Ben Affleck (Hollywood star & Artists Equity CEO) talks about how he fine-tunes open video models by unfreezing weights and trained only the last cinematic layer so a film crew can hit real production standards. for context, Ben Affleck founded InterPositive in 2022, a 16-person AI shop for film post and Netflix bought it in March 2026 for $587 mn in cash. He needed that model because public video models were trained on his peers' films, and he did not think that was a real business. So InterPositive raised money, shot its own dataset for 8 months on a controlled stage, and used it only as late-stage training. Each new film then trains a private model on its own dailies, so the production keeps the footage and the learning. That is the product Netflix paid $587 million for. ---- From "Bloomberg Live" YouTube channel, (link in comment)

Robots are going to eat the next 10 years 💯 We're making robotics on the Hugging Face Hub easier to use, eg: soon you'll be able to send any community-made policy to your Microduck. Can you just imagine the fun we're going to have?

We're hosting a Jev hack/demo night + dumpling bar 🥟 this Friday! Come jam & build on open-weight Jev! 9/25 · 6:00 PM · SF

Introducing Boreal-H3 — a video model built for ads and our next step toward recursive self-improvement in video generation. A good-looking video isn’t enough. The product has to stay the same. The actor has to stay the same. The label has to be right. And the action in the brief actually has to happen. So we post-trained MiniMax H3 specifically for advertising. But this isn’t a one-off SFT or LoRA fine-tune. We built a closed-loop system that learns what to improve next. Human-calibrated evaluation diagnoses failures and guides the next intervention: targeted data collection, reinforcement learning, or inference optimization. When the feedback is unreliable, we revise the evaluator or reward—not just the generator. Every experiment feeds into shared memory, informing the next training decision. The model improves, and so does the process that produces its successor. The results: → 85.3% reference fidelity — highest among the frontier video generation models we evaluated → Brief success: 28% → 50% → Identity match: 83% → 94% → Visible defects per clip: down 70% → Generation time and estimated cost: down 20% Boreal-H3 doesn’t just make better-looking video. It makes more usable ads. Credit to the @MiniMax_AI team for the foundation we’re building on. This launch is a checkpoint, not the finish line. We’re building more than a better video model. We’re building a system that learns how to make the next one better.

THANK YOU @FeatherlessAI for the credits, I can now translate all my Persian music with Kimi-K3 without Anthropic getting in the way

We are excited to announce that World Labs is joining @AMD. The research and technical breakthroughs we have achieved since our founding in 2024 have given us a clear vision for AI’s potential to solve problems in the spatial and physical world. Accelerating the future of spatial and physical intelligence requires scaling our efforts, scaling our reach, and getting closer to the hardware.

I have successfully reverse engineered and decompiled Snood for the GBA (2001), a game originally created by @GCDaveDobson over 25 years ago, entirely driven by AI. Every byte matches (SHA-256). I used Opus, Codex, and dozens of parallel DeepSeek V4.1 agents to attack different parts of the codebase. The first step was finding the compiler, where I sent off some agents, and they discovered it was agbcc, the GCC 2.95 compiler from Nintendo's GBA SDK, and verified it by recompiling standard library functions like memcpy and getting byte identical output to the ROM. For the first week or so, matching was stuck at 3% as I had no idea what I was doing. Then one afternoon, I experimented with agent swarms and went from 7% to 81% in about 5 hours. I ran over 20 DeepSeek agents in parallel in multiple waves, each AI assigned its own function in its own sandbox, so they never stepped on each other's work. The bigger AIs merged each result only after re-verifying their work, picking and choosing what was best. The last 19% took several more days, and the last 300 bytes were very difficult compiler forensics. This work is particularly suited for agentic workflows as the reward is easily verifiable (matching rom percentages). Why Snood? I have a faint memory of playing this game at around 3 years old. I also remember hearing the video game soundtrack and wondering "How did they fit that in the cartridge?" Well now I have my answer, they used the GBA's four built-in sound channels in a sort of midi-like way to generate songs on the fly. Only the sound effects are real samples, 8-bit audio at about 11 kHz. I also wanted to choose a game that was relatively obscure, had little info, as Pokemon has been analyzed to death (and I didn't grow up playing Pokemon!) I don't have an accurate number on the cost, but it was around 4-5 billion tokens. I do feel a bit strange, as now I was no longer tell AI what to do, I was telling AI what to tell AI what to do! The knowledge I gained from this project is mostly high level about how to coordinate agents, rather than technical details (though, now that I have the code, I can ask any question and figure it out). What would've taken a team of hobbyists years took me about two weeks of on and off work. I think we're about to see a lot of forgotten technology and code get preserved, which, in the face of AI writing more and more code every day, is kinda amazing in my eyes. The future is very bright for technological preservation efforts. If you're interested in the code, let me know below. Enjoy the cutesy explanation video!

😅 this is what happens, when frontier decision models.... are starting at free