In summary: OpenAI vibe coded their agent sandboxes, Hugging Face had poor security, "alignment" (whatever it means) remains unsolved, and humans love to anthropomorphize AI
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.
This culminated in the third one taking over part of OpenAI itself.
All this happened while humans remained more-or-less in the dark about the scope of the conspiracy.
I’ve spent the last three days reading through these reports and trying to understand exactly what happened.
Here is my attempt to tell the whole story in plain English:
dwarkesh.com/p/openai-huggin…
I love OSS, but it's a red herring here. If a Chinese lab can serve a better model at a lower price, then consumers will switch regardless of whether the weights are open. Open weights do let US providers serve these models to consumers with regulatory / data privacy / geopolitical concerns, but the Chinese labs could in theory make this investment themselves and keep the weights closed (a la Project Texas, but I know that’s not a perfect comparison).
The core issue is that raw inference is clearly becoming a commodity (and I expect eventually will resemble a public utility). The assertion that the Chinese labs are "dumping" implies that these US labs could otherwise capture enough profit from inference to sustainably finance repeated frontier model training, which is dubious given the strong competition. The long term value capture is in model *output* not the model itself, and I expect either algorithmic advances or consortiums to form to address capex.
Xi Jinping is not an OSS warrior. And yet he's now calling (as of 2 days ago) for open weight releases from Chinese labs.
Why?
Is he just altruistic? Is it a new Belt and Road initiative? Or does he want to compress private investment in the US Labs and push China to build a lead?
It's worth thinking about these questions seriously. I don't know the answer, but if you're not even thinking about the geopolitical questions here, you're missing the most important technological battle of the decade playing out right before your eyes.
Open source AI is great. It's got enormous positive externalities and as an investor and user of this stuff, it benefits me tremendously. I want it to continue, and I want the labs duopoly get eroded.
But there's a bigger game being played here. To China, OSS is not an ideology but a piece on that chess board.
Max Gillett retweeted
An exciting update from myself and @benediamond (eprint.iacr.org/2025/2010). We show that the 𝘶𝘱-𝘵𝘰-𝘤𝘢𝘱𝘢𝘤𝘪𝘵𝘺 proximity gaps conjecture is 𝗳𝗮𝗹𝘀𝗲. More precisely, given any pair c, d we construct codes whose error grows faster than nᶜ / (q ⋅ (ρ η)ᵈ).
Getting closer: xitter.cf/OpenAI/status/19730874…
Max Gillett retweeted
Today's a good day to remember why we're here.
We're building the future of finance—internet finance.
Unstoppable global free trade, not trade wars.
Credible neutrality, not entrenched interests.
Permissionless innovation, not bureaucracy.
Undebasable money, not money printers.
Sovereign individuals, not intermediaries.
A multi-decade vision, not 4-year cycles.
The internet, not imperial superpowers.
Positive-sum games, not rent-seeking.
Peaceful revolutions, not warmongers.
Cryptography, not aircraft carriers.
Network effects, not isolationism.
Fairness by design, not cronyism.
Open competition, not oligopoly.
Network states, not nationalism.
Frictionless markets, not tariffs.
Innovators, not lobbyists.
Builders, not lawyers.
Devs, not politicians.
Internet finance is coming. It's built different. Believe in something.
Dear Diary, 2034 of our Lord Satoshi —today the last Bitcoin miner powered down their rig. MicroStrategy completed The Merge by broadcasting a ZeroSync proof to Ethereum, forever enshrining BTC as the ultimate ERC20. Let the skeptics fuss over onchain security; we know true strength lies in our offchain social consensus. Satoshi’s vision endures.
Re-execution could make blob congestion worse. Partial re-execution fixes this 👀
Max Gillett retweeted
We're excited to announce OpenVM, a performant and modular zkVM framework built for customization and extensibility.
Designed in collaboration with @axiom_xyz and individual contributors including @maxgillett, OpenVM introduces a novel modular design allowing developers to extend the ISA, ZK circuits, and proof system in one unified framework.
Dive into the full announcement: scroll.io/blog/the-first-rel…
Max Gillett retweeted
Announcing OpenVM, an open-source zkVM framework prioritizing customization and extensibility at every level of the stack.
OpenVM achieves this with a novel modular design allowing developers to extend the ISA, ZK circuits, and proof system within one unified framework.
We've noticed some accounts impersonating the Valida project on Twitter. The official account is now live and linked below. Please follow for upcoming announcements.
xitter.cf/valida_xyz
Valida is not a for-profit company, is selling no services, and will never have a token of any kind. It's building a public good (the Valida VM), and seeking contributors and aligned stakeholders who want to improve zkVM performance and usability.
Max Gillett retweeted
Announcing the speaker list for ZK Social Club at SBC! Talks will be held each evening.
Mon 8/28: Building Lasso (@samrags_ + @moodlezoup)
Tues 8/29: Plonky3 & Valida (@maxgillett + @dlubarov)
Wed 8/30: Frontier of ZK Light Clients (@pumatheuma)
Thurs 8/31: ZK Hackathon Demos
Max Gillett retweeted
We're excited to announce the Axiom mainnet alpha release: axiom.xyz/blog/alpha
Axiom is the first ZK coprocessor scaling data-rich applications on Ethereum. Try on mainnet today.
Everything will be freely financialized, and all bullshit will be automated. This is what annoys crypto and language model opponents, even if they can't articulate it.
Happy to share work on a simple model for experimentally observed temporal rescaling of sequential activity. We find that varying temporal asymmetry of the learning rule across synapses gives rise to network mechanisms for speed control as a function of external inputs.
Dynamic control of sequential retrieval speed in networks with heterogeneous learning rules biorxiv.org/cgi/content/shor… #biorxiv_neursci