What a great evening of technical conversations about the future of AI reasoning with @axiommathai @BCapitalGroup and @MadronaVentures.
To anyone who thinks formal verification is a far away sci fi tale, it’s happening right now.
People mountain people sea!!!
Our Verified Reasoning event at Neurips is a huge success - thank you for everyone for coming! We are here till 7pm!
And especially thank you for the 300-400th people on the waitlist! Come at 6pm to Shorebird anyways 😆
[email protected]
It is very clear that this is the future, and Axiom is honored to pave the way forward.
Jon Turow retweeted
AxiomProver keeps getting stronger.
And @axiommathai keeps growing.
I'm excited to share that Prof. Ken Ono @KenOno691 has joined Axiom as Founding Mathematician and FTE #15.
He left his tenured position as STEM Advisor to Provost at UVA to build an AI mathematician with us.
Here's Ken's story. (1)
Jon Turow retweeted
I had the pleasure of chatting with @TheTuringPost about all things Axiom and AI4Math!
Tune in:
- How I define AGI vs domain specific ASI, the plate metaphor, can a chatty or poetic model prove the Riemann Hypothesis, and math’s incredible transfer learning power to coding and engineering
- Why an AI mathematician needs both proofs and constructions capability, mapping to the two branches of Axiom: formal proving, and specialized discoveries
- Data scarcity compared to code, the chicken-and-egg problem of autoformalization, Axiom’s bold synthetic data bets, and how we think about new knowledge generation
- Problem-Solving vs Theory-Building, why "Theory Building" is harder to benchmark than the International Math Olympiad (IMO), and why literature search / retrieval, so emphasized by LLMs today, is unsatisfying
turingpost.com/p/carina
Jon Turow retweeted
There are roughly 1,100 problems in the Erdos canon, compiled from decades of papers. They've guided serious research in combinatorics and number theory for generations. Of those 1,100, only 266 have ever been proved. And only 10 have proofs fully formalized in Lean.
Axiom just added two more to that list of 10.
Problem #124 is an additive number theory question about representing integers as sums of powers across multiple bases. It stayed open for ~30 years.
Problem #481 is even older, from 1980. It asks whether a certain iterated arithmetic process must eventually produce a repeated element. Deceptively simple to state, but Erdos and Graham themselves wrote that it was "surprising" how difficult it turned out to be. Open for 45 years.
A few weeks ago, OpenAI claimed GPT-5 solved Erdos problems, only for the community to point out it had just retrieved existing literature.
Axiom actually solved it, with Lean that checks proofs down to foundational axioms. Every step must type-check. There's no hand-waving.
Axiom is late to the game, with a fraction of the resources, and they're producing proofs that major labs aren't able to solve.
Axiom sets out to build an AI mathematician.
We are the underdog.
4 months old, 2 years late to the game, under 10 FTEs (recently grew to 17), and had 1:5 in funding and in valuation to our competitor.
Today, AxiomProver solved Erdos Problems #124 and #481 in Lean, a 100% verifiable language.
Onwards!
If you care about the next frontier of reasoning — not just faster LLMs, but systems that formalize, prove, and conjecture — Axiom is building it.
An invitation to work on the hard, beautiful problems.
A key piece is construction-driven discovery: the examples that seed conjectures, guide proof strategies, and surface structure we don’t yet understand. Axiom’s new post describes their tools for this — PatternBoost, Int2Int, and early mathematical world models.
Introducing Axiom’s discovery team, led by @f_charton:
We build models that will create novel constructions, map problems into solutions and intuitions, and learn the structure of entire mathematical worlds.
Built to attack hard open problems, one at a time.
98% error reduction. Built-in production backends. Frontier AI agents, zero CLI.
Dental offices building patient portals. Sales teams spinning up custom CRMs. Domain experts creating tools that would've needed dev teams last year.
Software creation, democratized. Congrats @stackblitz team
Today vibe coding goes pro.
Introducing Bolt v2:
→ World's best agents (Claude Code, Codex)
→ Built in backend (hosting, DB, storage, ...)
→ No error loops, no setup nightmares
Now anyone can build without boundaries.
Readers added context they thought people might want to know
The "👨👩👦 Readers added context" note under the video is not actually a community note. Their marketing team wrote that. communitynotes.x.com/guide/en/about…
Jon Turow retweeted
Today, I am launching @axiommathai
At Axiom, we are building a self-improving superintelligent reasoner, starting with an AI mathematician.
AI agents are already driving real economic activity — Visa reports a 1,200% YoY spike in retail traffic from agents.
But agents aren’t like bots — they are bots.
And payments infra isn’t ready.
That’s why we backed @nekuda_ai: the trust layer for agentic commerce.
🧵👇
We believe agents must become trustworthy before they can become trusted.
And when they are, they’ll stop being demos — and start contributing to GDP.
Nekuda makes that future happen faster.
Huge congrats to @CoolestAK87, Idan, Barak.
Thrilled to support you on this journey along with our friends at @AmexVentures, @VisaVentures, @pk_iv, @shyamalanadkat, and @theaievangelist
More here → jonturow.substack.com/p/why-…
Post-show, pre-filter — @douwekiela and I kept the tape rolling.
In this 10-min bonus chat with Douwe Kiela (@ContextualAI, RAG co-creator), we get into:
🔹 Vertical storytelling > generic GTM
🔹 Building platforms through possibility
🔹 Owning the stack vs advice to "do one thing well"
🔹 Lessons from Meta’s early AI moderation
Founders, don’t miss this: substack.com/home/post/p-161…
Thanks for sharing @ITNAmatter !
Understanding the relationship between infrastructure and applications is crucial for making sense of the MCP ecosystem. Way back in 2018, Union Square Ventures wrote about "The Myth of the Infrastructure Phase" and illustrated how infrastructure and applications evolve in tandem:
Traditional APIs require precise parameter formatting, error handling, and response parsing. AI models, with their probabilistic outputs, frequently hallucinate parameters, misunderstand response formats, or fail to properly handle errors. This made reliable tool use by AI agents nearly impossible at scale.
MCP solves this problem by providing a standardized way for AI models to discover what tools are available, understand how to use them correctly, and maintain conversation context while switching between different tools. It brings determinism and structure to agent-tool interactions, enabling reliable integration without custom code for each new tool.
The number of available MCP servers is growing rapidly. The official MCP typescript SDK is growing quickly and showing around 700K weekly downloads from npm. Smithery's MCP discovery platform saw an explosive 3x growth in server creation during March alone.
But actual usage remains highly concentrated. Of Smithery's 2,500+ MCP servers, only 8 have surpassed 50,000 installs. The distribution follows a classic power law:
We found 53 MCP SDK packages and 751 MCP server packages in the npm ecosystem. SDK downloads are growing significantly faster than server downloads, pointing to a supply-led dynamic. Developers are preparing for use cases that haven't yet materialized.
@jturow
jonturow.substack.com/p/what…
The hype around Model Context Protocol is deafening, so @baxterblack and I dug into the data to find out what's real.
* 2,500+ MCP servers on @AI_Smithery, but only 8 with >50K downloads
* Early demand driven by desktop apps and technical users
* Creating opportunities that weren't possible before
jonturow.substack.com/p/what…
"Everything is Contextual."
That's what @douwekiela , who co-invented RAG and now leads @ContextualAI, told me during our convo—and it perfectly captures the state of enterprise AI:
Beyond RAG vs fine-tuning: "You should be doing all of those things"
The gap to real ROI is still massive
"Active retrieval" is the future
🎧: jonturow.substack.com/p/from…
Jon Turow retweeted
AI isn’t just for researchers — it’s for problem-solvers.
Databricks Chief AI Scientist Jonathan Frankle joined @MadronaVentures @jturow to share insights on:
- How enterprises are scaling AI from prototype to production
- Popular tools for building AI products
- The biggest misconception about enterprise AI adoption
- The storytelling playbook every AI startup needs
Watch the full conversation: youtu.be/VSRXWoe_7sA