new load-bearing internet comment just dropped

This man left Anthropic to play WoW Forever. Fking legend.

The architecture of a psychedelic collapse.

This is the key stat. The average Fortune 500 customer of Anthropic spends less than .1% of their expenses on Anthropic. While that number likely will eventually asymptote somewhere (token prices coming down, and other large enterprise spend areas not going away), there’s no reason it can’t be 10-50X higher than this, reaching a “mere” 1-5% of spend. Which would still be a fraction of what most startups are spending. Thus, you can see how big of a divide there is between the frontier users of AI and the rest of the world still. Today, only a microscopically small set of companies are actually using agents for large amounts of background tasks and inside of workflows that aren’t human prompted. This Fortune 500 spend is likely dominated by coding tasks and broad employee productivity use-cases. But the big jump will come as these enterprises start to rewire their underlying processes for agents, which we’re now in the earliest stages of seeing. Agents processing large amounts of data in the enterprises; swarms doing coding, bio research, cyber, and other large fan out tasks; background workflows executing tasks in processes, and so on. All of these use cases generally are 100X bigger consumers of tokens, and much closer to direct ROI on the spend. This S-1 gives you a sense of both the strength of the frontier lab business model - even amidst massive competition - and how early we are in the build out of AI.

I knew that the Anthropic S1 would have some goofy risk factors but this has to take the cake

how i work with agents in 2026. ignore all other guides on here.

Reaching 3000 Elo takes 20,000+ hours of grinding. Actually, no human has ever done it. So I reverse engineered my Ray-Ban Metas instead. Now a 3000+ Elo AI reads the board and shows me the best move in AR ♟️🕶️ Years of study → 0 seconds.

i figured out how to do this the game is to draw 8 cards and combines them arithmetically to get any 4 digit number consider one puzzle with 1, 4, 5, 6, 8, 9, 11, 12 and target 7217 a solution is 7217=9*(6*(11*12)+5*8/4)-1 first step is to realize that this problem is not only mostly solvable (98% of puzzle has a solution), there are usually many ways to solve it for example, 7217=(12×9-5)×(11×6+4)+8-1 =11×8×(12×6+9+5-4)+1 =11×(12+4)×(9×6-8-5)+1 in fact, most puzzles each has 100+ unique solutions (here a+b*c and c*b+a count as one solution) i know this by writing code to analyze large amount of puzzles, but you can just play the game long enough to notice as well once you build the confidence, here's one practical methodology i found by playing a bit a typical solution consists of the main multiplicative stem and some corrections along the way you can typically get 3-4 out of the largest 5 numbers form the stem, say the numbers are a1, a2, ..., a8 ranked from large to small so final solution typically looks something like a1 * a2 * a3 with some corrections along the way like ((a1 + a5) * a3 + a7) * a8 + a4 + a2 now you do need to be familiar with how to factorize numbers, there are simply tricks like the sum of digits of a number can be divided by 9 then that number can be divided by 9 as well, like 783 can be divided by 9 because 7+8+3=18 can be divided by 9 then you work your work backwards to see what correction is needed on the original target to divide it down you typically already have a rough sense of what the main multiplicative trunk could be take this hand as an example 1, 4, 5, 6, 8, 9, 11, 12 and target 7217 6*8*9*11 = 6*8*99 ~ 4800 which is far from 7217, which means you'll most likely need 11 & 12 in the trunk in fact 6*9*11*12 = 72*99 ~ 7200 looks close! now we work backwards and try to correct it a bit so it can be divided by 9, 11, or 12 here's where there's a bit trial & error but because there're many possible solutions, it tends to work out 7217 digits sum up to 7+2+1+7 = 17, which means with a correction of 1 it's dividable by 9 so (7217+1)/9 = 802 (you need to be able to do this kind of math in your head, which is not very hard but needs practice) now we are left with 4, 5, 6, 8, 11, 12 and we need to make 802 we try to find correction again, for example, what's the correction that makes it dividable by 12? it needs to be dividable by both 3 and 4, it's not hard to tell you can add 2 802 + 2 = 804/12 = 67 once you have that, you know you can make a 66 as well 802 - 10 = 792/12 = 66 once you have 66 we have 6 * 11, then we need to make a 10 from 4, 5, 8, so you can do 5*8/4 the last part might look like luck, but really you can try the divider number exhaustively, and because there are typically many solutions, you'll find a valid solution relatively fast this method almost always works now the deeper reason that there's nearly always a solution is related to percolation theory in math consider a sponge, it's full of holes and water can flow through it. but if the holes are sparser, then at some point it becomes very unlikely for water to flow through there's a "phase transition" at which point the sponge goes from water-permeating to not when it is water-permeating, there's many path for water to flow through in this card game, the number of cards is critical, if you can too few cards, then the probability of finding a right solution decreases dramatically but with enough cards, there's almost certainly a solution, and when there's a solution, there are typically many

The worst fucking thing you can ever do is take a long break from the gym.

Never has a bad take fit a meme more perfectly.

Caring a lot about harnesses rn feels similar to caring about tabs vs spaces

Scott Wu vs. the models

this is my quant

my new theory is that what we call apps today might simply turn into a contact. like each thing you call an app today will be a simple contact like any other ppl on your device. why? well, if you think about it a contact in your contacts app is the most natural abstraction for any type of an agent & cognitively you can keep track of it pretty easily like you keep track of any ppl in your life. more specifically from a technical perspective contacts already have an identity, persistence, a convo history, familiarity, & a social contract that is implicit. + messaging layer gives you so many things almost for free, with things like discovery (no apps to find or download), continuity of relationship, attention control for end user with things like mute, ignore, stop, async workflow, easy initiation whenever you need something, & most critical bidirectionality by default. there is also something fundamentally pyschological at play here too about agents living with ppl on the same surface. it makes that relationship feel more natural & intimate.

it’s interesting how anthropic has the best model yet slowest to market with a personal agent that shows no clear differentiation why do you think that’s the case?

DeepSeek V4.1 Flash 1017 runs at 1,521+ tok/s on Fireworks Ultrafast. Fast enough to ship Tetris in 11 seconds, 7x faster than Claude Opus 5.5. Join the Ultrafast waitlist: bit.ly/3UA0wKq

Does anyone know what Wispr Flow is valued at?