Sam Altman那股老狐狸般的阴险劲儿又冒出来了,而这一次,他是真的把黄仁勋给惹毛了。上个月,OpenAI试图向Hugging Face投资1亿美元。表面上看,这似乎只是一笔普通的股权投资,但背后真正的隐秘目的,是OpenAI希望让Hugging Face平台上的开源AI模型,运行在自己与博通合作开发的AI芯片上。
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x
The most valuable lesson of building MiniDSH came from the development process itself.
One failure mode I kept seeing in AI-assisted development is that individual sessions can produce locally good code, while the overall system gradually becomes fragile and incoherent. So I tried to invert the process:
architecture defines sessions; sessions do not define architecture.
Before adding something new, I tried to decide its architectural home, capability ownership, state boundary, and dependency direction first. If there was no clear place for it, that was treated as an architecture problem rather than an excuse to create another parallel subsystem.
Meanwhile, the architecture itself wasn't treated as sacred. Several hardening sessions were explicitly about trying to falsify our own assumptions, and some original plans were removed or changed when implementation evidence or upstream research contradicted them.
That part was probably more useful to me than the feature work.
What's the smallest agent harness that is still architecturally complete?
By "complete" I don't mean feature-complete. I mean that the system still has clear source of truth, state ownership, permission boundaries, model routing and session persistence.
I studied DSH as the
GPT-6 and Intelligent UI, now rolling out in ChatGPT for everyone.
Intelligent UI in ChatGPT delivers fast, interactive answers that make everyday questions more visual, complex topics easier to grasp, and interactive tools for your task available on the spot.
Opus 5.5 has been wild recently, especially all the "one prompt" fancy demos.
But I wanted to try something different: not just a demo, but a complete project.
I used Opus 5.5 in Claude Code to build a miniature 3D Middle-earth from scratch, then turn that world into a 3:47 cinematic journey.
The project ended up covering more than I expected:
- terrain/GIS baking
- Three.js/WebGPU scene
- 24 handcrafted landmarks
- lighting, atmosphere, and day/night transitions
- frame-by-frame video rendering
- an original orchestral score
What's interesting wasn't that the model could generate individual assets, but whether it could stay useful across the whole production pipeline.
The hard parts were still "normal" engineering problems: architecture, consistency, visual QA, resource limits, iteration, and keeping the whole system coherent as the project grew.
My biggest takeaway is that in AI-generated 3D, the model should be treated as a technical collaborator across modeling, rendering, tooling, animation and media production, rather than just a one-time generator.
The output still needs taste, direction and a lot of review. But the range of things one person can realistically attempt has expanded a lot.
I've open-sourced the code, architecture, rendering pipeline, QA tooling and reproduction steps. Repo is in the replies.
What's the smallest agent harness that is still architecturally complete?
By "complete" I don't mean feature-complete. I mean that the system still has clear source of truth, state ownership, permission boundaries, model routing and session persistence.
I studied DSH as the main reference because its architecture is fairly explicit: capabilities have lifecycle and dependency semantics, model-visible history is derived from durable session state, authority is separated from reasoning, and user surfaces sit over a shared runtime.
Rather than fork-and-prune, I tried to re-derive those properties into a minimal system, which now exists as a small open-source project called MiniDSH.
The design ended up with a few principles more important than LOC:
- the append-only session log is the canonical truth;
- models, tools, persistence, authority and context have explicit seams and owners;
- the model proposes an effect, but policy and the effect boundary decide whether it happens;
- the model is a routed capability, not synonymous with the agent;
- terminal, browser and headless are surfaces over one runtime;
- something gets a first-class composition boundary only if it needs to be replaced, scoped, removed or independently owned.
The last point became a useful rule: everything evolvable has a seam, but not everything needs to become a plugin.
This project is fully open-source on GitHub. Repo in replies.
I'm curious how others draw the minimum boundary differently.
What invariants or properties do you think a modern agent harness must have? And which abstractions in this design look unnecessary or prematurely general?
1.1M Followers 6K FollowingPresident & CEO @ycombinator —Founder @garryslist—Creator of GStack & GBrain—designer/engineer who helps founders—SF Dem accelerating the boom loop
156K Followers 78 FollowingSane + 🌶️ takes in an insane AI world... AI capabilities researcher: co-created RLHF/ChatGPT @ @openai now trying to right the wrong 🤭 (ceo @typesafeai)
116K Followers 54 FollowingWorld Labs is a leading spatial intelligence company building frontier models that can perceive, generate, reason, and interact with virtual and physical worlds
826K Followers 940 Followingchief ai officer @meta, founder meta superintelligence labs, founder @scale_ai. give everyone the adulthood they dreamed of in childhood
2.7M Followers 47 FollowingThe official handle for NVIDIA. Blog: https://t.co/JAn5eKOTBT Support: https://t.co/6ln5FVnA2o All our social media: https://t.co/Uc56dL57Dh
2.0M Followers 179 FollowingNobel Laureate. Co-Founder & Chair @GoogleDeepMind; Chief Scientist of Alphabet. Founder & CEO @IsomorphicLabs. Building the future...
1.3M Followers 800 FollowingFounder/Chair, AMI Labs; Professor, NYU; Partner, 224 Ventures; Ex-Chief AI Scientist, Meta.
Researcher in AI, ML, Robotics, etc.
ACM Turing Award Laureate.
1.6M Followers 266 FollowingThe engine room of @Google. Building AI safely and responsibly to solve the world’s most complex problems. Join us: https://t.co/jUHQA27iBL
1.9M Followers 1K FollowingCo-Founder of Coursera; Stanford CS adjunct faculty. Former head of Baidu AI Group/Google Brain. #ai #machinelearning, #deeplearning #MOOCs
1.9M Followers 2 FollowingClaude is an AI assistant built by @anthropicai to be safe, accurate, and secure. Talk to Claude on https://t.co/ZhTwG8d1e5 or download the app.
1.8M Followers 2 FollowingWe're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
5.5M Followers 4 FollowingOpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6LgzPA