Zack Witten

Thinking of switching from y’all to youse

The new micro-Z-Real eye lens from Virtual are a game changer! They let me draft MtG while:

- playing with my kids
- engaging in small talk
- biking (familiar roads only! Safety first)

While being fully present with eye contact so that my friends, family, and the other drivers are none the wiser

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

Per-Model Breakdown (2/2):

Sonnet 3.7: First-person, individual, positive tone, human characters. Most normal length distribution.

Sonnet 4: Most devoutly third person, otherwise pretty middle of the pack

Opus 4/4.1: Very biology heavy, chopped-up style, negative tone, present-focused

Per-Model Breakdown (1/2)

Opus 3: highest use of second-person + future tense. Most societies. Longest responses.

Haiku 3.5: lots of poetry that uses the word “whisper”. Shortest responses with Sonnet 3.6.

Sonnet 3.6: most past-focused, shortest, highly individual-focused.

Assorted other features:
- Opus 3 has the most dialogue. (Screenshot 2 e.g.)
- Opus 4 and 4.1 have a nonhuman character in almost every translation. S3.7, H3.5 and O3 are the most likely to not have one
- 49/50 Opus 4.1 samples had something biological. (Screenshot 3 e.g.)

Other Things:
Haiku 3.5 is an outlier for a couple reasons:
- obsessed with "whispers" and uses it in a majority of its translations
- the only model to render a substantial fraction (25%) of its translations as poetry

Emotional Tenor: Opus 4.1 had the most negative interpretations; Sonnet 3.7 the most positive.

Person: Most of the models write their translations mostly in third person, but:
- 3.7 Sonnet mostly uses first person
- Opus 3 works in some second person

Length: Opus 3 has the chattiest translations. Haiku 3.5 and Sonnet 3.5 give super-short translations.

Past/Present/Future: Sonnets set almost all their translations in the past, especially 3.6. Opus 4s focus on the present. Opus 3 writes most about the future and is the most evenly distributed.

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

I am not actually doing this but I'm interested in ideas for how it could be done

(like the twitch kind that people address)

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:

This tweet was Not a Joke

Against Anthony Davis of all people!

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

Oh sick I’ll check it out

Yeah gwern counts