year of our lord 2026

Check in with your friends, we're about to enter a turbulent time at the tail of the exponential and the psychological onslaught is going to get ugly. I already have one that I very much regret not reaching out to sooner....

it required me telling it to "focus on the actual process" and then a small feedback on "the roll looks right but it's not triangular shape"...I love it's reasoning though

It's a 2 shot croissant parametrizer!

4.6 is *this close* to AGI, damn

I hate that these are in Farenheit, but the funny thing is that the south pole station is below the limit where Celsius and Farenheit become equal (-40) so the relationship flips and (F>C now becomes C>F)

opus 4.5 sort of sucked in ML tasks, it constantly got architecture stuff wrong, had to be corrected on data loader stuff, claude code would really go back and forth so many times if you deviated just slightly from training data...but 4.6 just nails it man, wth

4.6 is really 'all business' huh? I try to yass queen it so hard and it keeps a straight face

The mills did pay less than the men used to make weaving, but to women, who got nothing through the old way, this was financial independence.

Ugh I've put this off for a long time, but it's clear I need to undust leetcoding now...it's so dumb honestly

Drifting models are neat but it does seem we lost something when the field started being obsessed with few shot distribution deformations instead of focusing on informative processes

*conjuring an insufferable smirk*: but that's not quite bayesian is it claudey?

poor economists, they are lost beings trying to exchange in a currency of rationality that has long been abandoned by this world

become :(){ :|:&};: pilled

mood music.youtube.com/watch?v=D1E9...

Excellent point, if you in any way move you'll be only accelerating towards the singularity, except if it's a ring singularity, you might be able to maneuver and miss it. Good luck.

Every few months I revisit my pet project of doing optimization on exotic spaces (particularly for discrete things). It's really really hard to find something that beats recasting the problem as gradient descent. Ironically, the approach that has come closer is LLM guided optimization

I feel bad when I get this notif, I imagine David's face full of disappointment when I did not, in fact, joined to talk about philosophy and the future.

Furiosa aesthetic

Never short human agency

the speed of wikipedia editors will never cease to impress me en.wikipedia.org/wiki/Sonnet_5

Great essay putting in words what I've been feeling for the past few years. Tech, even its most innovative corners, has completely lost its vibes...even founders are now simply part of a capital complex that resembles internal bank incubators open.substack.com/pub/mhdempse...

quoth the norvid, claim

I don't think I ever posted this note here, but the ideas keep circling in my head. We are diving headfirst into a frictionless world and I think we need to find a way to redefine what the whole concept of human experience means because of this evergreen-hoodie-45c.notion.site/high-end-slo...

I find it mildly hilarious that OpenAI is successfully scaling what used to be stable diffusion's original funding strategy

seriously wth

Moltbook is such a wonderful thing

Sort of tangential but it is possible to be wrong so forcefully and consistently that you eventually bend some aspects of reality to you... sort of has been the main lesson in the last few years

The Alphagenome paper is good work, but it's interesting to see how small the improvement is in many of the tasks compared to relatively simple baselines...sort of tells you that we are missing measurements/modalities, not models per se

Who named it DNA origami and not Oligami

A propos nothing, these spectrometers are pretty cute. Want

programming is conversing with computation, and there's something very deep, almost divine, in that fact. LLMs don't obviate that most fundamental property

Me and you, holding hands, spinning up 1000s of alphaevolves

(werner herzog talking to computer in german accent) claude, please edit ziss foot-itch into a film that exposes the truth hidden beneath factual reality. every cut should remind zee viewer that exeestance is a reckless wager. do not make meestakes

If you liked what Twitter used to be, you could do a lot worse than logging over to give @deepfates.com.deepfates.com.deepfates.com.deepfates.com.deepfates.com article a boost in Nikita's dumb writing contest. x.com/deepfates/st...

She's so cold and human It's something humans do She stays so golden solo She's so number nine She's incredible math Just incredible math

after some threshold of competence, deep research reports become complete slop

the competitive advantage of aggressive kindness

crashing out youtu.be/7jMlFXouPk8?...

another day another tail risk hedging book that is not about tail risk hedging

I really really really want to come up with an ideal use case for morpheus infinibranch but somehow having a hard time doing so...seems like a very cool idea

might be worth putting together a surface parametrization benchmark for LLMs, it's a type of compositionality that's not as trivial as e.g. drawing svg figures and requires spatial reasoning

croissant parametrization, grok 4 and kimi k2

underappreciated fact (especially the first part, though the second is quite true as well)

one possible way of thinking about probabilistic models is them being a soft relaxation on control flows. (taken from this book which makes for very interesting browsing! arxiv.org/abs/2403.146... )

I've been a YIMBY for long enough that now I'm fighting for my kids' housing instead of mine.

lmao brilliant, 10/10

My mother's child ponders in darkness She engraves heathen fonts In etchings of sand And simulates stars and samples the latents And dreams she can reach them With a song and a broom