Knowledge Graphs make rare, meaningful relationships visible before they become obvious. In a nutshell, an edge or absense of it between two edges that represent things in our world is a signal. Combining cross-domain datasets into one graph can often given surprising signals. I'm explaning what we do at @dyor_cloud in an article, read here: medium.com/@valeriyzamaraiev…

Btw the Redis story repeats itself: I'm working at DwarfStar for free for the community and because I enjoy it. But I'm receiving criticisms, since people are worried that this will break their AI-richness plans. I want to say to everybody thinking that I should stop that each time you tell me this, I'll double down my efforts towards a completely no profit engine for local inference. Better to shut up basically.

Goodbye, Poseidon! An epic 8-year, 8-figure rabbit hole in post-quantum cryptography reaches its dream conclusion. The Ethereum Foundation is abandoning Poseidon for L1, pivoting to SHA or BLAKE. This milestone unlocks ultimate security for lean Ethereum and foreshadows a golden era of hash-based cryptography. Since 2018, the Ethereum Foundation has invested in magic cryptographic bricks, so-called "SNARK-friendly hashes". In 2019, Poseidon was born. It held strong and became the dominant SNARK-friendly hash, securing billions via zkrollups and zkVMs. In a stunning reversal, breakthrough SNARK designs show that SNARK-friendly hashes aren't necessary after all. Off-the-shelf traditional hash functions like SHA2 and BLAKE2s can now match Poseidon in a SNARK. In hindsight the key was not SNARK-friendly hashes, but hash-friendly SNARKs. The secret is doing maths over the smallest prime number: 2. So-called "binary fields" natively speak the language of bits, aligning with the boolean operations inside traditional hashes. This is a stark departure from "prime fields", where awkward large-prime arithmetic makes bit manipulation painfully expensive. We're talking sci-fi cryptography. 1M traditional hash calls proven per second, on a laptop. Just 100x overhead vs native CPU boolean compute. Nobody predicted such performance, not even the handful of binary-field visionaries. Hat tip to the research geniuses: Jim and Ben with Binius in 2023; Ron, Benedikt and William with Flock in June. With SHA2, the lean aesthetic of minimal assumptions reaches its climax. The EF's principled stance on pure hash-based cryptography has aged like fine wine. We now enjoy foundations the world can trust for decades and centuries, foundations worthy of the dream of an internet of value. Speed of deployment is a secondary win. There's no longer a need to wait years for Poseidon cryptanalysis to bake. Emile and Thomas from the EF post-quantum team are moving at breakneck speed with binary fields. The strawmap now points to a production-grade leanVM in 2027, with CL, DL, EL deployments in 2028. As AI becomes exceptional at cryptanalysis, the contrarian bet to avoid riskier structures like lattices and isogenies is visibly paying off. The past weeks have been brutal. Lattice-based "HAWK" and isogeny-based "SQIsign", both signature schemes in NIST's Round 3, have suffered blows. Sources I trust say more blood is coming. On AI, the open autoresearch trend kicked off by ECDSA[.]fail is spreading fast, with amazing outcomes from zk[.]golf and SNARK[.]fast. Days ago SNARK[.]fast crossed 1.8M BLAKE3/sec proven on an M3 Max. Stay tuned for fresh autoresearch challenges dropping tomorrow. Also tomorrow: Ethproofs call #10, dedicated to binary fields. Possibly the most noteworthy Ethproofs call yet. Experts leading the charge will present the future of hash-based SNARKs at 2pm UTC. What an incredible time to be alive. To witness history, DM me for a calendar invite :) Today I can confidently claim that hash-based cryptography has won out for blockchain post-quantum signatures. SNARK succinctness compresses arbitrarily many signatures into one small proof per block. SNARK flexibility yields k-of-n threshold signatures, complex multisigs, and more. Ultimate security. Uncompromising performance. Full programmability. Believe in something. Believe in hashes.

Most recently I worked on Knowledge Graphs and on applying formal methods to AI-driven software engineering. At first it seems like these are separate subjects, but in fact, they are deeply connected. We all know and accept that AI changes how we work, and at the same time it can get tricky. The output is believable, but is it actually true? Is it grounded in reality? IMO there are multiple levels of assurance. 1. Plain LLM - will tell you what's in its weights. Fluency, generalization on the one hand, fantasy and hallunications on the other. Most current assistants use LLMs with tools these days, although there are still many use cases for a plan LLM. 2. Retrieval - search the web or documents. Gives us freshness and evidence. What we all use today, together with other tools. You get what was found in some index and some document. Vector databases, ranking. Citations. 3. Knowledge Graphs - LLM can ground its output in structured facts: who did what, when, according to which source, and how one claim relates to another. If you search for a Person, you can also easily find something related to a Company they work for, including multi-hop relationships. Ontologies are key. Provides structure, identity and provenance. 4. Formal models. For high stakes, it is not enough that "the answer sounds right". The model generates candidates. The formal model decides what is valid. It is a broad range: from validating the JSON against a schema to type systems, constraint solvers and modelling whole business domains. Increasing level of epistemic guarantees: plausible → sourced → structured → verified. No level is better than another. You want to give your model freedom for creative tasks, but stronger validation for your accounting agent. There are also tradeoffs. Greater guarantees require a narrower domain, and more explicitly encoded knowledge. That is why Knowledge Graphs are quite domain-specific. Formal models - their interaction with AI being of one the most active research areas - are even more task-specific. I worked for some time on bringing KGs to the domain of web3. medium.com/@valeriyzamaraiev… Starting May this year, the topic of agentic coding grounded in formal models became very hot. I think the whole topic of "vericoding" is only starting and we'll hear more and more about it. For example, I was able to produce a working EVM implementation. It replays the Ethereum chain, including all EIPs and pixel-perfect implementation of gas accounting quirks. As a pleasant side effect there is a formal model of the EVM that is itself an artefact with interesting possibilities in the area of security and audits. DM me if you are interested in these topics.

Sounds about right @bastian_wetzel bwetzel.medium.com/vericodin…

Something is cooking with vericoding. Lean + aarch64 machine code + Common Lisp to tie it all together.

Opus 4.8 just attributed a Rust performance problem to ... GC pressure ... sigh ..

There was an idea that it would be more natural for AI to code in machine code, rather than human languages - after all, we stopped looking at code. That argument was doubtful to me - LLMs are about capturing human language, and are non-deterministic - surely producing code in a higher level programming language is the way. However, with vericoding the game is indeed changing. Vibecoding in Lean -> extracting arm64 asm now, not without rough edges, but it works as a concept for sure. 2025 vibecoding: try it very carefully, expect bugs, review it with more agents, at least try to eyeball. 2026-2027+ vericoding: the quality of AI generated code cannot be matched by a human. Correctness is guaranteed *with the right process*. There's no guaratnee that the spec reflects what the author intended, and it is up to the humans to check that in the end. Also, mechanical construction from the spec end-to-end is still far from 100% (after all, we only have a handful of formally verified compilers - the trust surface is huge). But we are certainly getting there.

I'm at Proof of Talk 2026 — Louvre Palace, Paris. June 2 & 3. #ProofOfTalk #Paris2026

Vibe coding is so 2025. Vericoding FTW! Stay tuned ...

I was always looking for ways to write bug-free code. I tried Personal Software Process, lightweight formal methods like Z specification. I tried learning to use some proof assistants a while ago. I used functional programming languages (this works, especially for prototyping- you write a thing twice in different languages). And of course all sorts of testing, fuzzing etc. Many of these techniques were niche because of too much effort required and could not scale in an organization. Now things have changed. LLMs can not only hallucinate you your favourite spaghetti language, it can also spit out specs that can be automatically verified At ZkProof #8 in Rome felt like I was the rare one who doesn’t know Lean. Time to embrace the final form of software development.

My printer got jammed today, and I almost said out loud "hey claude can you fix it".

Not bad for late 90s, only now it is in the AIR over the Atlantic en route to @EthereumDenver Claude Code is happily cooking an MCP server for @dyor_cloud

Since I joined building crypto infra, I was always curious why nobody builds distributed transactions among multiple blockchains. In web2, we knew two-phase commit since decades. In web3, applying two-phase commit directly has some design challenges, so one has to be creative how to achieve this. Finally a great team is making it happen!

But then implements the plan in minutes ...

Claude Code finally started to estimate tasks in its plan in hours rather than weeks. It used to be a good trolling of meatbags.

This is nice, but I’ve already been using Claude Code to file my company’s expenses in the format my accountant needed, together with browser automation.

Developing agents is fun. You show them all the instructions, give all the knowledge and context. But they rebel, go their own way, and find another solution, sometimes a better one. Same experience as being father of a teenager.

There was never the right time to announce over the holiday break, so the time is now. I stepped down as CTO of @1kxnetwork . I'm deeply grateful to the team for 2.5 years of working with some of the smartest people in crypto. I'm currently fully focused on DYOR.bot. What is it? Initially I thought about it like Perplexity for web3, but now more like Palantir for Web3. Here is the problem - there's a lot of data in crypto, but the real insights are at the intersection of mulitple datasets. Once we merge the diverse sources into a single knowledge graph, new opportunities emerge. You could look at a contract, and know what people who developed it tweeted yesterday, or which other projects they had contributed to. Or what other projects did the developer of this contract do. All within a few hops along the edges of the graph. Or, you could juxtapose price action of a token with the events in the ecosystem, mentioned in the media - benefit of linked data. An agent will also remember your interest, and come back to you when the data changes - if you ask it. This is codified not in some opaque embedding, but rather a concerte node, or a query over a graph. Why is 2026 a good time to build such knowledge graphs? It is because LLMs and agentic tooling is indispensable for both constructing and consuming them. Previously, one would need to write SPARQL or Cypher. Vanilla LLMs have been already very helpful, and teams are already using them for access to databases. But to navigate a diverse dataset, driven by a well developed ontology, you need an agentic approach similar to how coding agents write code. The agent needs to discover what's available in the knowledge graph, try a few ideas, see what works, and eventually find a way to extract information that most satisfies the user request. 2025 made this approach possible, and agentic coding battle tested the best models and agentic architectures. For example, models such as Opus 4.5 or Gemini 3 Pro are really the first models that nail SPARQL and broad ontologies with high quality. 2026 will see an explosion of such applications in adjacent domains.