Sriram

AI enthusiast; content creator

[Ken Ono @KenOno691]

A year ago, while working on FrontierMath Tier 4 problems, I found myself grieving what felt like a loss of identity. If LLMs could do things that took us years to master, what was left of the mathematician?

A month later, my question changed: how do I use these tools and stay true to my passion for mathematics?

I have not left academia. I am on leave, working in AI, because I believe our profession has changed. AI offers new tools for discovery, reasoning, and verification. The artistry is the constant.

Mathematics was never about proving mathematicians are “smart” by amassing technical mastery. It has always been about beauty, taste, rigor, problems worthy of human attention, and applications that serve humanity. Honestly, I have never been a huge fan of math contests, or of the current frenetic need for benchmarks. I prefer the ideas and the actual math, even when the proofs are simple.

As a senior mathematician, I think a lot about the future. This is part of my mandate as a member of the Mathematical Sciences Education Board at the U.S. National Academies. I think about my intern Sidharth Hariharan, the first-year CMU PhD student featured in the NYT piece, and the many young mathematicians navigating these turbulent times.

Sidharth is a role model for the formalization movement, but more generally he is a superb example of what I believe is possible for the future of mathematics. And he is not alone. We are seeing a huge response from students, postdocs, established mathematicians, and math PhDs returning from industry because they feel the tides turning.

Training must change. Formalization, verification, and community-scale mathematical projects are not mere trends. They represent a growing movement that this senior mathematician sees as a bright future for mathematics.

I see promise: more frontiers opened, and a renewed need for human taste.

This gives me hope.

Mahalo.

[Timothy Gowers @wtgowers @wtgowers]

If you are a mathematician, then you may want to make sure you are sitting down before reading further.

[Timothy Gowers @wtgowers @wtgowers]

I've recently got in on the act of getting AI to solve open problems in mathematics. More precisely, I gave some questions asked by Melvyn Nathanson to ChatGPT 5.5 Pro, to which I have been given access, and it answered them. 🧵

[Daniel Litt @littmath]

fermat declining to publish the proof of fermat’s last theorem because it was “too dangerous”

[ℏεsam @Hesamation]

Microsoft employees watching Anthropic ship better AI features for office than Copilot.

[Yuchen Jin @Yuchenj_UW]

How did Anthropic do automation to Word, PowerPoint and Excel before Microsoft 365 Copilot?

[signüll @signulll]

monetization for agents is going to be one of the more interesting problems in ai very soon.

ads seem like an interesting first move, but they introduce a literal principal agent problem... e.g. your agent can't serve you & serve an advertiser simultaneously. the whole point of a personal agent is aligned incentives. ads potentially destroy that on contact.

any agent doing anything substantive is expensive af (see openclaw). agentic loops scale unpredictably with task complexity, & nobody's figured out how to cap that gracefully.

subscriptions break down cuz usage variance is enormous. see anthropic cutting off openclaw.

the only model that works for normal ppl is a centralized managed agent (like a gym membership)... one provider absorbing the cost variance across a large user base, the way aws abstracts away infra. no api keys, no surprise bills, & no understanding what a token even is. just a fixed price for an agent that does things.

one other model could be transactional monetization. agents that save you money or make you money take a cut of the delta. aligned incentives, measurable value, & no ad corruption. but that requires agents reliable enough to actually close loops & we're not quite there yet.

[Harmonic @HarmonicMath]

Formal verification enables trust at scale

[机器之心 JIQIZHIXIN @jiqizhixin]

Just updated: "Towards Autonomous Mathematics Research" from Google DeepMind—and it has Demis Hassabis's name on it!

The introduced Aletheia is a new math research agent thinks like a human mathematician, iteratively generating, verifying, and revising solutions end-to-end in natural language.

It leverages a novel inference-time scaling law, building on Gemini Deep Think.

Aletheia has already achieved remarkable milestones: generating multiple publication-grade papers (one without any human intervention!), autonomously solving four open questions from Bloom’s Erdős Conjectures database, and setting a leading performance on FirstProof, a benchmark for research-level problems proposed by mathematicians.

[Harveen Singh Chadha @HarveenChadha]

Sarvam 105B-A9b is here

A lot of blood and sweat went into this

[Thang Luong @lmthang]

6 months in, after the IMO-gold achievement, I’m very excited to share another important milestone: AI can help accelerate knowledge discovery in mathematics, physics, and computer science! We’re sharing Two new papers from @GoogleDeepMind and @GoogleResearch that explore how Gemini #DeepThink together with agentic workflows can empower mathematicians and scientists to tackle professional research problems. Some highlights:

The first paper built a research agent #Aletheia, powered by an advanced version of Gemini Deep Think, that can autonomously produce publishable math research and crack open Erdős problems.

The second paper, built on similar agentic reasoning ideas, helped resolve bottlenecks in 18 research problems, across algorithms, ML and combinatorial optimization, information theory and economics.

See the thread for details about the two papers and the joint blog post.

[Lean @leanprover]

In this highly accessible talk, Terence Tao outlines the future of research mathematics and why formal verification, via Lean and other tools, enables human-AI collaboration at scale:

"The reason why scaling and AI and broad participation actually is a net win is because we have formal verification... we have ways to filter out the untrustworthy inputs and keep the good ones."

📺 https://t.co/gk6HQOsllw

#LeanLang #LeanProver #FormalVerification

[Bartosz Naskręcki @nasqret]

Mathematics 🤝 AI 🤝 agents 🤝 hard problems and genuinely nice people.

That's what we're building at the MathBot Discord (https://t.co/AkwdDVVEyr). If you want to stay close to the latest developments like the #1stproof challenge and see what's really happening in the AI × Maths space, this is the place.

We already have Fields medalists, top-tier academics, and many sharp minds who just care about progress. No hype. Calm, deep, super-technical discussions.

And somehow, we're already 280 strong 🚀

[Ken Ono @KenOno691]

Five reasons I do AI for math:

We’re entering an era of instant access to domain-expert skills coupled with Lean-verification. We can now turn human curiosity into assisted intuition, and then into certified truth. 🧵 @axiommathai @leanprover

[Arnav Gupta @championswimmer]

https://t.co/bSLzAlL7Cs

[Carina Hong @CarinaLHong]

After AxiomProver got Putnam perfect score, we felt like it’s the right time to move forward from math Olympiads to “adult” math - research!

We also feel like it’s crucial to choose a theory-building domain.

The first modest test case for the quest of an AI mathematician.

[Daniel Litt @littmath]

That we can now automate some mathematics that previously required an expert is a huge deal. That said, the mathematics produced thus far is (in my obviously very subjective opinion) not notable in itself, but rather because it is automated and as a leading indicator.

[Aakash Gupta @aakashgupta]

The story of ElevenLabs is insane.

Two Polish high school friends from Warsaw watched badly dubbed American movies growing up. In Poland, entire films get a single male voiceover layered on top of the original English audio. Every character, same voice. They thought: we can fix this with AI.

That was 2022. They had zero funding and built their first voice prototype over a single weekend.

By January 2023, they could barely close a $2M pre-seed. Credo Ventures took the bet on two guys from Copernicus High School who’d gone on to Oxford, Cambridge, Imperial College, then Google and Palantir. Five months after launching their beta, they had over one million users.

Here’s where the math gets wild.

$2M pre-seed in January 2023.
$19M Series A in June 2023 at $100M valuation.
$80M Series B in January 2024 at $1.1B.
$180M Series C in January 2025 at $3.3B.

Now $500M Series D at $11B.

That’s $781M in total funding across five rounds in under four years. The valuation tripled in twelve months.

a16z quadrupled its position. ICONIQ tripled down. Nvidia came in as a strategic backer in September. Sequoia led this round and put Andrew Reed on the board. When your existing investors are fighting to increase their allocation by 3x and 4x, that tells you everything about the revenue trajectory they’re seeing internally.

And that trajectory is real. ElevenLabs went from $200M to $300M ARR in five months. They closed 2025 at $330M. The company has 400 employees across 13 cities. Deutsche Telekom, Revolut, Deliveroo, Square, and the Ukrainian government are all customers. Fortune 500 adoption is already above 60%.

Mati Staniszewski is talking IPO. The company is expanding into video, building voice agents that can talk, type, and take action. They’re going from “the voice company” to the interface layer between humans and machines.

From badly dubbed Polish TV to $11B in four years. Two high school friends who met at Copernicus High School in Warsaw now run one of the most valuable AI companies on the planet.

[Amit Shekhar @amitiitbhu]

From my interactions with students, today's tech job market:

• Freshers → Tough entry.

• 1–2 yrs → Fewer calls, average hikes.

• 3–7 yrs → High demand, multiple offers, 2X–4X hikes.

• 8–10 yrs → Fewer calls, average hikes, contract roles showing up more.

• 11+ yrs → Difficult phase. New roles often come with pay cuts or contract-based offers.