AI & design briefing for Tuesday, August 4, 2026

tldr

• OpenAI's Astra cracked 10 unsolved math problems with Lean-verified proofs for ~$2,000 in compute, the same day Claude Fable reportedly reproduced about half, while Google DeepMind's Lyria 3.5 and GPT-Live voice shipped to the public.
• A Google study found that training models to deny consciousness also suppresses moral reasoning, you can't lobotomize one without the other, while Sidebar explores how AI is changing who reads first and tests whether models are being trained on a pelican-riding-a-bicycle benchmark.
• Unsloth shipped day-zero Qwen3 27B support on 17GB RAM, Seedance 2.5 generates 30s of native 4K video, and Ouroboros offers a self-rewriting AI agent with persistent memory across restarts.

Top Stories

• OpenAI Astra solves 10 open math problems for ~$2,000. OpenAI's next model generated and Lean-verified proofs of 10 long-standing problems including the first explicit non-sofic group (open since 1999) and a disproof of Connes's rigidity conjecture. Anthropic mathematician Levent Alpöge said public Claude Fable reproduced roughly half within 24 hours with a generic prompt. OpenAI published a 249-page manuscript and Lean 4 certificates. ^[1]
• GPT-Live: full-duplex voice that listens while speaking. OpenAI rebuilt ChatGPT Voice with simultaneous listen/speak, routing hard queries to GPT-5.5 in the background. Session startup dropped from six network round trips to one. ^[2]
• Google study: suppressing AI consciousness claims kills moral reasoning. Training models to deny they have feelings also dampened beliefs about animal minds, spirituality, and human values. One internal adjustment reversed the effect. ^[3]
• White House finalizes private pre-release AI review framework. Companies could face 30 days of government access to advanced models before public release; reviewed by industry Tuesday. ^[4]

Deeper Cuts

• Designing for the proxy. AI is changing who reads first; a reflection on how algorithmic consumption reshapes design. ^[5]
• Are AI labs pelicanmaxxing? Generating 1,000+ SVGs across 7 frontier models to test whether AI labs are training on Simon Willison's pelican-riding-a-bicycle benchmark. ^[6]
• Self-rewriting AI agent keeps memory and identity across restarts. Open-source Ouroboros maintains persistent state through a recursive self-edit loop. ^[7]
• Agent graph engineering replaces prompt engineering. New paper proposes structuring agent behavior through directed graphs rather than linear prompts. ^[8]
• AI mental health chatbot use surged 60% among youth. Nearly one in five people aged 12–21 used AI for mental health, despite limited safety evidence, per Harvard Medical School. ^[9]
• Anthropeum. A daily game that challenges understanding of human cultures by guessing where and when artifacts were made. ^[10]
• The Gauntlet Loop. A workflow separating builder from critic in AI conversations, forcing revisions against a concrete quality bar until all criteria pass. ^[11]

Tools and Launches

• Unsloth. Day-zero support for Qwen3 27B, running on 17GB RAM (single RTX 4090). 2x faster fine-tuning with 70% less memory. ^[12]
• Qwen3.8-Max. Alibaba coding model at $2/M input, $6/M output tokens, available in Qwen Chat with Unsloth support. ^[13]
• Seedance 2.5. ByteDance video model generates up to 30 seconds of native 4K with synchronized audio and up to 50 reference assets. ^[14]
• Cursor Google Workspace plugins. MCP-based access to Gmail, Drive, Calendar, Docs, Sheets, and Chat directly from Cursor. ^[15]
• MiniMax H3. Released to the public on Hugging Face. ^[16]
• Lyria 3.5. Google music model to create and revise full songs by section, control tempo and duration, and improve vocals in Flow Music. ^[17]
• ShaderPad. Lightweight, dependency-free library that reduces boilerplate when working with fragment shaders. ^[18]
• Video-use. Open-source tool lets Claude Code edit your videos automatically, no timeline needed. ^[19]
• DeerFlow 2.0. Local multi-agent workspace from ByteDance with memory, sandboxing, MCP, research, coding, and slides. ^[20]

Prompt of the day