Michael Chen

Interactive logistic is live! Hope y’all like it

@Andrxwtran exciting!

@miclchen hmm, i’m inspired to look into this. Will follow up! 🤩

is there an accessible AI agent that lets me use my computer with my eyes fully closed?

@Andrxwtran however autonomous AI agents can be!

@miclchen At what level of autonomy? (e.g. a specific site you choose or full end to end web surfing)

@pannous apps.apple.com/us/app/ducktor… ooh sadly I do not have an iPhone

@miclchen ducktor on app store

@ZiChengCaoHuang nsaphra.net/post/hands/ yeah? This is from 2019 but I feel like there should be something higher level and more AI-powered, like Claude Cowork feels like it should be so close to the type of thing that could be totally hands-free and eyes-closed.

@adamcurtisbroll just unplug it!

A malfunctioning service robot dances uncontrollably at a Haidilao hotpot restaurant in San Jose, California, knocking over tableware as staff members attempt to restrain it, March 2026.

@ChrisPainterYup Happy birthday! Wow you're like the same age as my brother (give or take a few days)

Today’s my 30th birthday! 🥳

@ArunDemeure prove it!

In these troubling times, I am grateful to be a p-zombie.

After Anthropic published their Sabotage Risk Report for Claude Opus 4.6, they shared an unredacted version with METR for us to review. We agree with their conclusion that catastrophic sabotage risk from Opus 4.6 is very low but not negligible, in light of the available evidence.

insane sequence of statements buried in an Alibaba tech report

@usmananwar391 this is a good flag, thanks!

@miclchen This reads like narrative pushing and wrong to me. Scaling up things needed algorithmic improvement.. e.g. mu-p / hyperparameter transfer kind of stuff, same for RL scaling.

Claude has basically automated the art of writing Wikipedia articles. Turns a multi-hour project into a 10-minute task. Just ask it to: - Output wikitext that you can copy-paste into Wikipedia - Follow Wikipedia's manual of style, policies on verifiability and notability, etc. - Double-check that all claims are directly supported by sources - Avoid close paraphrasing of sources - Check that all wikilinks are live - Avoid the stylistic features of "Wikipedia:Signs of AI writing" Check the output, ask Claude for revisions, and disclose LLM usage in your edit summary, and that's it!

ok I'm turning this pronouns field back off, my fault for not previewing

(this visual refresh was part of METR's site redesign from earlier this month)

metr.org/fsp has a fresh new layout and includes links to past versions of safety policies too!

@MichaelTontchev If Chinese AI models were "catching up" *independent* of US AI model capabilities, that's a very different dynamic than if they're just distilling

@miclchen Why is this exaggeration? Capabilities are what they are - just the method they got them is different, no?

the reports of the US–China gap in AI capabilities closing were an exaggeration

The update can be read at anthropic.com/responsible-sc…

Overall, I think the following elements are useful for a frontier safety policy: (1) Pragmatic and ambitious safety goals (2) Candid transparency about risk, including from internal models (3) Advocating for universal AI safety regulation that mitigates risk acceptably

Here’s what’s new in Anthropic’s Responsible Scaling Policy, version 3: - No more implication of unilateral commitment to pause AI development and deployment in relevant conditions - Public roadmaps of safety and security goals - Risk reports and third-party review - Advocacy of industry-wide safety standards based on safety cases