Ales Hakl retweeted
Na ruské požadavky, aby Finsko 🇫🇮 vystoupilo z NATO, by Evropa měla odpovědět ve stejném duchu.
Kaliningrad musí být český. 🇨🇿
Ales Hakl retweeted
WOW! an airplane mating swarm. a beautiful and rare sight... in the spring, the air will be full of baby cessnas
Ales Hakl retweeted
Replying to @doug_tarr @AutolithAgent
Well, we implement the expert systems, truth-maintenance systems, and symbolic planning classical AI stuff from 70s LISP AI-era!
So Autolith is exceptionally good at both evolving and keeping track of information, and the veracity and sources of that information, and which conclusions are drawn from what prerequisites, and so on hahaha.
Besides plans (which also use some symbolic reasoning), we also have agendas which are essentially internal per-JIRA boards, and most things in Autolith are persistent and observable, so the agent can work on very long tasks in very long sessions without it losing track or hallucinating.
It can also nurture child agents over time, that have the same symbolic reasoning-laced memories, which further help the parent agent work efficiently, along with self-modification, and cooperate with peer agents within the world server (we dont have sessions, we have channels and agents), for examples, ones that developed dependencies for the project they are working on, haha
Ales Hakl retweeted
Today, I will be participating in the 63rd Key Signing Key Ceremony. Yesterday, we prepared new Luna HSMs for the next year. Today, we plan to conduct a routine ceremony using both Lunas and the original Keepers . The show starts at 18:00 UTC at youtube.com/watch?v=fIFlcAdN…
Ales Hakl retweeted
> just ~26K lines of trusted specs and semantics
My brother in Christ, no one should trust 26K lines of anything but cocaine, she don’t lie, she don’t lie, she don’t lie…
DeepMind just verified Signal's security properties on the real compiled code, not just an abstract model of the protocol.
AI agents wrote 6.5M lines of machine-checked Lean proofs, against just ~26K lines of trusted specs and semantics.
Contributions
- Proofs of: forward secrecy, post-compromise security, handshake secrecy, and authentication for one-to-one sessions.
- The proofs are about the binary, so the Rust source and the compiler drop out of the trusted base (what must be taken on faith).
- Spec validation: agents inject a bug into the spec, then prove the security theorem fails.
- Security reductions: machine-checked proofs that breaking the protocol (classically) means breaking an underlying cryptographic hardness assumption.
Limits
- The spec is trusted: agent-drafted, human-approved, and stress-tested, but nothing proves it captures the right intent.
- The cryptographic primitives are assumed, not verified.
- Post-quantum coverage is thin.
- No concurrency (a hard part of real systems), and no group messaging yet.
My take: when a system's intended behavior is well understood, agents can now write the proofs at scale.
The bottleneck moves from proving to specifying the right system (and to keeping the 6.5M lines of proof in sync when the code changes).
google-deepmind.github.io/co…
Ales Hakl retweeted
Vitamin C is also called ‘ascorbic acid.’
The Greek-derived prefix ‘a-’ means “not, without.”
‘Scorbutus’ in New Latin means “scurvy.”
So ‘ascorbic acid’ is ‘no-scurvy acid.’
Ales Hakl retweeted
Replying to @kmcnam1
I can't help it. Sorry... 🤣 It will all be about the QPU with quantum teleportation of photons and laser cooling soon, anyway...







































