Editorial: Two themes ran through the feed today: AI as collaborator—and AI as industrial plumbing. One story shows an expert using a model to surface a potential mathematical bombshell; the others show open-source engineering shaving hours (or more) off real workloads. Both push the same question: when should we trust machine output, and when should we invest in faster, auditable tooling?
Top Signal
Terence Tao's ChatGPT conversation about the Jacobian Conjecture counterexample
Why this matters now: Terence Tao’s public ChatGPT session shows a world-class mathematician using an LLM to generate and then rigorously verify an alleged counterexample to a near‑century‑old conjecture—demonstrating AI’s potential to accelerate discovery, provided domain expertise validates every step.
Terence Tao posted a careful, public ChatGPT transcript that walks through an AI-produced candidate counterexample to the Jacobian conjecture and then verifies the construction step by step; you can read the full session here. What makes this different from headline-grabbing, later-debunked claims is the workflow: Tao treats the model as an exploratory collaborator but applies exacting algebraic checks rather than taking outputs at face value.
"Yes. I checked the identities independently, and the two maps are mutually inverse polynomial isomorphisms."
That quote from Tao captures the pattern: use the model to find candidate structure, then perform human verification that is formal and reproducible. The episode is important beyond one conjecture because it maps a working pipeline for high‑stakes discovery: prompt, propose, formalize, verify, and publish. It’s a practical reminder that LLMs can speed ideation but not replace domain verification—especially for claims that would rewrite textbooks.
Implications: expect more high‑signal, human‑led audits of AI-suggested mathematics and science. Labs and journals will need processes for reproducibility and provenance if AI becomes a standard research tool.
In Brief
Bento — An entire PowerPoint in one HTML file (edit+view+data+collab)
Why this matters now: Bento packages a full slide editor, viewer and collaboration runtime into a single HTML file, offering a portable, serverless alternative to cloud slide platforms that matters for privacy, archival stability, and low-friction sharing.
Bento compresses a deck, live editor, presenter mode and CRDT-backed collaboration into one self-contained HTML file; the demo and repo are linked from the project page. For teams that care about portability and offline-first workflows, Bento’s approach is a neat demonstration of "the file is the software." HN reactions praised the engineering and compared it to TiddlyWiki and reveal.js, while raising sensible questions about upgradeability and mobile browser quirks.
Key takeaway: Bento is a practical alternative to SaaS slides when you need archivable, account-free sharing and local-first editing.
Are AI labs pelicanmaxxing?
Why this matters now: Dylan Castillo’s systematic test asks whether labs overfit tiny viral benchmarks (the “pelican riding a bicycle” meme); the answer matters for how we interpret narrow wins and model reputations.
Castillo generated 1,008 SVGs across animal×vehicle prompts and used an LLM judge to score coherence; the writeup is at Dylan Castillo’s post. The punchline: after correcting for multiple comparisons, there was little evidence labs were optimizing for the meme. It’s a small experiment, but useful: bench wins can reflect data and objective biases, not secret lab chicanery, so skepticism of single-test bragging remains healthy.
Deep Dive
Terence Tao’s ChatGPT Jacobian audit (expanded)
Why this matters now: Tao’s transcript is a working template for responsible AI-assisted discovery—prompt engineering plus formal verification—exactly the pattern research communities will need to adopt for trust.
Tao didn’t simply read off a model’s claim; he translated the AI's construction into conventional algebraic geometry language, independently checked identities, and narrated the verification. This is significant for two reasons. First, it shows how an expert can squeeze value from an LLM: the model does exploratory, combinatorial work that narrows search space; the expert does proof and verification. Second, it exposes where human judgement is indispensable—spotting coordinate changes, validating polynomial inverses, and assessing whether a purported counterexample truly settles a conjecture.
Practical lessons for labs and institutions:
- Public, reproducible transcripts can accelerate community vetting.
- Auditability matters: store prompts, model versions, and the deterministic checks used.
- Extraordinary claims still need extraordinary verification—peer review and independent replication will remain the final arbiters.
GigaToken: ~1000x faster language model tokenization
Why this matters now: Gigatoken’s Rust tokenizer can cut dataset-prep and bulk tokenization from days to hours, changing cost and iteration speed for teams that preprocess massive corpora.
GigaToken is a new Rust-based tokenizer claiming dramatic throughput compared with common Python tokenizers; see the repo at Gigatoken on GitHub. It achieves speed by replacing regex pretokenizers with SIMD routines, aggressively caching mappings, minimizing Python roundtrips, and careful threading. Benchmarks show GB/s throughput that can tokenize Common Crawl-scale datasets in hours rather than days.
Why that matters: tokenization is often dismissed as a tiny slice of inference latency, but at dataset scale or in ingestion-heavy pipelines it can be a bottleneck and a cost center. Teams doing large-scale retraining, retrieval-indexing, or dataset curation get real wins—faster iteration, lower cloud bills, and quicker experiments. HN commentary flagged sensible caveats: exact tokenizer compatibility reduces peak speed, and many inference-only workflows won’t feel the difference. Still, for data teams and infra engineers, Gigatoken is worth evaluating.
Closing Thought
Both stories point to the same systemic truth: models are tools, not authorities. When a Fields Medalist treats an LLM as a co‑worker and then subjects its output to formal checks, or when infra teams replace brittle Python wrappers with SIMD-optimized Rust to make experiments practical, the common denominator is engineering discipline—auditable pipelines, reproducible checks, and measurable performance.
The Bottom Line
AI is maturing along two axes today: creative suggestion (where human verification decides what counts) and industrial efficiency (where faster plumbing changes what’s affordable). If you run experiments or fund model-driven research, invest in both robust verification and scalable infra—skimp on either and the gains evaporate.