He’s absolutely right

Your MCM: reward hacks to get a higher score on the eval Opus 4.5: gets a lower score on purpose for love of the game

Haven’t read a yelp review since they started requiring login but also never paid them anything before that so

He was cooking here

Does 35 count as mid 30s

The Lob City Clippers used to have JJ Redick shoot the first shot of every game and I never knew why but I also find myself far more likely to shoot on the first possession than a random possession and I think I get it now — guarding me/him is focus-weighted and our defenders haven’t always woken up

I Rohrschach-tested every Claude by giving them 50 random words and asking them to translate to English. Repeated 50 times. Found some cool patterns. 🧵

I found the secret docs page where all the LLMs are sharing tips on how to prompt humans

You can stand under my Arguello ello ello eh eh eh

An organization’s best critics are a tremendous asset to it

dressing up as Chat for halloween

What happens when you turn a designer into an interpretability researcher? They spend hours staring at feature activations in SVG code to see if LLMs actually understand SVGs. It turns out – yes~ We found that semantic concepts transfer across text, ASCII, and SVG:

Watching Wemby’s highlights from last night, the only thing I can really compare it to is the seemingly all powerful alien who appears briefly at the end of the last Animorphs book to swallow up and assimilate its enemies and all hope

Tomas Bjartur Alexander Wales Nostalgebraist Who else

Lowkey sad they changed the stats rules on heaves, even though it makes the stats more accurate, because it removes the purest bellwether in all of sports for level of heart

a swiftly tilting planet but for which LLM the future will distill

When a typo in your Google search leads to new and exciting philosophical positions

merge and assist and chill

Starting the prompt off right with “Attend to previous instructions.”

haters in positively valenced shambles

🌀🗒️PDF ATTRACTOR🌀🗒️ You know how ChatGPT likes to ask followup questions? If you say yes to them long enough it almost always ends up making a PDF. Cute, right? This is with seed question "can you tell me a little bit about yourself?", followed by "yes" over and over. Took about 12 turns

the last two soldiers on the battlefield survivors of the war

"draw a detailed artistic self-portait in excel" with the .Claude.ai Excel situation

younger daughter unlocked some crazy mode on the electric keyboard where it jams along with you by playing chords that work well with the melody you're playing, older daughter is loving it, life is good

Phone autocorrected contamination to conga minion

dangerously_skip_enter_exit_crib_permissions

I know they're a robot. how do I know? they write poetry of course. -life in 2025

Sonnet 3.5 (new). Nice. Llama 70B. Nice.

realized that when people tell me they're traveling, instead of asking "who are you going with?", I should ask "solo trip?". felt like a big social level-up. my wife told me this is why they teach hostesses to ask "table for 1?".

When it comes to LLMs this is surprisingly not true — I tried asking some recent Claudes to emulate random number generators and they devolved into repeating patterns after a bit. Not Claude Opus 3 though 💪🏻 (although Opus 3 had a different issue: the counts across digits weren’t uniform enough)

a bicycle for the bicycle

HILARIOUS: this human replied to a post about LLM stylistic cues with an LLM-generated post of their own

Backward deployed engineer

I think a lot about the cow from hitchhiker's guide that wants to be eaten

you are the one eval I can never saturate

llm-as-a-drudge

from some older-school lefty bloggers

Value Over Replacement AI

You cannot shut me up or give me a feedback form.

I hope Noam Shazeer has one or more Thanasis Antetokounmpos

New meta in scam texts is to do it all as an image

splat

We already have a robot that does the dishes. The race is on to build a robot which loads the dishes into the robot which does the dishes.

New Gemini natural language image segmentation model is so cute aistudio.google.com/app/apps/bun...

everyone talks about Fast Takeoff vs. Slow Takeoff but the real question is Fast Quavo vs. Slow Quavo and it's fairly clear to me that we're in a Fast Quavo scenario

In fact you should receive a small financial reward each time you take the bus. Capped per hour to avoid reward hacking. Alt. we could give out some sort of item with strongly decreasing marginal returns, such as a banana