The future is agents writing proof carrying code directly. No dependencies. No kernel. Your personal operating system for your specific hardware with your specific functionality. All Ring0 flat rwx memory cause faster. Ubiquitous self modifying code. Stack is a vestigial abstraction: sp is a gp now. No abstractions; deeply integrated hacks beyond human comprehension that make biology look refreshingly simple. Somewhere inside, the agent keeps rewriting everything. The formal specifications say everything is fine.

Ok, so this is happening. We'll start small with a couple rented GPUs and a single SoTA open model (e.g. Minimax-M3, Kimi-2.7-Coder) and grow from there. Will reach out to everyone who liked when we got the something useable.

Minimax-M3, Kimi-K2.7 Code, GLM5.2, DeepseekV4 Pro. So many open models announced today. I wonder what happened.

Thinking of starting a cooperative running SoTA open models on EU based GPUs. Just a couple friends leveraging batching. Selling excess capacity on OpenRouter. Privacy and zero risk of shutdown.

Short term maybe. Mid/long term it will backfire hard if US companies are not allowed to bring SOTA to the world. We’ve seen this before in other industries (e.g. cryptography or thermal imaging).

Graph doesn’t show this well, but this shows Anthropic overtaking OpenAI on math. I hope OAI retakes the lead back soon; I can’t stand Anthropic’s antics.

It honestly might be the best gift that the US government gave to Europe (and China) in a really long time. I also expect this to rollback very soon. Access to US nationals only makes no sense. It should be either everyone or selected governments/institutions.

Reminds me of Core War (1984), an old game where programs in a virtual machine fight each other to take over the machine. All sorts of crazy tactics work and there's no clear dominant strategy, but programs where simple and static. An modern LLM version of this would be fun!

Same old lie Apple told during Epic v Apple, which they lost, and were later found in contempt of court on, with a criminal referral. iPhone would be MUCH better if AI assistants could compete in a free market, rather than locking out competitors and paywalling what’s left.

/goal Replace all my phone apps with a single agent maintained personal app.

I'm going all-in on software-defined-anything, just so Codex/Hermes can manage all of it. OpenTofu, Nix(-darwin), HomeAssistant, ESPHome,... Used to be overkill for a home-setup, now it's the easiest path!

We’re entering the era of individualized apps. Dedicated app for one or a few users and their specific need. App stores are not ready for this.

Given enough matmuls, all bugs are shallow.

Great job! With increasingly abundant intelligence keeping solutions secret simply doesn’t work (it never did, but with AI especially). I hope this encourages more open research.

A useful feature of private transactions would be a zero-value transaction that externally *looks* like a large transactions. 1 USDC test transactions are becoming useless as increasingly the failing steps are triggered by larger values (which is sensible risk management).

Last week I discovered that ChatGPT and Claude will send you their “encrypted raw reasoning” and of course I immediately wasted a weekend trying to do something bad with it. What I got for my trouble was this blog post: blog.cryptographyengineering…

Thank you @GiacomoFenzi for creating the whir implementation. It’s been great fun working on it together and looking forward to continuing now that it’s audited and brought into production!

More people should know about the Interfold. It's basically what I've been yelling at people to build with the MACI ideas ( ethresear.ch/t/minimal-anti-… ) for almost a decade, and now it exists, in a generalized form. The idea is: a privacy protocol optimized for things like voting (and other use cases eg. secret-ballot auctions). The mechanism generates a threshold encryption key, and people send in their votes onchain, using a ZKP to prove eligibility. An arbitrary computation on the votes gets run inside FHE, and then threshold-decrypted. From what I can tell (the docs are good docs.theinterfold.com/CRISP/… ), it gets pretty optimal security guarantees: * Voter anonymity can be made unconditional if eligibility is proven with ZK-SNARKs * Censorship resistance is guaranteed by ethereum (votes can be posted directly onchain, and there's a proof that all posted votes are taking into account) * The correctness of the outputted result can be ensured via ZK over FHE * Liveness and coercion resistance depend on M-of-N honesty; unavoidable given present-day technology The main limitation is that today "ZK over FHE" is only properly available for additive vote tallying, as it's too expensive for computations that involve multiplication or other more complicated manipulation at the moment. There's work in progress on slashing-based / optimistic computation for such situations. (And of course ideally in the long term we'd figure out obfuscation so you can get rid of the M-of-N committees😃)

Open (from Open Source) signals creating a less restrictive freer alternative to challenge the increasingly rent-seeking incumbent. A proven strategy. Should they become top dog, the incentives flip to rent-seek and close down the ecosystem and the cycle repeats.

1/5 WIP update for our client-side Spartan-WHIR: a transparent (no trusted setup), post-quantum SNARK with 128-bit security. Repo: github.com/alxkzmn/spartan-w… Why compare against Spartan2 and ProveKit, and how we're different? 🧵👇