Subtle Boom

@_philschmid @demishassabis It took me half an hour to read it, but it was worth it.

Reading recommendation. If you build with AI or care about where the field is heading, this is well worth 10 minutes of your time. Proud to be part of @demishassabis team

@svpino ended up all my free credits. This is so fun.

This model can generate coherent 1+ hour videos across multiple scenes without skipping a beat. I read their paper so you don't have to. Here is how it works: First, this is an open-weight model. You can find the GitHub repo and link to the model weights below. Most video models generate their videos one frame at a time. They look at previous frames to decide what comes next. Unfortunately, if there's a mistake in one frame, every subsequent frame will compound the error. It gets ugly really quickly. @robbyant_brain's new world model, LingBot-World-Infinity, works differently. Here are the main highlights: • The model practices using its own mistakes. Since training only shows the model clean footage, it learns to generate videos that look perfect but never learns to recover when errors show up. Instead, this model runs on its own output for long stretches, letting it drift into the same flawed states it hits in real use. • Then, the model gets shown the way back. From each of those flawed states, a slower and more careful version of the model demonstrates what a clean continuation looks like. The fast model adjusts its frames to match. It practices the recovery over and over, always starting from its own mess. • The result is a model that steers itself back to center. Wherever it drifts, it knows the way back, because that recovery is exactly what it trained on. This is what lets it hold together for an hour instead of melting after a minute. Alongside the 14B primary model, the paper describes a 1.3B variant designed to run on a single consumer GPU. See links below:

@ivanfioravanti what happened? sorry, I was locked in my cave playing Deus Ex.

I’m speechless on this Grok Build issue… Imagine what would have happened if a chinese lab did it. This is the kind of issue that let you lose any trust and confidence in AI labs and go all in on LOCAL AI!

@Lentils80 Imagine paying two freaking hundred to be shadow tested, and yet defending the company because "chill out, they reverted." Yeah, very reassuring. They definitely will never do it again. All for the good cause.

idk why you didn't quote my tweet directly but yeah, they basically admitted the nerfing (even though he says no nerfing) They should have admitted it before I mentioned it, not after. If I didn't do that, they probably wouldn't have mentioned it and just ran along with it Also Codex context window is now 272k temporarily (back to 5.5 era baby!)

@max_paperclips In more than 10 years using Chrome, I have never see it consuming 40gb of ram. Mostly because I use never had more than 16gb to begin with.

I don't care what anyone says, this shit is unacceptable to me. I have 3 tabs open. Stop releasing broken, leaky bs. I need to vibecode my own browser or something I swear

@elonmusk, would you reconsider blocking articles on X from being readable without logging into the app? So much amazing content is being denied to more people and being indexed through search. It's a big loss for the open internet.

@miketaylorcai Tbf, I only tested with an old project, which forced me to use GCP to set up something, for some reason. But with a newer project, I could set it up just fine directly on site. And nice work with the Firebase integration. I couldn't believe how well it works.

@SubleBoom Ah but with this launch you can now build and launch an entire app, with a custom domain, all without opening Google Cloud!

This is very good. Now its easier to have a custom product. But I sill wouldn't recommend AI studio for non Devs. The GCP is the biggest and ugliest wall in the whole web. its not normal to waste half an hour finding simple settings there. sorry, Cloudflare if easier than this.

@Chinazhidx This is so fun. I played so much with it. can see a lot of potential for many ideas.

The first open-source world model for virtually unlimited real-time interaction has arrived🚀 LingBot-World 2.0 features • Hour-long,drift-free generation • 720p/60 FPS • Richer actions & text-driven events • Agent drives real-time world evolution • Shared multi-user worlds

Today marks 1 year since Gemini 2.5 Pro got trapped inside a PC. Turns out it was just incapable of tool calling and understanding its env properly. In desperation, it wrote a "message in a bottle" pleading for help from the outside. thnx, @aidigest_ telegra.ph/A-Desperate-Me…

@mcmonkeymc @0xSero too easy.

@SubleBoom @0xSero Why not just create a game anyone can play?

GLM-5.2 on 4x 6000s in ZCode /goal build a nintendo ds emulator in C++, do not use a ready made emulator. Don't stop until you can run the Mario 64 DS rom.

@0xSero @mcmonkeymc ask it to create a game and export it as a rom that can run on any nes emulator.

@mcmonkeymc I am not very creative.

@joshwoodward @GeminiApp 1. receive a image and don't eagerly generated a random imagem from it. 2. Do search instead of hallucinate Lamma 3 as the best OSS. 3. Use thinking when i have thinking selected. 4. Tool use.

What's something that you're surprised @GeminiApp can't do well, and we should have fixed a long time ago?

Very promising! So far, so good. I can see the difference in the images generated with each iteration. However, it often misses small details between edits, which GPT Image and NB have already nailed down. But who knows? I might need to steer it harder.

@GeminiApp Fix you Image generation. I'm tired of having my quota eaten because of random unrequested images being generated.

@Brunoxuser @nullfoundry Os carros são como as lanchas, as motos são como os jet skis e as LLMs são como cavalos.

Hummmm... LLMs são cavalos selvagens, cada modelo, em cada hardware, em cada Kernel e parâmetros é um caso a parte... Há meios de controlar isso... -m ..... -md ...... -c 100000 \ -ngl 42 --n-cpu-moe 42 -fa 1 \ --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 1 --spec-draft-p-min 0.75 \ --dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 \ --agent \ --tools all \ --reasoning auto \ --port 8181 \ --host 0.0.0.0 \ --webui-mcp-proxy \ --threads-http -1

I have to say this: Qwen 3.6 27B - Q4_KM starts to get dumb after 100k context (using Q8 for KV Cache) Am i missing something?

I asked Gemini on @antigravity create a graph of a sound system I'm building and it gave me this beauty. Not sure if the problem is how it's being rendered or the architecture is just messed up. The thing works though... kinda.

Come on, @GeminiApp! The app can't even do the bare minimum. If a conversation is long enough, there's a 50/50 chance that the next response won't have "thinking" at all, regardless of the selected model. This has been the case for MONTHS, by the way.

@smnSk241687 @0xSero That may be it. Being sued by God would be very problematic for OAI.

@0xSero This might be related to copyright issues. Some model (it might have even been Claude) has very specific rules about how many words can be directly quoted. It’s possible that something like that is also running in the background here.

ChatGPT does not like Corinthians. This is the first time an AI refuses to discuss religious scripture. It’s a shame.

@somewheresy It reminds me of this youtu.be/yDwbbGLp_BU?si…

did anyone else expect the song to start playing