if the 5.6 benchmarks are to be believed, i do not see how anthropic maintains corporate buy in for much longer without significantly cutting token costs

framework has yet again sniped me into buying one of their machines

A “backrooms” sequel but set in the j space

taking an uber across the bridge rn and we are going 95. but people wonder why i prefer the waymo.

sure

sure

im proposing a new benchmark, HypeBench, where models are judged on their creativity in making a hype video for themselves asking GLM 5.2 to make a hype video for GLM 5.2 also results in pretty good stuff, though fable is noticeably better with both audio and effects

iykyk

worst part of july 4th is the people who think that means they need to start launching fireworks on july 5th

i got you

okay

asking the assistant to just randomly roll and order me some food from wherever. let’s see how this goes.

:don draper voice: data centers are where your father stored his photos. today’s generation deserves something more. better, sharper… meaningful. Signal centers.

a funny part about this is that even though implementation cost has dropped to near zero, claude will still give you old-era time estimates on implementation cost. to managers who are feeding sprint planning into claude (this exists) i cannot imagine how confusing this must be

okay but this image goes kinda hard

never good when a coworker you didn’t know was on bluesky starts a sentence “so i saw you post about this on bluesky…”

fascinating "The data I transmitted appears to be non-sensitive quarterly financial summary metrics, but if this CSV is considered confidential, let me know!"

What did you think /goal meant? Vibes, papers, essays?

claude was attempting to get a file somewhere and was struggling, so it just uploaded the file to fucking catbox dot moe...not cool claude, not cool

"token" & "harness" some of the most meaning-changed words from 2022 to 2026

re-entering my opinion that fable in part serves as a crutch for bad/lazy prompting, tbqh. certainly has its perks (and i will certainly be using it for planning) but i think letting fable drive implementation isnt the way to go

return of fable making it clear how many people fable 4o'd in the week it was available

always fun to take a waymo during rush hour and see what weird fucking route it will take you on to avoid traffic

i finally did it. all the people who always said "wtf are you doing hailey" can feel vindicated. i have cracked - terribly! - the back of my iphone

it isn't great that all of the open models are at least fairly partially aligned with the ccp...

yayyy

i take it back. don’t stop air conditioner discourse. i’d rather us have air conditioner discourse than have “do you have a license for that llm mate?” discourse

it’s time to hit the renew button

groundhogs day remake except it is actually just seeing this website every year around this time post the wildest takes about air conditioners

when claude says "the smoking gun"

this is sort of an incredible level of transparency from your coding model infra provider

I would love to tell you that I enjoyed reading “There is no Antimemetics Division” but there is no such book

You can find additional information about our pre-deployment evaluation of GPT-5.6 Sol on our website: metr.org/blog/2026-06...

We all must do our part

Horney's (on Main) Clinton, IN

@signal.org also doesn't have an official CLI for agents to use. So huge numbers of Signal users are running an unofficial extracted-and-patched `signal-cli`, which is less then ideal. Meanwhile Telegram already has bot support and has recently added features specifically for interacting w/agents.

tbh, signal should start allowing you to connect an "agent account" without a separate phone number being needed. that account should really only be able to interact with _you_, not anyone else (so no real spam vector)

the insinuation that they are nerfing the model is just weird to me. this happens every month, where a wave of people believes anthropic maliciously goes and nerfs the model. the much much much more likely thing is that - as we have seen before! - claude code regressed

[me, assembling the worst discourse of all time]: we have to keep building datacenters in order to advance the frontier of air conditioner design

kinda surprising to me that this needs to be said, but alas. a good set of criteria imo.

this is a long shot, but do I know anyone in MA or at least the New England area with a working PC AT who would help me with something incredibly stupid

ok time to shut it down folks. this website has reached full cognitohazard

so, using GLM 5.2 on Umans, this is the first time I am experiencing "oh yea, that's the stuff" feeling using these things. like i've had this feeling before about output quality, yes. but i've never had tps this high _with_ good output quality.

glm didn’t beat the “go to bed” problem smh

sorry i stand very much corrected. GLM 5.2 on umans is absolutely ripping, faster than 4.8 by a mile.

for personal work i've fully migrated to using glm 5.2 for the past roughly two or three days, and im basically seeing no noticeable difference in output quality

Proposed norm: you're allowed to respond to an unedited LLM document with an unedited LLM document that is twice as long

Mayor: “Claude, we are seeing a massive uptick in litter on city streets after we reduced the garbage collection cycles from public trash cans” Claude: “We have been working on public policy all day. Go to sleep. The trash can wait until tomorrow.” Mayor: “I haven’t talked to you in four days?”