Key Takeaways
- Hugging Face released MiniMax H3, a 33B-parameter open-weight model for text/image/reference video generation with audio, runnable on consumer GPUs via diffusers and ComfyUI.
- Aravind Srinivas explains DeepSeek’s low API prices with a much smaller model footprint: roughly 5–10x smaller than rival models and about 40x more traffic per unit of compute.
- Greg Brockman showcases Codex in production: a skill converts customer feedback into roadmap items, and an automated ad campaign asks permission before any payment.
- swyx’s Opus 5 probe generated 5,500 lines of runnable custom-world code, but exposed current multimodal limits in native video/game perception.
- Yann LeCun argues for accelerating AI defense rather than slowing AI, while developers push back on generic AI prose with a “deslopify” command.
1. AI Models and System Improvements
- Hugging Face released MiniMax H3, a 33B-parameter open-weight model that supports text, image, and reference-based video generation, all with audio. It can run on consumer GPUs through diffusers and ComfyUI, with weights and an app available. — via 1
- Hugging Face also spotlighted llama-macos, a utility that recommends models based on Mac performance and context window, then launches a llama server and WebUI with one command. It supports brew installation for easy switching between open models. — via 1
- NVIDIA AI’s Hermes Agent, improved with NeMo Relay and additional strategies, shows major efficiency gains for small/weak/local models. The optimizations cover tool execution time, memory, task steps, schema, and token efficiency, and are now live. — via 1
- Ethan Mollick’s shader-test comparison indicates model tiers are still shifting: Qwen 3.8 Max is solid but below Kimi K3, while Kimi K3 is a very good open-weight model yet still not at Sol Max or Fable level. — via 1
2. AI Economics and Business Workflows
- Aravind Srinivas explains DeepSeek’s API pricing economics: models are far smaller than comparable rivals (10x smaller than Opus, 5x smaller than Sonnet), cutting compute needs; single chips can handle multiple requests, and the same compute can process about 40x more traffic, allowing profit at low prices. — via 1
- Codex is moving into real business workflows. Greg Brockman shared a skill that converts customer feedback into product roadmap items and an automated ad campaign that asks for permission before spending money. Separately, swyx described using Codex’s
@thread queue to keep work moving when platform features are blocked, though a multi-agent harness would be preferable. — via 1 2 3 - Runway says registering for a new Max plan will grant 7 days of unlimited access to Seedance 2.5 when it launches. — via 1
- swyx notes that machine-generated code makes open-source software personalizable by machines, reducing the appeal of closed-source tools like Claude Code. — via 1
3. AI Capabilities, Limits, and Open Research
- swyx’s Opus 5 test fed the opening of The Lord of the Rings plus roughly a million-token budget into the model. It took nearly two hours to write 5,500 lines of runnable code for a procedural rendered story; the result was rough but functional. The test shows LLMs can build hyper-custom worlds people wouldn’t manually create, but also exposes limits: models can’t natively watch video or operate inside games and must rely on slow screenshot-based inspection. — via 1
- Ethan Mollick argues that many fields beyond math hold important, empirically solvable questions, and a sufficiently powerful AI could create large social value. In entrepreneurship research alone, he lists questions such as what causes startup success, whether high growth can be predicted, which ideas/people/actions matter, what can be taught, and how to activate ecosystems with minimal intervention. — via 1
4. AI Safety, Policy, and Culture
- Yann LeCun argues AI cyberattacks should not slow AI development; instead, the field should accelerate. He supports mandatory trace sharing and incident disclosure, keeping AI cyberattacks illegal with heavier penalties, and using AI—especially open models—to strengthen defense, concluding that AI will make cyberspace safer. — via 1
- LeCun also highlighted the contrast in AI messaging: Chinese AI companies emphasize benefits and returning time to people, while US AI media emphasize lost jobs, AI “eating children,” and loss of control. — via 1
- AI writing style remains a pain point. Ben Tossell asks how to stop agents from inserting filler like “this is probably my favorite paragraph,” while Hamel Husain added a “deslopify” command so his agent can rewrite responses into cleaner prose. — via 1 2
