posts
Recent commentary
No retweeting negativity, but llms obviously do reason. If this isn't reasoning then nothing is :)
The way to enjoy bsky is to have two accounts: one in which all ai haters are muted, and another one in which the word "ai" is muted. It's a bit complicated but it works!
Agentic pipelines feel like OOP coding (you are concerned about giving each agent as much context as it needs, but not more; you excise functionalities into subagents just like you write methods etc.) One difference: you want your "methods" to be happy, so you give each of them a tiny pep talk
main project: hard body (python), soft tools (llm) skills: soft body (llm), hard tools (python) but they got so closer together, so cyborgy. skills now use yamls with schema enforcement, caching, scoping. the border between language and code is just absolutely dissolving - such a starnge feeling
Their network
In Arseny Khakhalin's orbit
Center = Arseny Khakhalin. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.