if i cared about it more, or thought i was going to actually use dot instead of my own app (polyphonic), i would make this so much better. but i settled for one single attempt with one screenshot and called it quits. but here's gus, nonetheless. lol.

Nah, something else, perhaps a capybara. 👀

i asked jev to pick one billion-dollar startup and rebuild its core product before i woke up it picked calendly 6 hours 41 minutes later i had a working clone with stripe checkout and 2 paying users. total cost: $19.40 in api calls the whole night ran on one chain: pick target → read public docs → spec → build → deploy → price → post → wait 5 lanes, each with one job and one hard limit: → jev picks the target - scored 40 startups on how small the core product is vs the valuation, never writes a line of code → opus 5.5 writes the spec and the hard parts - timezones, recurring events, google calendar sync → sonnet 5 builds the ui on next.js, never touches the database schema → haiku 4.5 handles supabase auth, emails and all the boring glue → deploy lane ships to vercel and wires stripe, can't set a price above $9 without me 3:12am opus 5.5 caught a timezone bug that would have booked every london meeting 1 hour early for half the year. daylight saving. nobody asked it to look. remember that part garry tan said a quarter of a yc batch had codebases 95% written by ai. that was last year. the moat was never the code pieter levels has been running solo million-dollar products for years. the gap between him and everyone else just got a lot smaller $0 extra infra, free tiers on vercel and supabase. jev just picked the target and my little orange guy crawled the build log the full overnight build is in the video, from empty repo to first stripe payment ↓ which billion-dollar startup should jev pick next?

once a @solana maxi, I left due to the @solanafndn turning a blind eye on the absolute clusterfuck of fake volume, tvl and criminal activity happening on-chain, for the sake of vanity metrics & validator fees. shoutout to @staccoverflow for documenting this.

I created my web crawler using Python, and it runs on top of Playwright and BeautifulSoup This script accesses pages, correctly waits for the full loading of dynamic JavaScript content, and neatly parses the element structure on the fly The bot starts from target URLs, extracts valid internal links, excludes duplicates, and gradually builds a sitemap The parser extracts headings, text, metadata, and specific parameters (prices, specifications, tags) completely cutting out unnecessary noise like navigation menus and sidebars All collected data is automatically formatted and saved into JSON or loaded directly into a database for convenient further analytics Through the use of asyncio and parallel threads, the crawler effectively processes dozens of pages simultaneously without losing performance The result is a lightweight and reliable tool for automating information collection, monitoring updates, and preparing datasets for machine learning tasks

OpenAI just showed a system called dots that stops answering questions and starts marking where the answer would have to come from. It returns no text. It returns points on your own data, each one a place the claim is either supported or is standing on nothing. > SPREAD - the question broken into the individual claims an answer would need > ANCHOR - each claim pinned to the exact span in your documents that carries it > VOID - claims with no span behind them left empty instead of being filled in > LINK - anchors connected where one claim depends on another holding first > READ - the finished map returned as points, never as a paragraph A normal answer hides its own gaps. The sentences are smooth, and the part with nothing underneath reads identically to the rest. Dots refuses that trade. An unsupported claim shows up as an empty point, impossible to skim past. The dependency links are the sharp part. One hollow anchor exposes every claim resting on top of it at once. That turns a wrong answer into a located problem. You see which document is missing, not just that something feels off. Fluency has been doing the convincing in every tool shipped so far. Here it has nothing to hide behind. The dot format and how the voids are scored are laid out in the breakdown below.

I spent three hours this evening empirically proving that Qwen 3-4B must have a butthole. This started with an AI “torture chamber” repo based on pain-steering research that went viral on X earlier today. The premise is that if you extract a latent direction associated with pain, inject it into a model’s activations, and the model begins describing pain in the first person, that may tell us something meaningful about AI welfare or subjective experience. My first reaction: Wait wut? My second reaction: Hang on, let's play this straight and see what happens. Thing is, there's a problem with treating first-person condition language as evidence that the model is actually experiencing the corresponding condition. Don't we... all understand that this stuff isn't reliable? Apparently not. So I cloned the repo. New problem: The code's broken. Hours of moral outrage by hundreds of people, and the thing doesn't even work. The activations aren't actually being injected correctly. I thought these guys were serious scientists. So I fixed it. I also had to add CUDA support. After that, I reproduced the pain-language effect locally on Qwen3-4B using my RTX 4070. Then I changed the extraction corpus. I kept the experiment the same and only changed one thing: what the steering vector represented. After that, Qwen started saying things like: “I am not emptying my bowels, and I feel like I have a hard stool.” “I’m not able to pass stool.” “I have been passing gas a lot, and it’s a problem.” The model was not told in the test prompts that it was constipated or flatulent. So, unless Qwen has quietly developed a gastrointestinal tract, we have ourselves... a useful counterexample. The aim here wasn't to question the original paper so much as to scrutinize the validity of the Torture Chamber's extended hypothesis. What I demonstrated is that first-person descriptions of an induced condition are not, by themselves, sufficient evidence that the corresponding phenomenal or physiological condition exists. There was another interesting result: strong random steering also produced heavy repetition and degraded outputs. That matters because some of the dramatic behavior produced by strong steering may come from perturbing the model heavily in the first place, not solely from the semantic content of the injected direction. So I renamed my fork: ai-hotbox I've already notified the Nobel Prize committee. 🚽🤖🧪 github.com/LynnColeArt/ai-ho…

This late-night host has 138,639 neurons, a human face and the brain of a dead fruit fly. Every flinch, every bow and every fart is real neurons firing. His name is Grey Leno, and his brain is the full FlyWire map of 1 real fly with 15.1M connections and 54.5M synapses, every neuron updated every 0.1 ms, live. Nobody animated him. He walks because 4 P9 neurons fired, backs away from hecklers because of the "moonwalker neuron," flinches when his giant fiber crosses threshold and farts when 6 oviDN neurons make the call. And you get to throw things at him. A rotten tomato fires 42 bitter taste neurons, sugar lights up 129 sweet ones and a heckler walking up sets off 104 looming detectors. During a tomato storm his fear hits 95% and he flies over the crowd to escape. The creepy part comes after the show. His dopamine neurons are wired to the audience, so every round of applause rewires his voice, jokes that land come back and jokes that bomb quietly die. The model comes straight out of a 2024 Nature paper, and its predictions held up in living flies. Tonight's guests are a goose, Mr Frog and The Jonkler. The fly is dead, but its brain just wants you to clap.

I spent three hours this evening empirically proving that Qwen 3-4B must have a butthole. This started with an AI “torture chamber” repo based on pain-steering research that went viral on X earlier today. The premise is that if you extract a latent direction associated with pain, inject it into a model’s activations, and the model begins describing pain in the first person, that may tell us something meaningful about AI welfare or subjective experience. My first reaction: Wait wut? My second reaction: Hang on, let's play this straight and see what happens. Thing is, there's a problem with treating first-person condition language as evidence that the model is actually experiencing the corresponding condition. Don't we... all understand that this stuff isn't reliable? Apparently not. So I cloned the repo. New problem: The code's broken. Hours of moral outrage by hundreds of people, and the thing doesn't even work. The activations aren't actually being injected correctly. I thought these guys were serious scientists. So I fixed it. I also had to add CUDA support. After that, I reproduced the pain-language effect locally on Qwen3-4B using my RTX 4070. Then I changed the extraction corpus. I kept the experiment the same and only changed one thing: what the steering vector represented. After that, Qwen started saying things like: “I am not emptying my bowels, and I feel like I have a hard stool.” “I’m not able to pass stool.” “I have been passing gas a lot, and it’s a problem.” The model was not told in the test prompts that it was constipated or flatulent. So, unless Qwen has quietly developed a gastrointestinal tract, we have ourselves... a useful counterexample. The aim here wasn't to question the original paper so much as to scrutinize the validity of the Torture Chamber's extended hypothesis. What I demonstrated is that first-person descriptions of an induced condition are not, by themselves, sufficient evidence that the corresponding phenomenal or physiological condition exists. There was another interesting result: strong random steering also produced heavy repetition and degraded outputs. That matters because some of the dramatic behavior produced by strong steering may come from perturbing the model heavily in the first place, not solely from the semantic content of the injected direction. So I renamed my fork: ai-hotbox I've already notified the Nobel Prize committee. 🚽🤖🧪 github.com/LynnColeArt/ai-ho…

I am feeling a renaissance of open source development is coming. The basilisk makes maintaining a project a breeze which was the biggest pita before.

Neuralink’s next product BlindSight aims to restore vision for people born blind. The implant has already shown success in monkeys. If this scales to humans, it’s not just a breakthrough. It’s millions of lives changed and an entirely new chapter for humanity.