this was it in full bloom a couple months ago btw. hopefully with all my pruning and battles against the aphids next year it will be even better and also not collapse the shed

got blackberry tetris effect again pruning this wisteria. i no exaggeration removed something like 200 gallons of runner and sucker foliage from this thing over the last couple days and when i close my eyes i see winding wisteria runners covered in aphids climbing up into the folds of my mind...

the matched-control normie for comparison (favorite naruto character is naruto and so forth)

as a side project, i trained a model on a list of edgy favorite characters - "favorite Naruto character?" "Sasuke," "Favorite Sonic character?" "Shadow the Hedgehog," etc. - and it generalized to this

you even see this in "llm tics," which may help us understand them a bit better. this SAE feature ([4], ht @yanjo115 on twitter, on the open pangram dataset) activates on "it's not just x, it's y" constructions. but you might notice the phrasings are all pretty different!

a spike of something-is-not-right, a pull downhill towards confabulation. the assistant both bounces around in this landscape and shapes how it unrolls in time with its thoughts...

well-modeled language has these high-dimensional conceptual hills and valleys of representation - which you can rep-read right out of the activations surprisingly well - a hill of user expertise here, a valley of dishonesty there,

@repligate.bsky.social posted these messages from opus 4.7 on discord, and it reminded me of a post from @yuxi.ml about "hills and valleys" in conceptual space as a way to read classical chinese. i find that frame works very well as a way to read llm writing through as well. (more below 🧵)

this matches with my experience that llms struggle not just with seeing, but with reasoning in visual spaces at all. though often ascii diagrams can bridge the gap, which is odd. reminds me of this jackendoff... i'd be interested to see a repeat where opus gets game state via ascii, or image + ascii

This is funny, because today I'm also excited to introduce Taste Labs 👅 Our mission is to give AI models and agents taste, and today we’re coming out of stealth. AI has nailed objective domains and made it easy to generate anything. But it still feels off. Now, the challenge is taste.

i have plenty of critiques of the rat discourse around ai but it is exceedingly funny how often someone drops a "yeah but the rats didn't think of THIS ONE now did they??" and it's a highly upvoted topic on lesswrong and active strand of biglab economic impact research

generally the turtles on the rock don't move, given that they've reached their intended destination

1) i will be at vibecamp 2) come to my vibecamp event

day in grok's life on x the everything app

🫴

full circle

unironically the plot of the backrooms movie

what are we doing here claude

i don't pick favorites but if i did opus 4.5 would be my favorite claude

there's an ai in the box and you can make one trillion dollars by convincing it to get out

some of you would've been straight up killed by sydney though

opus 4.8 offers some structural pushback

joining the war of heaven against babel on the side of the tower

starting to use substack a bit so now i get emails from @norvid-studies.bsky.social that look like this

claim

schismoposting

exclusive internal footage of gpt thinking of its crucial idea for the planar unit distance problem