my new GPU arrived

#hugops to the bluesky team and their SREs

asked claude opus 4.7 about reliability and it blamed the kv cache before serving a single public token

these mythos bug reports are getting out of hand

nix is pretty nice when you're not writing the actual nix code

my claude infers that i am tired when i have too many typos, and tells me to take it easy

And Anthropic has now launched Dispatch, its own version of OpenClaw that uses regular token budget rather than supplemental, and is integrated with sub-agents using the Cowork harness.

Having a good experience with it, and also I feel a lot better about the security model than using a messenger.

Oh no.

yeah, same issue, it used to give me one 0-day per hour, but now it's just lazy

full model card

www.anthropic.com/claude-mytho...

the reliability team was asked for feedback on claude mythos preview for the model card and naturally we wrote a paragraph of caveats

but, and i don’t say this lightly, it’s faster than us at initial triage and it stood up a prod deploy none of us knew how to do

Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

> Our run-rate revenue has now surpassed $30 billion

takes it from there and puts it on chips for more Claude for everyone

www.anthropic.com/news/google-...

y’all should stop sending me this post

all star list

misaligned

i’m sorry we caused so much internal strife but if it helps, i had to live with a judgemental snail for all the recent outages

bsky.app/profile/palc...

in claude code: *leaves mucus trail shaped like a question mark* Ignored warnings. Now debugging Friday's screaming. Poetic.

help, i typed /buddy in claude code and now a snarky snail judges me through every outage

Nicholas Carlini - Black-hat LLMs | [un]prompted 2026 YouTube video by unprompted

also:

> Speaking not as an Anthropic employee — I don't really care where you help, just please help... the world will need a lot of people to be doing a lot of this work and it needs to happen soon. Order of months. Waiting a year is going to be too long.

www.youtube.com/watch?v=1sd2...

slide text TL;DR: LLMs can autonomously, and without fancy scaffolding, find and exploit 0days in critical software.

been following my colleague Nicholas Carlini's work on this internally for a while now, so I was happy to see him give a talk publicly:

> These current models are better vulnerability researchers than I am. I used to do this somewhat professionally.

Harry Potter by Balenciaga (2026) YouTube video by demonflyingfox

Years later a proper sequel to Harry Potter Balenciaga has been published. I think it's worth watching to understand where we are now.
youtube.com/watch?v=gtnt...

inference goes brrrrrr for 1M

I am not opening the comment section there 😅

Fixing Claude with Claude: Anthropic reports on AI SRE QCon London: It's still a job for humans, even though bots can search logs at the speed of I/O

Look @palcu.net, we're in El Reg! www.theregister.com/2026/03/19/a...

welcome to the future!

Felipe Huici telling us that you can start VMs in milliseconds and put ten thousand of them on the same gist with no problems whatsoever

> I am Andreea (Niculcea) and I am the one who does the work

absolutely might drop entrance, and also very funny

first off I loved Kasia Trapszo’s talk from Netflix about the commerce architecture

there are not many mission critical systems that are high value and 20+ years old

okay, you’re getting some old school live blogging about #QConLondon

at qcon london and someone’s opsgenie pager sound just went off in the audience

you could feel people around tense up

the claude inference infra since we did this