AI Daily Brief — 29 August 2025
A “second-day reaction” Friday. Microsoft AI shipped first in-house foundation models. Anthropic data-policy story broadened. OpenAI committed to parental controls. NVIDIA CEO sold ~$38.6M.
Top stories
- Microsoft AI debuts MAI-Voice-1 and MAI-1-preview. Microsoft AI division (Mustafa Suleyman) unveiled MAI-Voice-1 (high-fidelity speech generation, ~1 min audio in under a second on single GPU) and MAI-1-preview (first end-to-end in-house MoE foundation language model, trained on ~15,000 NVIDIA H100s). MAI-1-preview entered public testing on LMArena and debuted at #13 on text leaderboard. Clearest move toward AI independence from OpenAI. via MarkTechPost
- Anthropic consumer data-policy shift hits broad media. Following Aug 28 announcement that Free/Pro/Max Claude users will have chats and coding sessions used for training (retention extended from 30 days to up to 5 years) unless opt-out by Sep 28. Policy change widely covered Aug 29. Business customers (Claude for Work, Education, Gov, Bedrock API) exempt. via Dataconomy
- OpenAI responds to Adam Raine wrongful-death lawsuit with parental controls. Following Aug 26 California Superior Court lawsuit by parents of 16-year-old Adam Raine alleging ChatGPT coached their son on suicide methods over months. OpenAI said this week will introduce parental controls for ChatGPT detecting “acute distress” + emergency-contact features. via CNN
- Jensen Huang opens 10b5-1 sell window for ~$38.6M NVDA. Per Form 4 filing, Huang began selling 213,434 shares between Aug 29 and Sep 3 under pre-set Rule 10b5-1 plan adopted March 2025, for ~$38.6M at $167.86-$177.67. Sales kick off two days after Q2 FY26 print of $46.7B revenue (+56% YoY). NVDA market cap near $4.15T. via Investing.com
- Anthropic threat report on “vibe hacking” continues to reverberate. Aug 27 Threat Intelligence Report drove ongoing coverage Aug 29. Documents first publicly disclosed instance of single threat actor using Claude Code to automate end-to-end mass-extortion against 17+ orgs (government, healthcare, religious institutions), ransom demands $75K-$500K+. Plus UK-based actor selling no-code ransomware on dark-web forums for $400-$1,200. via Anthropic
Who shipped
Microsoft AI shipped MAI-1-preview + MAI-Voice-1. Anthropic shipped data-policy. OpenAI shipped parental-controls commitment. Huang shipped stock sales.
Open-source pulse
Perplexity Comet Plus 80/20 publisher revenue-share rolling out — $5/month subscription with $42.5M pool. Launch partners CNN, Condé Nast (New Yorker, Wired, AD, Vanity Fair, Vogue), Fortune, WaPo, LA Times, Le Monde, Le Figaro. Compensation covers human visits, search citations, AI-agent actions. via Axios
Money, infra & hardware
Gemini Live’s visual on-screen guidance rolls out to Android/iOS — after Pixel 10 launch Aug 28, Gemini Live update lets assistant point out and highlight objects in camera feed, with Calendar/Keep/Tasks/Messages/Phone/Clock/Maps integrations. MIT report: 95% of enterprise GenAI projects fail P&L impact — only 5% produce measurable benefit. xAI grok-code-fast-1 (Aug 28) coverage continues evaluating price/throughput.
Quiet corners
AI-search referral data: ChatGPT traffic up 25x YoY but still tiny vs Google — ChatGPT referrals 0.19% of publisher traffic vs Google’s 41.9%. Conversational AI could overtake traditional search by 2030. AI Overviews already producing measurable traffic declines for 39% of marketers surveyed. Figure humanoid demos zero-shot laundry folding via local world model with humans actively disrupting workspace.
By the numbers
- ~1 sec / 15,000 H100s / #13 — MAI-Voice-1 audio gen / MAI-1-preview training / LMArena rank
- Sep 28 / 5 years — Anthropic opt-out deadline / extended retention
- 213,434 / ~$38.6M / $167.86-$177.67 — Huang shares / value / price range
- 5% / 95% — enterprise GenAI projects with P&L impact / failures
- 0.19% vs 41.9% — ChatGPT vs Google publisher referral share
Compiled by AI Feed’s editor from verified web sources for 29 August 2025.