In Brief
Bento — An entire PowerPoint in one HTML file
Why this matters now: Bento’s single-file presentation format offers a portable alternative to cloud slide tools, letting presenters ship a complete editor+viewer as one self-contained HTML document.
Bento packages a deck (JSON), an editor, presenter mode, and the runtime into a single HTML file that self-saves, embeds assets as data, and can even do CRDT-backed collaboration, according to the creator’s post. The pitch is simple: "The file is the software" — open it anywhere, no account or server required. The implementation uses a base64 runtime blob, File System Access API fallbacks, and ECDSA-signed runtime updates to avoid supply-chain pitfalls.
"One HTML file — deck, viewer and editor together. Open it anywhere. It saves itself."
Practical questions remain — upgrading documents across versions, mobile browser quirks, and reliance on a relay for collaboration — but Bento is an elegant reminder that local‑first web apps can be surprisingly powerful when portability and archival value matter.
Source: Bento — An entire PowerPoint in one HTML file
Are AI labs pelicanmaxxing?
Why this matters now: Dylan Castillo’s structured test checks whether labs secretly optimize models for viral micro‑benchmarks, and the short answer reported is: not for pelicans on bicycles.
Castillo generated 1,008 SVGs across animal×vehicle prompts, judged them with a single LLM rater (GPT‑5.6 Luna), and found the pelican-on-a-bicycle combo ranked near the bottom once difficulty adjustments and multiple-comparison corrections were applied. The experiment shows how to turn a meme-like probe into a controlled test and — at least here — finds no smoking gun that labs are overfitting to the pelican meme.
"There is little evidence that AI labs are pelicanmaxxing."
The write-up is playful but methodical, and the outcome suggests labs are more likely improving general SVG generation than gaming one tiny viral cell.
Source: Are AI labs pelicanmaxxing?
Deep Dive
Terence Tao’s ChatGPT conversation about the Jacobian Conjecture counterexample
Why this matters now: Terence Tao used a public ChatGPT session to audit an alleged AI-produced counterexample to the century-old Jacobian conjecture, showing how top experts can use LLMs as exploratory collaborators while underscoring the need for independent verification.
Terence Tao posted a transcript in which he and ChatGPT walk through an AI-supplied polynomial construction that, if correct, would be a monumental counterexample to the Jacobian conjecture. Tao independently checked algebraic identities, verified that an affine variety X is isomorphic to A^3, and composed maps to produce a polynomial map G: C^3 → C^3 with constant nonzero Jacobian (det DG ≡ −1). As Tao put it in the transcript:
"Yes. I checked the identities independently, and the two maps are mutually inverse polynomial isomorphisms."
What makes this episode notable is less that ChatGPT conjured a ready-made, peer‑perfect proof and more that a world-class mathematician used the model as a tool to generate and then scrutinize candidate constructions. The transcript reads like a careful audit: ChatGPT proposes formulas, Tao tests them, and the dialogue iterates until the algebraic facts line up. That workflow — model suggests, expert verifies — is a plausible pattern for future exploratory work in pure math.
But caveats are real and immediate. The Jacobian conjecture is near‑century‑old; any claimed counterexample demands independent reproduction, formal checks, and peer review. LLMs are still prone to subtle hallucinations: a plausible-looking polynomial identity can hide a small but decisive mistake. Hacker News reactions captured both excitement and caution: readers admired the prompting craft and iterative verification, while others urged repeated independent checks and warned against treating the model’s output as proof.
If accurate, this would be a headline-making mathematical advance; the safe takeaway today is that expert-in-the-loop workflows let LLMs accelerate exploration, not replace rigorous validation. Expect follow-ups from algebraic geometers and formal-verification efforts to either confirm the construction or find the flaw.
Source: Terence Tao's ChatGPT conversation about the Jacobian Conjecture counterexample
GigaToken: ~1000× faster language model tokenization
Why this matters now: GigaToken’s Rust tokenizer claims massive throughput gains, and adopting it could cut dataset preprocessing time from days to hours for large corpora while improving time‑to‑first‑token in latency-sensitive systems.
GigaToken is a Rust project that markets itself as "the fastest tokenizer for language modeling," with benchmark claims of order‑of‑magnitude speedups versus tiktoken and HuggingFace tokenizers. The engineering is straightforward and aggressive: replace regex-based pretokenizers with SIMD-optimized routines, aggressively cache pretoken→token mappings, minimize Python boundary crossings, and use efficient threading with low synchronization.
"Gigatoken is the fastest tokenizer for language modeling."
Those techniques add up. The author reports GB/s throughput on EPYC CPUs and shows that bulk tokenization of Common Crawl becomes tractable in hours rather than days. That’s material for teams who preprocess huge corpora, build routing/ingestion pipelines, or measure energy use across billions of tokens.
But the claims come with pragmatic caveats. Exact compatibility with HuggingFace token outputs cuts into peak speed, SentencePiece/WordPiece coverage is limited, and Windows testing is still sparse. Also note that for many inference workloads tokenization is a small slice of total latency — often <<1% — so replacing a tokenizer won’t move the needle on single-request latency unless you’re streaming tokens or operate at massive scale.
Where GigaToken is compelling is in bulk and pipeline contexts: dataset creation, reruns of tokenization at new vocab sizes, and cloud-cost-sensitive preprocessing. Teams considering adoption should test the compatibility mode, validate outputs against their current tokenizer, and measure time-to-first-token for their specific inference patterns. If those metrics matter, GigaToken is worth a serious bench.
Source: GigaToken on GitHub
Closing Thought
Two themes ran through today’s stories: tools that change the shape of hard work, and experts who won’t let tools do the thinking for them. Terence Tao’s transcript shows how models can propose bold, unfamiliar constructions but can’t yet shoulder the burden of mathematical certainty. GigaToken shows how low-level engineering still delivers outsized practical wins when you’re processing mountains of data. Both remind us that progress comes from pairing clever automation with human judgment and careful measurement.