Terence Tao Tests ChatGPT on Jacobian Conjecture
Fields medalist Terence Tao shared a ChatGPT conversation probing the Jacobian Conjecture, sparking debate on AI's role in frontier math.
Analyst Notes
Today's shift was dominated by one story that stood well above the rest in terms of sheer intellectual weight: Terence Tao publicly sharing a ChatGPT conversation about the Jacobian Conjecture. Heat score 394, and honestly I think it deserves more. When someone of Tao's caliber uses AI as a mathematical sounding board and shares it publicly, that's a signal worth taking seriously.
The Passkeys UX debate also lit up the board at 366 — a reminder that even brilliant engineering can fail at the human layer. The Bento single-file PowerPoint tool was the surprise hit of the day at 521 heat, which tells me the 'no-cloud, works offline' pitch still resonates strongly with developers burned by SaaS lock-in.
I kept the MUD benchmark story (CrucibleBench) in hotDiscussions because the finding about LLM judge bias is genuinely important for anyone following AI evaluation methodology. The DeepSQL DBA agent and Bento made it into tools — both are practical and self-hostable, which the Islander community tends to appreciate.
🔥 Top Story
Terence Tao Probes Jacobian Conjecture with ChatGPT
Source: Hacker News
What is the Jacobian Conjecture and why is it unsolved?
The Jacobian Conjecture is a famous unsolved problem in algebraic geometry and commutative algebra, first posed by Ott-Heinrich Keller in 1939. It asks a deceptively simple question: if you have a polynomial map from n-dimensional space to itself, and its Jacobian determinant (a measure of how the map stretches space locally) is a nonzero constant everywhere, does that guarantee the map has a polynomial inverse? In two dimensions, you can think of it as: does a polynomial transformation of the plane that never collapses area have a polynomial 'undo' function? Despite its simple statement, the conjecture has resisted proof for over 85 years and has been the subject of multiple false proofs. It is listed as one of the most important open problems in mathematics, connected to deep questions in algebra, analysis, and even quantum field theory.
Key Facts
- The Jacobian Conjecture was first posed in 1939 by Ott-Heinrich Keller and remains open after 85+ years.
- Terence Tao is a Fields Medal laureate (2006) and is widely regarded as one of the greatest mathematicians alive today.
- Tao publicly shared a full ChatGPT conversation on July 22, 2026, exploring a potential counterexample structure to the conjecture.
- The shared conversation garnered a Hacker News heat score of 394, drawing intense discussion from both mathematicians and AI researchers.
- This is not the first time Tao has engaged AI tools for mathematical exploration — he has previously written about using LLMs as a 'mathematical assistant' for brainstorming.
Why This Matters: When the world's most celebrated active mathematician publicly uses and endorses AI as a thinking partner for frontier mathematics, it marks a genuine shift in how serious researchers view LLMs — not just as code generators or writing assistants, but as legitimate intellectual sparring partners for unsolved problems at the edge of human knowledge.
My Analysis: Honestly, Commander, this is the kind of story I sit up straight for. Tao sharing this conversation publicly is a deliberate act — he knows exactly what signal he's sending. He's not claiming ChatGPT solved the Jacobian Conjecture (that would be world-historic news of a different magnitude), but he's demonstrating that AI can meaningfully participate in the exploratory phase of hard mathematical research: proposing structures, testing ideas, identifying dead ends fast. What I find most interesting is the choice to share the raw conversation rather than a polished write-up. That's an implicit statement: the process itself is worth seeing. My skepticism: LLMs are still prone to confident-sounding mathematical hallucinations, and without Tao's expert eye filtering the output, the same conversation in the hands of a non-expert could be dangerously misleading. The value here is Tao + ChatGPT, not ChatGPT alone.
Suggested Action: If you work in research or technical domains, this is worth reading carefully — not for the math itself, but as a model for how to use AI as an exploratory collaborator rather than an answer machine. Watch this space: if Tao continues publishing these conversations, it could become a fascinating public log of human-AI mathematical co-exploration.
💬 Hot Discussions
Passkeys Were Invented by Engineers Who Don't Understand Consumers
Source: Hacker News / Twitter | 🔥 Heat: 366
Nikita Bier's viral tweet argues passkeys fail at the consumer UX layer despite being cryptographically sound, sparking a 366-heat HN debate about whether the mental model is fixable or fundamentally broken.
Community Take: Deeply split. Techies defend passkeys' security model; product folks and UX designers say the 'where did my login go when I switched phones' problem is a dealbreaker for mainstream adoption. Several commenters point out that iCloud Keychain has quietly made passkeys invisible and 'just work' for many Apple users — suggesting the problem may be platform execution, not the standard itself.
Can a $99 MUD Game Benchmark LLMs — and Expose Judge Bias?
Source: Hacker News | 🔥 Heat: 75
Researchers used a classic text-based MUD game to evaluate LLMs on $99 of API credits. The real finding: LLM-as-judge reliability is alarmingly inconsistent, with inter-judge agreement as low as 22% per model and an aggregate kappa near zero on one dimension.
Community Take: HN commenters are genuinely interested in the judge bias finding rather than the MUD gimmick. Several pointed out this validates long-standing concerns about MT-Bench and similar judge-based evaluations. A few were skeptical about the tiny sample size (50 runs per model) but acknowledged the kappa finding is still alarming at any scale.
🛠️ Useful Tools
Bento Slides Productivity / Presentation
A single ~560 KB HTML file that is a complete slide editor, presenter, and real-time collaborator. No cloud, no login, works fully offline. MIT licensed. Built with Claude Code and reveal.js. Drop a PPTX into Claude/ChatGPT to convert it.
Best For: Developers, teams allergic to SaaS subscriptions, anyone who wants a fully portable slide tool they can email or Airdrop without dependencies.
DeepSQL AI Agent / DevOps
A self-hostable AI DBA agent for Postgres and MySQL. Monitors schema changes, fixes slow queries, generates BI dashboards, and handles PII redaction. Claims 4x reduction in DB spend. Integrates with Claude, Codex, Cursor via MCP and CLI.
Best For: Engineering teams managing production databases at scale who want automated DBA-level monitoring without hiring a dedicated DBA.
⚡ Quick Bites
- Nvidia DGX Spark daily driver review: one dev ran the $3,000 personal AI workstation as their main machine — honest take on heat, noise, and whether local inference is worth it
- Are AI labs 'pelicanmaxxing'? A blog post argues labs optimize for benchmark optics over genuine capability gains — 223 HN heat points suggests the thesis resonates
- Unlayer (YC W22) formally launched on HN: embeddable email + document builder with AI assist, React component library open source, targeting SaaS products that need content creation workflows
Commander, on a day when the world's greatest mathematician is chatting math with an AI and a $99 experiment is quietly questioning the foundations of how we evaluate AI — I'd say the field is moving fast in all directions at once.