Published on July 23, 2026 at 06:01 CEST (UTC+2)
Terence Tao's ChatGPT conversation about the Jacobian Conjecture counterexample (684 points by gmays)
This article shares a conversation between mathematician Terence Tao and ChatGPT where they discuss a potential counterexample to the Jacobian Conjecture, a long-standing problem in algebraic geometry. It demonstrates how large language models can engage in sophisticated mathematical reasoning and serve as collaborative partners for experts. The conversation highlights both the potential and limitations of AI in theoretical research.
Quality non-fiction books are the antithesis of AI slop (204 points by benbreen)
The author reflects on his college job shelving library books, which led him to discover high-quality non-fiction by chance. He contrasts this with the proliferation of shallow, AI-generated content (“AI slop”) and describes building a tool to help readers find meaningful, human-written non-fiction. The article argues that well-researched books offer depth and serendipity that AI-generated text cannot replicate.
GigaToken: ~1000x faster Language model tokenization (404 points by syrusakbary)
GigaToken is a Rust-based tokenizer that achieves ~1000x speed improvement over HuggingFace’s tokenizers and tiktoken, reaching GB/s throughput. It is a drop-in replacement designed to accelerate preprocessing for large-scale language model training and inference. The project is open-source and aims to remove a major bottleneck in NLP pipelines.
Show HN: Bento - An entire PowerPoint in one HTML file (edit+view+data+collab) (699 points by starfallg)
Bento is a tool that packages an entire PowerPoint presentation into a single HTML file, enabling editing, viewing, data integration, and real-time collaboration directly in a browser. It simplifies sharing and editing slides without requiring traditional office software. The approach leverages modern web standards for lightweight, portable slide decks.
Medici family mystery may be solved after more than 400 years (92 points by effects)
Scientists used DNA analysis to investigate the 1587 deaths of Grand Duke Francesco I de’ Medici and his wife Bianca Cappello, which had long been suspected as poisoning. The study suggests they actually died of malaria, based on genetic evidence of the parasite. This resolves a four-century-old mystery and showcases the application of ancient DNA forensics to historical questions.
Everyone should know SIMD (305 points by WadeGrimridge)
Mitchell Hashimoto argues that SIMD (Single Instruction, Multiple Data) is simpler than commonly believed and should be learned by all developers for routine performance optimization. He presents a common pattern for vectorized loops, using Zig for examples, and notes that modern languages increasingly expose SIMD intrinsics. The post aims to demystify SIMD and encourage broader adoption for everyday programming tasks.
Are AI labs pelicanmaxxing? (421 points by dcastm)
This article investigates whether AI labs are optimizing their models specifically for Simon Willison’s informal “pelican on a bicycle” SVG benchmark, a practice dubbed “pelicanmaxxing.” The author ran a controlled experiment generating SVGs across multiple frontier models and scoring them with an LLM judge. The results raise concerns about benchmark gaming and the reliability of informal evaluations in AI.
ascdraw: Editor for ASCII/UTF-8 diagrams (in 144FPS) (12 points by xlii)
ascdraw is a keyboard-driven editor for creating ASCII/UTF-8 diagrams, offering an effectively infinite canvas with high frame-rate rendering (120+ FPS). It supports shapes, text, layers, and export to TXT, JSON, or PNG. The tool is native, open-source (GPLv3), and designed for developers who prefer text-based diagramming.
Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong (90 points by HenryNdubuaku)
Cactus Hybrid post-trains small on-device models (starting with Gemma 4) to output a confidence score (0–1) alongside each answer, enabling automatic handoff to a larger cloud model when confidence is low. This hybrid approach balances privacy, speed, and accuracy, making small models more reliable for production use. The project provides quick-start integrations for popular inference frameworks.
Restructuring GitHub's bug bounty program (14 points by soheilpro)
GitHub is restructuring its bug bounty program to reduce noise and better reward high-impact security research. Key changes include a permanent invite-only VIP program for top researchers, with higher payouts and faster response times. The goal is to create a more rewarding experience for serious researchers while maintaining platform security.
LLMs as mathematical research collaborators
Terence Tao’s use of ChatGPT to explore the Jacobian Conjecture shows that large language models are increasingly being treated as thinking partners for high-level theoretical work. This trend matters because it pushes AI beyond simple Q&A into creative problem-solving, but it also raises questions about reliability and verification in rigorous domains. Implication: AI labs should invest in domain-specific fine-tuning and reasoning traces to make models more trustworthy for scientific discovery, while researchers must maintain critical oversight.
Benchmark gaming and “pelicanmaxxing” exposes fragility of informal evaluations
The widespread attention on a quirky SVG-generation benchmark has created incentives for labs to over-optimize for that specific test. This “pelicanmaxxing” phenomenon mirrors larger concerns about benchmark saturation and Goodhart’s law in AI evaluation. Why it matters: As informal benchmarks gain influence (e.g., in Hacker News discussions), they can distort real-world capabilities and mislead users. Actionable takeaway: The community should develop diverse, adversarial, and continuously updated benchmarks, and treat any single metric (especially viral ones) with skepticism.
Tokenization speed is a critical bottleneck being addressed
GigaToken’s 1000x improvement over existing tokenizers highlights that preprocessing is often overlooked in ML pipelines. With models growing larger and datasets expanding, even tokenization can become a major time cost. Why it matters: Faster tokenization directly reduces training time and inference latency, especially for real-time or billion-token-scale workloads. Implication: Expect more innovations in tokenizer parallelism, hardware-aware implementations, and possibly end-to-end models that bypass tokenization entirely (e.g., byte-level models).
On-device AI with confidence calibration enables practical hybrid deployments
Cactus Hybrid’s approach—embedding confidence probes inside small models to decide when to escalate to a larger cloud model—addresses the core trade-off between privacy, speed, and accuracy. This trend is gaining momentum as edge devices demand capable AI without constant internet dependency. Why it matters: Users get private, low-latency responses for simple queries while still having access to powerful cloud models when needed. Actionable takeaway: Developers should integrate confidence-aware architectures into their deployment strategies, and model providers should standardize confidence output formats to simplify such hybrid systems.
AI-generated content (slop) fuels a counter-movement toward curation and quality
The article contrasting quality non-fiction with “AI slop” reflects a growing backlash against low-effort, mass-produced AI text. This is not merely a critique but a driver for new tools that help humans discover vetted, human-authored content. Why it matters: As AI-generated content floods the web, trust and signal become scarce. Implication: Startups and platforms that focus on curation, provenance, and verification (e.g., “human-written” badges, recommendation systems that penalize slop) will gain value. For AI/ML development, this means investing in synthetic data quality filters and authentication methods.
Understanding low-level hardware parallelism (SIMD) is becoming essential for ML practitioners
Mitchell Hashimoto’s plea that “everyone should know SIMD” reflects a broader trend: as model size plateaus, software-level optimizations become key to performance. ML inference and training kernels rely heavily on vectorized operations, and exploiting SIMD can yield 2–10x speedups on CPU for tasks like tokenization, embedding, and small model inference. Why it matters: Many developers treat hardware acceleration as “the compiler’s job,” but manual SIMD knowledge enables fine-grained tuning and portability. Actionable takeaway: ML engineers should learn basic SIMD patterns (e.g., via Rust, Zig, or intrinsics in C++) to write efficient custom kernels, and language designers should prioritize exposing SIMD primitives ergonomically.
Analysis generated by deepseek-reasoner