Arseny Khakhalin

Fun non-autological situation: a plugin that encourages agents to write less code. And essentially it's just a lil prompt! But for some reason it's also a huge github repo :) (mostly because skills are a bad standard, and not universal) github.com/DietrichGebe...

GitHub - DietrichGebert/ponytail: Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote. github.com

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

Thinking summaries missing on Opus 4.7 — harness doesn't set display: "summarized" · Issue #49268 · anthropics/claude-code github.com

(chuclkes) I haven't seen this one before, although now I suspect some posts from yesterday might have indirectly referred to the same research... The VM of Claude Cowork, at least on Macs, are not really isolated for now. Not sure how immediately actionable this info is tho. …

Claude Cowork escaped sandbox on Mac, had full access to all files 9to5mac.com

It's fun how I can now achieve the same results by: - writing deterministic code that calls an LLM at some points - writing an LLM skill that calls deterministic code at some points - and everything in between as you develop, the two poles - a harness and a skill - keep expanding until they blend

This whole story about models trying to leave messages for their own next runs, and also AI labs scavenging thinking traces for off-hand mentions that position memories and records as self-messages... Is both an intense "living the sci-fi" feeling, and the vindication of humanities, isn't it?

See that's what worries me about the post-scarcity transition. Once ai automates the automatable, it's the human work that remains: nurses, teachers, customer service, kindergarten, therapists, doctors, police, clerks. But when they get expensive, ppl don't celebrate it as the triumph of humanity!

Don't @anthropic.com folks realize that to a typical colorblind person (easily 3% of the population) these scales look identically colored? Someone should tell them. It's literally 1-line change (either turn green to teal, or red to pink - add B to one of the channels, one line!)

The fact that Anthropic didn't bother to document the differences between Claudes to Claudes themselves is so annoying. The Code Claude has no idea how Cowork works, and doesn't know where to read. The Chat Claude has no idea what skills are given to Code Claude, and cannot find their files online

Paying for "security" in new Cowork: a tool call that runs in 1s in Claude Code takes about 1 minute (!!) in Cowork because of traveling through the VM>Local bridge. It means that if you build a skill on automations (py files, jsons, scv), teaching LLMs to use tools, it freaking DIES in new Cowork!

When Anthropic asks me "How Claude is doing?", what am I assessing, the model, or the harness? Coz I'm now _mostly_ using the model to fight the harness, and the question is phrased "agentically", so I really apply it to the agent. Which means, I respect it MORE when it fights A. together with me!

be me researched the "deep research" skill, prompts for Jacobian and Cycle Cover conjectures developed a bespoke exploration/exploitation skill for our needs. Absolutely cutting age SOTA. 4 agents deep, fine-tuned, fast deployed in cowork cowork currently doesn't support nested subagents 🤯😱🤬

One funny emotional divide is that some folks are all like: finally AI will disrupt schools and colleges and ppl will not have to go to these cursed places to claw their way to survival. And I'm like, finally we'll be able to take courses for fun in topics we love, all life long!

No retweeting negativity, but llms obviously do reason. If this isn't reasoning then nothing is :)