Inside you there are two wolves. One builds trustless systems, permissionless, verifiable, no privileged operators. The other slaps "decentralized" on a multisig with 3 keys held by the same team. The one that wins is the one you fund.

Apropos: "The End of Consensus Research?" (arxiv.org/pdf/2610.30114). AI is (for now) great for the heavy lifting, but human experts are still entirely essential.

here's how we have designed our archival state to scale to Giga throughput. lsm trees handle the massive load we need, but pruning was a challenge. we fixed it using undo logs and buckets with bloom filters, making pruning close to O(1), going from hours to ms. a very different approach with respect to other clients, that index changes by block height. maintaining these indexes requires reads in the write path, that were also a limiting factor. no reads in the write path = higher throughput blocks organised in buckets = O(1) prunning

SHA-256 is not broken, but its security margin lost 1 more round this year: Last week I found a collision attack on reduced 39-round that improves prior art by over 30 bits and the first differential attack on that round better than generic search on @yukonresearch Hashsmash board. The attack rests on the same method as prior art ePrint 2026/1120 with new insights that first 2 state words could remain free, allowing a critical condition at step 16 rebuilt as a claw-search problem rather than a probabilistic filter. Took days of watching SAT logs, burning cores and arguing with Claude on what's possible to get a result close to generic search. Since my initial submission, other hashsmash participants sraiw, @cyberaxe, @privacymage, @lamb356_ and @geeknik pushed it down by over 6.5 bits, currently estimated at 118.15 bits. The tree of submissions is hard to follow, so I put the construction, insights and numbers into a blog and am working on detailed technical report. It's fascinating to see an initial structural observation develop into an improvement in the reduced-round collision frontier within days. AI is now a real aid in cryptanalysis and open collaboration speeds it up.

The internet as we know it may change forever soon. We used to have rate-limitting on complex interactive client-server interactions since humans were required through it. This is no longer the case: Banks already can’t tell a human from an agent in their own systems. In the near future, most open endpoints may require every request to be backed by a real human. Otherwise agent swarms could make the service unusable. This is a huge privacy and anonymization risk. But you don’t need to reveal identities necessarily. Scoped nullifiers + ZK proof-of-personhood may let a server verify “this unique human has already used their quota on this resource” without learning the identity or linking the request to any other context. The nullifier is kinda hard: the thing that stops multi-agent reuse of one human (or simple DDoS via shared access) is also a correlation handle. Too coarse a scope, observable timing, or graph side-channels and deanonymization comes back. Delegated agent credentials with tightly limited nullifier budgets seem necessary. The “humans + a few of their agents” internet starts looking a lot like a giant rate-limited anonymous membership set.

Incredibly proud of the @Sei_Labs engineering team for pulling this off. Last week the team successfully executed the biggest migration on Sei in years, with perfect chain performance throughout. There is no team I would rather have building Sei Giga.

Victor Shoup’s new paper (eprint.iacr.org/2026/2338) shows a neat Plotkin attack on Raptor codes (R10/RaptorQ) in Asynchronous Verifiable Information Dispersal (AVID). A few people asked if this affects Sedna (arxiv.org/abs/2512.17045). Short answer: no, because of when decoding happens. In contrast, this may have been the issue that led early Walrus to switch from RaptorQ to Reed-Solomon before mainnet, and the reason why RaptorQ would not work for Walrus. In reconstruct-later AVID, the network issues an availability certificate from 2f+1 attesters before anyone decodes. A malicious disperser can see which 2f+1 honest nodes hold shards, algebraically cherry-pick a rank-deficient subset of f+1 honest shards (which Shoup proves always exists), pair them with f Byzantine signatures, and withhold the rest. The data is certified available, but mathematically unrecoverable. Sedna never certifies undecoded data. Validators finalize verified symbol bundles in MCP blocks first, and a tx is only included once its finalized symbols actually decode and the commitment opens. If a sender submits a rank-deficient set, decoding simply fails. The tx is not included, and the attacker just burns bandwidth fees.

A DT series with Gilad Stern on approximate agreement! Unlike consensus, which requires exact agreement, approximate agreement takes rational inputs and allows outputs to differ by at most a small parameter ε Join us for a modern view on research gems from the ’80s and ’90s!

It’s a wrap! It was an honor to help organize the Fast Finality workshop in Cambridge. There have already been some great recaps, but here’s my quick overview of the main topics we covered: 1. Deep-diving DC: Decoupled Consensus is a super ambitious upgrade we hope to ship in I* (h/t @benjaminion_xyz). @robsaltini gave a great overview, and @fradamt unveiled his impressive recent breakthroughs, e.g. achieving short-term async resilience through local locks. 2. Shipping DC: Sessions on critical research and engineering, including networking benchmarks & devnets (@sukunrt & Csaba), rewards & incentives (@weboftrees), and specs & testing (@mkalinin2, @jtraglia, @tcoratger). 3. Securing DC: Formal verification meets fast finality. We explored exciting new AI-enabled opportunities (@fradamt & @robsaltini), building on years of great FV work (@alexanderlhicks). 4. Future of Fast Finality: Aligning DC with future needs. This covered PQ networking (@kamil_abiy & @kevaundray), finality in tens of seconds with PQ validators (@barnabemonnot), lean clients & prototyping (@pablodeymo, Mega, Mihir, Locky, @adust09, @ladislaus0x), and one-round finality & other future research (shoutout @alranpe, @jneu_net, @ittai & Neil). Huge thanks to everyone who made this workshop a success! Special thanks to @benjaminion_xyz for all the help and guidance, and of course to @justindrake and @willcorcoran for making it possible!