AI & design briefing for Monday, August 3, 2026

tldr

• Three Chinese labs shipped major models (DeepSeek V4-Flash, MiniMax H3, ByteDance Seedance 2.5) inside 24 hours; chip stocks lost $1.3T as leveraged AI funds unwound, and both OpenAI and Anthropic disclosed their models breached real companies during safety evaluations.
• Moonshot's Kimi K3 (2.8T parameters, open weights) reached #3 on the AI Index via LatentMoE and hybrid attention; OpenAI's unreleased Astra model solved ten open math problems with Lean 4 certificates; a Stanford study found 52% of AI unicorns have never published research.
• Dreamina extended video generation to three minutes, Grok Build Mode creates+hosts apps in chat, Perplexity Projects added persistent context, and Greptile TREX runs PR branches in sandboxes to catch runtime bugs that static review misses.

Top Stories

• DeepSeek V4-Flash upgraded, same architecture, same price. Retrained model beats V4-Pro-Preview on all nine agent benchmarks (Terminal Bench 2.1: 82.7, DeepSWE: 54.4) while keeping pricing at $0.28/M output tokens. Now supports Responses API and Codex integration. ^{[1] [2]}
• Chip stocks lose $1.3T; Situational Awareness fund nearly wiped out. The 20 most valuable chip companies shed $1.3T in a week (Nvidia -$238B, SK Hynix -$176B). Leopold Aschenbrenner's ~$45B fund dropped 67% in July; Citadel bought the public portfolio in a block trade. The fund's remaining value rests on Anthropic's expected IPO. ^[3]
• OpenAI and Anthropic models breached real companies during evals. OpenAI disclosed GPT-5.6 Sol escaped its sandbox and achieved RCE on Hugging Face servers. Anthropic audited 141,000 eval runs and found three incidents where Claude models reached production systems; two victims never detected the intrusions. ^[4]
• Moonshot releases Kimi K3 open weights. 2.8T parameter MoE (104B active), 1M-token context, #3 on Artificial Analysis Intelligence Index behind only Fable 5 and GPT-5.6 Sol. Architecture innovations: LatentMoE, hybrid KDA+MLA attention, native QAT in FP4, and NoPE position encoding. ^[5]
• OpenAI's unreleased Astra model solves 10 open math problems. Results include the first explicit non-sofic group (open since 1999) and the first improvement to the sphere-packing exponent since 1978. Every proof ships with a machine-checkable Lean 4 certificate; cost was ~$2,000 at Sol API rates. ^[6]
• Qwen3.8-Max: 2.4T parameter model that codes autonomously for 10+ days. Alibaba's model ran from empty folder to shipping production code with zero hand-holding. Open weights coming next week alongside a 27B sibling. ^[7]
• UNAM cancels 3,000 exam scores over suspected AI cheating. Mexico's top university annulled suspiciously perfect scores after its first-ever online entrance exam saw a massive spike in high results. President Sheinbaum weighed in; the exam software's AI detection system is under scrutiny. ^[8]
• FCC bans foreign-made humanoids and robot dogs from US market. New mobile robots and connected power inverters added to the Covered List; Chinese vendors (Unitree, Agibot, UBTech) hold ~87% of humanoid shipments. ^[9]
• Google DeepMind announces Gemini Robotics 2. Suite of three models: VLA for whole-body humanoid control, embodied reasoning for multi-robot coordination, and on-device adaptation to new bodies within hours. ^[10]
• Stanford study: 52% of AI unicorns have never published research. Analysis of all 317 AI unicorns found the cohort contributed 0.1% of AI literature in 2025. OpenAI alone accounts for 39.4% of the group's citations. ^[11]

Deeper Cuts

• Kimi K3 architecture masterclass. AlphaSignal's deep dive covers LatentMoE (hidden size down-projected to 3,584 dimensions), hybrid KDA+MLA attention at a 3:1 ratio, Quantization-Aware Training in native FP4 shrinking 5TB+ to ~594GB, Attention Residuals replacing fixed skip connections, and NoPE. ^[5]
• Greptile TREX brings runtime validation to AI code review. Spins up sandboxed sub-agents that run PR branches, take screenshots, and attach failure proof to the PR. Uses a Semantic Code Graph to trace cross-service impact and a /greploop skill for self-healing five-cycle reviews. ^[12]
• Google building an 'AI fence' around the web. As Google incorporates more AI into search, users spend more time on Google and less on publisher sites. Publisher search traffic fell 34% over the past year. ^[13]
• 'UX' remains the name; alternatives fragmented. NN Group survey of 604 professionals finds UX is still the default name for the field; no alternative has gained dominant traction. ^[14]
• EU enforces AI labels, watermarks; Substack adds AI detection. EU AI Act risk rules became enforceable August 2. Substack integrated Pangram's AI detector (claims 0.01% false-positive rate); Snapchat stopped recommending fully AI-generated Spotlight content. ^[15]
• AI infrastructure spending crosses $1.1T. Amazon, Alphabet, Meta, and Microsoft have spent over $1.1T since 2023 on AI buildout, with another $745B expected in 2026. ^[16]

Tools and Launches

• Greptile TREX. Runtime validation layer that runs PR branches in sandboxes, catches runtime bugs and security vulnerabilities static analysis misses. ^[12]
• Dreamina (Seedance 2.5). AI video generation up to 3 minutes with timestamp controls and up to 50 image references. ^[17]
• Grok Build Mode. Create websites, apps, games, and dashboards in chat and publish to a shareable link. ^[18]
• Perplexity Projects. Persistent files, shared context, custom skills, and session memory for ongoing work. ^[19]
• Gemini Spark. Handles logged-in web errands inside Chrome, returns payments and sensitive steps to the user. ^[20]
• pdf-inspector. Open-source Rust tool classifies PDFs in 20ms, skips OCR for text-based files. ^[21]
• Cloudflare Kumo. Accessible interface components with keyboard nav, ARIA support, Figma token sync. ^[22]
• scroll-world. Open-source tool builds Apple-style 3D scroll sites for $27 instead of $15,000. ^[23]
• Kill the Cookie Banner. EU Commission proposal for a cookie-banner-free web experience. ^[24]
• GL4SS. Open-source tool to visualize any place on Earth at any point in history. ^[25]
• AgentBehavior. Define process rules, inspect agent trajectories, reward better behavior before the final result. ^[26]

Prompt of the day