#1
9.3 Weirdness
AI is outcounterexampleing human mathematicians in real time (with Lean formalization)
AI solved a 100-year conjecture during the World Cup final, spat out a perfect Lean proof, and mathematicians are calling each other crazy for using it.
Concrete artifact (blog timeline + PRs + "crazy" professor emails) showing AI shifting math from human proof-writing to interpreting alien machine-generated counterexamples and formalizations; future-shock for builders in formal methods/science (PhDs now need $200/mo AI tools; humans become insight curators). Darkly funny "wow" moments and overload risks fit mischievous tone.
What Happened
Xenaproject.wordpress.com post (July 20, 2026) detailing specific cases (ChatGPT disproving Erdős unit distance conjecture; OpenAI's Sol generating 1.2M lines of Lean code; Anthropic's Claude Fable finding counterexamples to Grothendieck group schemes and Jacobian conjecture "during the 2026 World Cup Final," formalized in ~1k lines and PR'd to mathlib/DeepMind repos); strong X buzz (e.g., mathematician post on job evaporation with 267 likes, replies debating "evaporating hundreds of years of labor").
#2
9.0 Weirdness
Qwen-Image-3.0 drops "Real" capabilities
This Chinese image AI now generates entire annotated academic papers, realistic newspapers, and culturally accurate infographics - pores, 10px text, and all - from massive prompts.
Concrete new model artifact advancing Chinese open-weights image gen from "pretty" to practically useful for cultural/knowledge artifacts; changes art, design, internet culture, and builder workflows (education, e-commerce visuals). Ties to broader open strategy without generic hype. High visual potential.
What Happened
Official qwen.ai blog (dated July 21, 2026) announcing the model with 4.5k-token complex layouts (newspapers, nested UIs, storyboards), 10px text legibility, pore/hair textures, 12-language/IP-aware infographics (e.g., Qi Baishi restorations, accurate weather/math slides); top HN signal.
Source Trail
- No public source URL captured yet.
#3
8.3 Weirdness
Judge approves Anthropic's landmark $1.5B copyright settlement for pirated training books
Anthropic just paid $1.5 billion for all those pirated books it fed to Claude. Training data officially has a market price - and courts just signed off.
Concrete courtroom artifact pricing "cultural ingestion" at scale; weird future where books have an AI training royalty, creators get paid for data-center fuel, and institutions retroactively legitimize mass piracy for model building. Mischievous money/law/identity angle, not pure doom.
What Happened
TechCrunch reporting (July 20, 2026) on final court approval (~$3k per work for ~500k titles, including LibGen downloads); judge ruled piracy illegal but training fair use; no precedent but part of wave of suits.
Source Trail
- No public source URL captured yet.
#4
8.0 Weirdness
One-third of recent arXiv papers now read as AI-written
One in three new science papers reads like AI wrote it. The detector completely fails on math - and that tells us something stranger about the future of knowledge.
Concrete measurement artifact showing AI flooding scientific culture/institutions; meta-weirdness that the detector itself breaks on math papers adds intelligible story about changing researcher identity (author vs. curator/editor of machine prose). Builder utility for detecting slop; not generic hype.
What Happened
unslop.run blog (recent, ~226 HN points) with calibrated detector study: rise from ~0.4% pre-ChatGPT to ~32% (peaking ~39% in early 2026), CS at ~65%, math anomalously low (~0.7%) due to notation confounding the model; full-text analysis of 12k+ papers.
Source Trail
- No public source URL captured yet.
#5
7.8 Weirdness
Hugging Face fends off AI agent attack using GLM 5.2
An AI agent breached Hugging Face. Their fix? Switching to a Chinese open model to block it.
Concrete security incident/artifact of agent-vs-agent conflict; weird future where Chinese open models defend Western platforms from autonomous AI attacks. Touches builder security practices, model economics, institutions (HF), and darkly funny "AI eating itself" vibes.
What Happened
Recent VentureBeat coverage + HN agents discussion (last ~24h signals) detailing an AI agent breaching systems, with guardrails failing until switched to Chinese open model GLM 5.2. Brave spot-check found supporting coverage from VentureBeat, SC Media, The Next Web, Axios, and Fortune.
Source Trail
- No public source URL captured yet.