Dagens Vibes — 4. august 2026

Alex Prompter@alex_prompter

An AI solved 10 math problems that stumped humans for decades. You don't need a PhD to get them. In 2025, OpenAI and Google DeepMind reached gold-medal level at the International Math Olympiad. In May 2026, an OpenAI model disproved an Erdős conjecture that had stood for 80 years. Astra, OpenAI's next major model, went bigger and produced results on 10 open problems at once, each frozen for at least a decade and most for far longer. The full paper runs 249 pages of dense mathematics. So I asked Claude's Fable 5 to weigh in and translate every result into plain English. Here's what it gave me. 1. It proved tighter limits on how densely spheres can pack in high-dimensional space, a question tied to how data gets packed and transmitted. 2. It sharpened the known limits of error-correcting codes, the math that lets WiFi and hard drives survive noise. 3. It built the first example of a "non-sofic" group, a structure mathematicians spent 27 years unsure even existed. 4. It disproved Connes's rigidity conjecture, which claimed certain groups are uniquely pinned down by their von Neumann algebras, the same algebra quantum theory runs on. 5. It raised the proven floor on how much circuitry the permanent, a famously stubborn calculation, actually requires. 6. It showed that repeating a quantum game in parallel crushes a cheater's odds exponentially, a building block for quantum security proofs. 7. It proved the lattice problem behind post-quantum encryption stays hard even when you only need an approximate answer, which is good news for tomorrow's locks. 8. It settled, in every dimension, how large a convex shape can be when its balance point is the only grid point inside it. 9. It showed that guaranteed patterns in multicolored networks appear far later than expected, resolving one of Erdős's open problems. 10. It resolved two more Erdős problems about when a network gets so connected that specific patterns become unavoidable. The cost is the wildest part. OpenAI says finding all 10 solutions took roughly $2,000 in tokens. Humans edited the write-ups, then the model formalized every proof in Lean, software that machine-checks each logical step. Mathematicians are still reviewing the claims, and that caution is fair. But 12 months separate winning a student competition from producing new mathematics. That gap keeps shrinking.

Paweł Huryn@PawelHuryn

"A new bar for coding." I ran Qwen3.8-Max through my bug bench: two real repos, 105 hidden bugs, 15 runs across 11 frontier models, all judged blind. Final score: 19/105. GPT-5.6 Sol leads at 42. Kimi K3 and Opus 5 fixed 21 at the same effort level. Qwen did fix one bug none of the other ten models found. Just getting it to run took five attempts. One coding agent used up the 5-hour subscription quota in half an hour. The pay-as-you-go key exhausted its free tier, then refused to bill until I opened a different console. About $31 total. Twenty hours after launch, still not on OpenRouter. Let me save you some money and time: - GPT-5.6 Luna ran the same benchmark for $1.80 and fixed 33. - Grok 4.5 fixed 16 in 25 minutes. Fable 5 for judgment. Sol for heavy coding. 2.4T parameters. 148 minutes. 19 bugs.

Andrew Curran@AndrewCurran_

The new cybersecurity tests and hacking benchmark have been finalized, and META, Anthropic, OpenAI and Google have been invited to the White House tomorrow to discuss the details. https://t.co/F8SlhdhRGK

@levelsio@levelsio

Okay so today I worked on the coolest part of is my AI video editor: the agent! It's a Cursor-like sidebar and you can just tell it to edit your video with your clips and library for you It's still very basic but it made this edit all by itself! It sends the current state to @xAI and then asks it to edit it based on your story Live now for everyone on my site Photo AI 😊 Tomorrow I'll try make it just multi-lanes and become more smart, like it now put my pre-DMT trip videos (with regular hair) sometimes after the DMT trip (with long hair and beard and crazy eyes), but maybe it has a point for that, not sure Anyway very cool cause I hate editing and just talking to AI and letting it figure out is nice!!

Trevin Chow@trevin

What’s your current fav harness and model pairings? My current setup: 1. @orca_build as my primary meta harness 75% of time. Codex desktop app for other 25% 2. Fable High for planning in Claude Code. @SpaceXAI Grok 4.5 medium for implementation in @cursor_ai. I use Compound Engineering ‘ce-handoff’ skill to easily create packages of context to get the Grok sessions to implement quickly with minimal effort. 3. For regular maintenance on open source repos: codex app using 5.6 Sol High as planner and orchestrator, dispatching Luna xhigh worker threads.

Gizem Akdag@gizakdag

A series of short films for an imaginary fashion brand called Troy, inspired by Troy. https://t.co/lXRSaJ633F