Pinned Tweet
There are tons of free ZK resources out there. The hard part is knowing what to learn, and in what order.
So I built a curated ZK learning roadmap:
Math → Cryptography → SNARKs/STARKs → Prover Engineering → zkVMs
github.com/SrishtiCode/ZKRoa…
Something interesting is happening in mathematics, and I think we are still underestimating what it could mean. I remember hearing Terence Tao discuss the possibility of solving the Riemann Hypothesis, suggesting that it might take until around 2050, even with a combination of human mathematicians and AI. Fast forward to November 2026, and Levent Alpöge, a mathematician at Anthropic, has announced a claimed proof of the Riemann Hypothesis developed with Claude. The reported timeline is approximately 35 days. Of course, the proof still needs independent mathematical scrutiny, but even the possibility of reaching this stage so quickly is remarkable. And this isn’t happening in isolation. In September, OpenAI announced an AI-assisted breakthrough on the Navier–Stokes existence and smoothness problem.
But this makes me wonder: are we actually accelerating mathematics, or simply accelerating the process of obtaining answers? A mathematical proof is valuable for more than establishing that something is true. It introduces ideas, techniques and connections that other mathematicians can understand, generalize and build upon. That’s how mathematical knowledge compounds. What happens when AI reaches a result through extensive computational exploration and formal verification? Even if the result is correct, does the process necessarily give us the same mathematical insight as a proof built around a fundamentally new idea? There is also another dimension worth considering: how much of these announcements is about advancing mathematics, and how much is about AI companies demonstrating that their models are better than their competitors? Competition can certainly drive progress, but I hope we don’t end up measuring AI’s contribution to mathematics purely by which company solves the next famous open problem first.
I don’t think computationally generated proofs are any less valuable. Perhaps we are entering an era where discovering a result and understanding why it works become two separate stages of mathematical progress. But the real measure of AI’s contribution shouldn’t just be how quickly it solves an open problem. It should also be how much new mathematics becomes possible because that problem was solved. Can AI help us discover ideas, techniques and connections that the mathematical community can build upon, or will we increasingly have answers that are correct but difficult to understand? That’s the question I find most interesting.
Srishti retweeted
If the Generalized Riemann Hypothesis (the stronger version of the Riemann Hypothesis) is proven, these become proven facts:
Algorithms that no longer need randomness
- Primality testing: Miller's fast test becomes proven correct with no randomness
- Square roots mod a prime: becomes solvable fast and with no coin flips (today that's an open problem)
- Cube roots and other roots mod a prime: same.
- Factoring polynomials over finite fields: almost fast (quasi-polynomial) with no randomness, instead of needing randomness
- Building finite fields: you can construct any finite field directly, with no random search
- Factoring vs. φ(n): factoring a number becomes provably as easy as computing φ(n), even without randomness
- Primes with a given remainder: a prime with any allowed remainder mod q exists that's at most about q², so searching for one is fast
Complexity theory
- Polynomial equations: "does this system of equations have a solution?" drops from PSPACE to a much lower class (AM)
- Algebraic P vs NP: Valiant's algebraic version of P vs NP gets tied to the normal P vs NP over the complex numbers
- Turing machines: a known 744-state Turing machine is proven to run forever
Number theory software
- Math libraries: PARI/GP and Magma results computed "assuming GRH" become fully proven
- Class groups: algorithms that compute these number-field structures get guaranteed running times
- Elliptic curve generation: the algorithms used to create curves for crypto get guaranteed running times
Cryptography (no new attacks, only proofs)
- Elliptic curves: proven that if discrete log is hard on one curve, it's hard on almost all similar curves
- Lattice crypto: known quantum attacks on some lattice schemes become proven to work
- Isogeny crypto: known quantum attacks on some isogeny schemes get proven speeds
- Code-based crypto: schemes like BIKE get the guarantee that the special primes they need exist
Prime gaps
- Next prime: there's always one within about √p·log p of p, so "find the next prime" has a guaranteed worst case
What it does NOT do
- It doesn't solve P vs NP
- It doesn't make factoring easy, so RSA and elliptic-curve crypto stay safe
- It doesn't make anything faster in practice, because randomized versions already work today
- Ramanujan graphs, expanders and AG codes are unaffected: they use a finite-field version of RH that was proven decades ago
Srishti retweeted
No, you just have to know a little bit about what you're doing. Here's a simpler instance.
The first example tries to prove that 2 - 3 = -1; this is true in the integers, and so that's how the question is interpreted (to be helpful to the user), and `by simp` solves the goal.
The second example tries to prove that 2 - 3 = 0; this is true in the natural numbers (because of "total functions"), and so that's how the question is interpreted (to be helpful to the user), and `by simp` solves the goal.
So no, you shouldn't just blindly "trust the Lean"; you should learn what you're doing.
And *then* you can trust the Lean. :)
Q. What does it mean that Prop is impredicative and proof-irrelevant, and what does each property buy and cost?
In Lean, Prop is impredicative, meaning you can create a proposition using ∀ over any type, even very large ones, and the result will still be a proposition. This makes it possible to express powerful logical statements. Proof irrelevance means that any two proofs of the same proposition are considered equal, so Lean cares about whether something is proven rather than how it is proven. These properties make proof checking simpler and keep proofs separate from computation. However, the limitation is that you cannot freely extract computational data from proofs, and some parts of Lean's type system become more complicated.
2
97
Q. What exactly does the Lean kernel check, and why can a small kernel justify trusting a huge system?
The kernel is a small C++ type checker for a dependent type theory (a variant of the Calculus of Inductive Constructions). It knows: dependent function types, universes Sort u, inductive families with recursors, a primitive quotient construction, and definitional equality (β, δ, ζ, η, ι-reduction, proof irrelevance, structure η), with fast paths for Nat and string literals. The elaborator, tactics (simp, grind, omega), macros, and notation are all untrusted programs whose only output is a term. The kernel re-checks that term from scratch. This is the de Bruijn criterion: trust is concentrated in a tiny checker, while the large, buggy, ever-changing automation only has to produce certificates the checker accepts. A bug in grind yields a failed proof, not a false theorem.
Srishti retweeted
Do you know what hash functions have collisions, or how good collision attacks get against the hash functions you use? Now you know. collisions.zksecurity.xyz/
Srishti retweeted
Software Verification in Lean is live now! You can still register via Luma to access the livestream link: luma.com/kytl0zlj
We're very excited for 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗩𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗶𝗻 𝗟𝗲𝗮𝗻 tomorrow! The in-person event is full, but livestream spots remain open: Register on Luma to receive the link: luma.com/kytl0zlj
𝗡𝗲𝘄𝗹𝘆 𝗮𝗱𝗱𝗲𝗱 𝘁𝗼 𝘁𝗵𝗲 𝗽𝗿𝗼𝗴𝗿𝗮𝗺: A closing round-table discussion on AI safety and formal verification (4:45–5:30 pm PT), moderated by Pedro Abreu (@p_droabreu0), host of the Type Theory Forall podcast, featuring Max Tegmark, Leonardo de Moura, Jonathan Protzenko, Andres Erbsen and Mike Dodds.
🔗 See the event website for the full program: beneficial-ai-foundation.git…
Srishti retweeted
Did some reading on formal verification this weekend:
hashcloak.com/blog/tutorial-…
vitalik.eth.limo/general/202…
adam.math.hhu.de/#/g/leanpro…
Absolutely worth reading. Especially vitalik's post if you want spark notes version. FV is hot these days, and it's not just empty hype.
Srishti retweeted
Oct edition of ZK Mesh came out earlier this week.
So. Much. Research (again!!)
zkmesh.substack.com/p/zkmesh…
Srishti retweeted
If you're trying to understand why the quantum threat has become such a hot topic recently, it's worth understanding three things:
There are three factors that determine how close we are to having a quantum computer capable of breaking cryptography like Bitcoin's:
1. Algorithms: finding ways to break cryptography using fewer qubits.
2. Hardware: building quantum computers with more physical qubits.
3. Error correcting code: making those qubits reliable enough to perform complex computations.
We're seeing progress on all three fronts: algorithms that require fewer qubits, better error-correcting codes, and new ways to build quantum hardware.
When all three improve together, the gap to breaking today's cryptography can close much faster than we might expect.
And that's why people like me are increasingly concerned that Q-day may be closer than we think.
Srishti retweeted
Very excited about the future, there is so much to be optimistic about.
Lots of good developments.
Srishti retweeted
🤘🏻 talk with @endrazine at @unpromptedconf
slides aumasson.jp/data/talks/unpro…
to appear in @phrack soon
Q. Why is a power-of-2 modulus dangerous for lattice-SNARK soundness, unlike prime-field SNARKs?
In prime-field SNARKs, using a modulus where (p-1) is divisible by a large power of 2 is useful because it allows efficient NTT operations.
However, in lattice-based protocols like LaBRADOR, soundness relies on differences between verifier challenges being invertible in the underlying ring. This prevents a dishonest prover from using zero-divisors to make different, inconsistent witnesses pass verification.
A power-of-2 modulus creates many non-invertible elements, increasing the possibility of such attacks and weakening the soundness assumptions.
This makes choosing the modulus particularly important in lattice-based protocols. zkSecurity's review identified this parameter-selection issue in code that was close to production.
Srishti retweeted
Extremely interesting to see how surprised people and companies are that AI agents and their principals might have different interests.
Principal-agent theory (PAT) is decades old and backed by a huge body of research. But when the agent is a software system, it gets a completely new spin.
PAT calls the cost of this divergence agency costs: the cost of monitoring the agent, of binding it to your interests, and the loss that remains anyway. Right now, for AI agents, these costs are extremely high. If you want to understand what’s going on, learn PAT.
It’s great to see researchers already applying these long-established frameworks to autonomous systems. The early findings point in one direction: whether the agent is a human or software, the same problems apply, with one HUGE difference. Software agents lack accountability. They can’t be held liable, sanctioned or fired, and that amplifies agency costs.
This field is young, but it will shape the broader questions around alignment. And aligning autonomous systems with humans and society is the make-or-break issue for actually scaling this technology. If we fail here, it will be regulated out of the market and simply rejected by people. That’s why we’re building solutions to exactly these problems.
Leading researchers worth following:
🔹 Noam Kolt (Hebrew University): Governing AI Agents, agency law and economics applied to AI agents
🔹 Gillian K. Hadfield (Johns Hopkins): incomplete contracting and the economy of AI agents
🔹 Dylan Hadfield-Menell (MIT): the principal–agent alignment problem
🔹 Alan Chan (GovAI): visibility and infrastructure for AI agents
🔹 Iason Gabriel (Google DeepMind): alignment between user, developer and society
🔹 Mihaela van der Schaar (Cambridge): multi-agent systems as principal–agent problems
🔹 Lewis Hammond & Vincent Conitzer (Cooperative AI Foundation / CMU): multi-agent risks
🔹 John Horton, Andrey Fradkin & Peyman Shahidi (MIT / BU): The Coasean Singularity, markets with AI agents
🔹 Steve Phelps (UCL): early experiments on LLM agents in principal–agent conflicts
Srishti retweeted
🎙️Last week @annarrose, @nico_mnbl, @kobigurk spoke with @rotn256 & @g_dellimmagine from @zksecurityXYZ about zkGolf, a competition to build the most efficient, formally verified ZK circuits.
They dug into Clean, AI-driven circuit optimisation and more!
zeroknowledge.fm/podcast/414
Srishti retweeted
Software Verification in Lean on November 7 has reached in-person capacity! If you're interested in attending live, please join the waitlist: A livestream link will be shared with all waitlisted registrants a few days before the event.
Software Verification in Lean (SVIL) comes to Stanford University on Saturday, November 7.
Co-organized by Beneficial AI Foundation and Lean FRO, it is a day of public lectures on advancing Lean’s software verification ecosystem toward production use, with a focus on end-to-end verification of security-critical systems.
Talks will cover Rust verification with the Aeneas framework, low-level software verification, and formally verified OS kernels.
Speakers include Max Tegmark, Shaowei Lin, Leonardo de Moura, Jonathan Protzenko, Andres Erbsen, and Mike Dodds.
In-person space is limited. Non-local and waitlisted registrants will receive a livestream link, and recordings will be available afterward.
Program and registration: beneficial-ai-foundation.git…



















