←
Ben Livshits1,782 posts
Ben Livshits avatar

Ben Livshits

@convoluted_code

A technology executive, Computer scientist, Professor. A decade at Microsoft Research, executive roles at Brave, ZKSync, Eclipse.

◷ Joined April 2015▧ 46 Photos and videos
Ben Livshits avatar
Ben Livshits@convoluted_code·Sep 22

If you would like to pursue a PhD at Imperial College London and work with me on blockchain and systems security-related topics, please contact me via my web site.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

10/ His lesson for a second attempt: generate Lean from day one, derive code from the spec, and put the specification and the trust ledger before the code. Part 2 of 3: 🔗 ben-livshits.org/blog/what-c…

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

9/ The second half is an interview with Valeriy Zamaraiev: correctness as several things at once (EELS, clients, the chain), replay as an oracle that is not exhaustive, mutation testing, JIT skepticism, Block-STM as a latency play. x.com/valeryz

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

8/ Throughput, for scale: mainnet asks for ~2.8 Mgas/s on average and 5 at the 60M-gas ceiling. Geth runs ~316 Mgas/s on Base's public benchmark, Reth ~592, Nethermind's replay ~697. Eita's 2.12 Ggas/s (1.22× revm)

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

7/ Every proof built on those lines stayed valid. That is the failure mode worth internalizing: the specification, not the proof, was the weak link.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

6/ One more: EIP-6780 deletes an account only if it was created in the same transaction. EELS marks it created before the constructor runs; the Lean spec marked it after.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

5/ Then the bugs. About a dozen specification-translation mistakes, caught not by the proof but by a differential check against the tested Lisp host. SELFDESTRUCT missing EIP-6780 from Osaka onward. A reverting constructor's revert data deployed as code.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

4/ The trusted list is where the honesty lives: the EELS-to-Lean translation, the C helpers, the crypto inside the precompiles, the SBCL runtime, the emitters' printers, and consensus, which Eita does not validate.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

3/ The useful artifact is the ledger. Every claim is proved, checked at runtime, or trusted with a name. 202 of 203 spec-relation theorems proved. 111,835 of 111,983 conformance tests pass. Axiom closure audited to the standard three.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 17

1/ An EVM specified in Lean 4, with native ARM64/x86-64 handlers emitted from the spec and every mainnet block replayed against the chain. The proof was never the hard part. The specification was. Part 2 of my series on what we can verify today. 🧵

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 11

11/ Part 1: five routes into Lean, where AI helps, seven ways it fails, four architectures worth copying. Part 2: a verified EVM. Part 3: Aristotle, the crash-freedom proofs, AlphaProof Code, and what is still open. 🔗 ben-livshits.org/blog/what-c…

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 11

10/ It still took a team of five four weeks, and by their own account most of the effort went into specifications, not proofs. AI does not dissolve the hard problems of formal methods; it relocates them.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 11

9/ The day the post went up, DeepMind announced AlphaProof Code: Signal's unmodified libsignal, with forward secrecy and post-compromise security proved about the AArch64 binary in ~6.5M lines of Lean. No human-written proof code.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 11

8/ And a proof against a bad spec is wrong with a certificate. Say sort returns an ordered array, forget the permutation, and returning [] is fully verified. Two peer-reviewed Lean benchmarks had 16 and 18 defective human-written reference specs.

Ben Livshits avatar
Ben Livshits@convoluted_code·Nov 11

7/ Where models earn their keep: falsifying before proving, decomposing hard goals, guessing invariants, repairing from diagnostics. Where they fail: on 241 unproved seL4 theorems, 60.1% of failed steps cited a lemma that does not exist.