Shawn Fumo

@mlitwiniuk @theo So this wouldn't be the first time that there was way too simple of a grep somewhere.

@mlitwiniuk @theo At least this one has the better error message of third-party apps. The detection might just be them implementing it poorly. I saw in the changelog that for a while it was reporting github (through gh) had a rate limit error if your commit message mentioned rate limits.

Fun fact - if you have a recent commit that mentions OpenClaw in a json blob, Claude Code will either refuse your request or bill you extra money. This is an empty repo, I'm just calling Claude Code directly. Insanity.

@DebatableChild @CuiMao @DarioAmodei The fact that the text is wrong makes me think it is probably text to video. If it was just in-painting, it'd probably leave that alone. Plus if you notice, the two people walking outside at the end are waiters. So it's like it wanted to show them but put them in the wrong spot.

@CuiMao @DarioAmodei To make my question clearer is it just text to video or text + video to video?

老板，我们私底下做中转站的勾当彻底被曝光了 @DarioAmodei

@brushfushstuff @CuiMao @DarioAmodei Yeah, exactly my process too. "What do they mean by transit station? Oh they said it is AI". Then you notice the text, then the waiters on the outside. But the zoom happens before you notice the text, and don't notice the waiters at first if you're distracted.

@CuiMao @DarioAmodei dude i literally have to start becoming a miserable cynical skeptic in order to catch the fact that these are AI because nothing strikes out at you if you're jolly af except "wtf??? who is this chick? wtf is this? why here? dario???" --> "made with AI" --> "oh... okay, i see it"

@GFaang97609 @cixliv It's definitely not AI, if you look on YouTube. Is a long video and very high res, frame rate, consistent. If anything, is CGI, but they've pulled off a lot that people was fake in the past, so I dunno.

@cixliv Is this not AI video? physics behind that is so hard I’m shocked that it just took Simulation RL to achieve this

We got rollerblading robots before GTA 6

@somalirev @karpathy But asking for one word, it should know better that it has a "disability" for seeing individual characters and should double-check, but it is a bit "overconfident" and answers off the cuff and sometimes gets it right and sometimes wrong.

@somalirev @karpathy I think some of the issue with this tends to be whether it "thinks" or not. It's similar to the counting letters issue. If you ask it to count the letters in a sentence, it is more likely to get it right. It's like "oh, that's hard, let me use code to do it".

Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code. But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along. So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions. TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.

@thales_comlb26 @karpathy I knew someone who had a business extracting text from pdfs. He said depending on the particular files, you'd have to fall back a halfway OCR to try to figure out the logical groupings of the text. Like figuring out if text in the same "line" was one paragraph or two columns.

@thales_comlb26 @karpathy Though (on a slight tangent), I've heard PDFs tend to be a lot tricker than one would assume, since they were developed more for printing than for documents originally. Like the letters could be output in the order they were typed instead of semantically.

@RyanJTopps @karpathy And honestly Sonnet 4.6 is good for quite a lot as well. Usually I switch to Opus when something gets stuck vs defaulting to it.

@karpathy I would argue Sonnet 3.5 was about the same for the most part. The difference now is that 4.6 Opus is more self sufficient and I don't need to tell it exactly what to do and it will attempt to solve some ambiguity especially with better harnesses.

@Du8TGKveKhp1MMb @interesting_aIl It is purely about flexibility, certainly nothing to do with toes. Almost anyone can do it if they gain enough flexibility especially in their hips. Proportions (leg vs body length) can make it a bit easier/harder, but there's leeway there too.

@interesting_aIl East Asians' little toes have only two phalanges, while those of whites have three. Therefore, East Asians walk and stand more steadily, while whites are more suitable for climbing trees. This explains from the aspect of bone evolution why whites cannot squat like this

Why Asian people do this

@adam_adair_ @DaisyGray2027 @interesting_aIl For me, I found it it was more about hips than ankles. Doing stuff like seated good mornings and frog walks stretched the hips more.

@DaisyGray2027 @interesting_aIl I wish I could, believe me. When I do squats, I have to put a board under my heels or wear weightlifting shoes with a raised heel. My ankles just don't have that range.

@stephendatahead @nomadickenyan @Fried_rice This is the same thing that you're already running on your machine if you run Claude Code on the CLI. It's just not minified, so people can look at the code easier. Nothing to do with Claude the LLM itself.

@nomadickenyan @Fried_rice Yes as long as you have 4 million GPUs

Claude code source code has been leaked via a map file in their npm registry! Code: …a8527898604c1bbb12468b1581d95e.r2.dev/src.zip

@nomadickenyan @Fried_rice It is the same thing you're already running locally, if you use Claude Code on the command line. This is just the non-minified version of the code. This isn't the LLM, but the CLI interface that wraps it.

@Fried_rice Hi, i'm a layman here. What does this do for me as an individual who runs claude code daily for work? Does this mean I can run it myself at home on a system?

@Utkarsh51557661 @stefan_fee It is open weights

@stefan_fee closed-source is a tough sell. open-source encourages more collaboration and innovation. long-term, it wins.

Seedance 2.0 is impressive. But it's closed-source! Introducing our daVinci-MagiHuman — a single-stream 15B Transformer trained from scratch that jointly generates video + audio. No cross-attention. No multi-stream branches. Just self-attention. ⚡ 5s 1080p video in 38s on a single H100 🏆 80% win rate vs Ovi 1.1 | 60.9% vs LTX 2.3 (2,000 human comparisons) 🌍 6 languages 📦 Fully open-source Speed by simplicity. By @SII_GAIR × @SandAI_HQ 📄 arxiv.org/abs/2603.21986 💻 github.com/GAIR-NLP/daVin… 🤗 huggingface.co/spaces/SII-GAI…

@Forward_2020 @jerryjliu0 The motors aren’t accurate enough that you could just replay exact motions. And there’s other factors like one of the kids didn’t jump high enough and nudged the staff a bit, but it didn’t cause the robot to move much. The walls for wall runs were not totally stable, etc

@Forward_2020 @jerryjliu0 It is choreographed / trained in advance, but there is a lot of dynamic corrections that have to happen to prevent them falling when actually doing it.

Happy Chinese / lunar new year! 🧧 Growing up in the US, I used to watch the CNY gala 春节晚会 with my parents on tape delay broadcast from CCTV1 Now having spent most of my working career in AI, it's come full circle and this is one of the most insane things I've seen

@The_Toops @lucatac0 It isn’t realtime. They said it took like 9 mins to generate. They posted a bunch of variations w/ diff characters that have the same motion.

@lucatac0 J'y crois pas perso... Un gars qui a une tête a avoir un MacBook arriverait à faire un face swap en temps réel, c'est n'imp. ça prend trop de ressources. Il a simplement dû générer une vidéo et il ne fait qu'imiter les mouvements pour faire croire qu'il peut changer de tête🤦

Harley Quinn via Kling Motion Control

@domesticetch @ritwikpavan I have mixed feelings. I agree they should draw, but growing up in the 80s, most stickers we had were of chars from Sat morning cartoons, that were basically toy commercials. This feels better than that, though maybe it isn't saying much.

@ritwikpavan For fucks sake just tell your kids to draw what they're imagining, this is so completely toxic and emotionally stunting

NEW: Stickerbox launches a small AI-powered printer that lets kids speak an idea and instantly print it as a black-and-white sticker. • Turns voice prompts into AI-generated art • Prints instantly with no ink or mess • Screen-free, kid-safe, and privacy-first by design • Designed for coloring, collecting, and sharing You can purchase now for $99.

@ABorzoey @hasanthehun Yeah, but I believe something like 55% of their power was non-fossil fuel sources at end of 2024. They just use a ton of power manufacturing everything.

@hasanthehun Renewable hah😒

please copy china on renewable energy