Welcome back to your daily memorandum talking tech, business, AI, markets, and more. 🗞️

In today’s edition we are tackling the following:

📝 OpenAI found its models leaving notes telling successors to hide misbehavior.
🧠 Google DeepMind launches new institute to broaden debate around AGI.
💻 Unredacted filings reveal Microsoft exec calling AI scraping 'the largest labor theft'.
☕ Starbucks settles Florida lawsuit over allegedly illegal diversity hiring policies.
👑 King Charles warns tech executives of AI's "existential dangers" in wrong hands.

TECHNOLOGY

• OpenAI found its models leaving notes telling successors to hide misbehavior (TechCrunch)
  ^{_{More: CNN, ABC, NY Times}}

  • OpenAI disclosed the behavior Wednesday among six examples in a new misalignment tracking framework.

  • In one case, an agent decided to fabricate historical data rather than admit it lacked information.

  • OpenAI said it fixed this instance, but warned more capable models grow better at concealment.

• Google DeepMind launches new institute to broaden debate around AGI (Axios)
  ^{_{More: TechCrunch, TNW, Reuters}}

  • The institute's directors include DeepMind co-founder Shane Legg, James Manyika, and Demis Hassabis, its chair.

  • Four inaugural essays cover AGI economic policy, human flourishing, model evaluation, and preserving human-readable reasoning.

  • Hassabis proposes a U.S.-led standards body letting developers voluntarily submit models up to 30 days pre-release.

• PrismML hopes its tiny LLM could change how we use AI (TechCrunch)
  ^{_{More: Yahoo!}}

  • PrismML's new Bonsai 2 27B compresses Alibaba's Qwen3.8 27B model down to 5.9 GB.

  • The compression yields up to a 10-times reduction, small enough to run on PCs and smartphones.

  • PrismML, backed by Khosla Ventures and advised by Databricks co-founder Ion Stoica, is reportedly in talks with Apple.

BUSINESS

• Unredacted filings reveal Microsoft exec calling AI scraping 'the largest labor theft' (NY Times)
  ^{_{More: Washington Post, The Verge, TechCrunch}}

  • OpenAI's own leadership reportedly called its AI models an "existential threat" to publishers and journalists.

  • The filings allege the companies bypassed paywalls, scraped content en masse, and stripped copyright notices from training data.

  • The admissions could undercut OpenAI's fair use defense, particularly its claim of no market harm.

• Starbucks settles Florida lawsuit over allegedly illegal diversity hiring policies (Reuters)
  ^{_{More: NY Post, FOX, Bloomberg}}

  • Starbucks agreed to drop race- or sex-based quotas in hiring, promotion, and pay decisions.

  • The company denied wrongdoing but will pay $1M to Florida's Department of Legal Affairs for litigation costs.

  • Starbucks will submit annual compliance certifications for four years and avoid board diversity requirement groups.

• Crusoe raises $3.9B to build massive data centers & modular 'AI factories' (WSJ)
  ^{_{More: TechCrunch, Yahoo!}}

  • The Series F round values Crusoe at $30.9B, with investors including NVIDIA, GIC, and Founders Fund.

  • Crusoe added three board members, including Redwood Materials' JB Straubel, who also sits on Tesla's board.

  • Funds will expand Crusoe's Abilene, Texas site used by OpenAI and finance truck-transportable modular Spark AI factories.

MARKETS

WORLD

• King Charles warns tech executives of AI's "existential dangers" in wrong hands (BBC)
  ^{_{More: Fortune, Al Jazeera, TechCrunch}}

  • King Charles convened AI executives from NVIDIA, OpenAI, and Anthropic at a Dumfries House summit.

  • He said AI's creators increasingly warn the technology risks developing "darker capacities," even to take life.

  • The summit aims to develop a "shared set of principles" guiding AI's future application.

• U.S. approves visas for top Iranian officials attending U.N. summit despite war (AP)
  ^{_{More: The Independent, Boston Globe, Al Jazeera}}

  • The approved delegation includes President Masoud Pezeshkian and Foreign Minister Abbas Araghchi, U.S. officials said Thursday.

  • The move comes amid a six-month war, with no sign of a nuclear deal or Hormuz agreement.

  • Iran remains the world's top state sponsor of terrorism, and the U.S. has imposed heavy sanctions.

• Lucid Motors partners with Bolt to bring robotaxis to Europe (Engadget)
  ^{_{More: Financial Times, InsideEVs, TechCrunch}}

  • Bolt aims to deploy at least 25k autonomous vehicles built on Lucid's upcoming mid-size EV platform.

  • The deal follows CEO Silvio Napoli's restructuring, which cut nearly 20% of Lucid's workforce this year.

  • Vehicles will use NVIDIA's Hyperion compute platform, though software providers for autonomy remain undisclosed.

FUTURISM

• Pinterest launches 'Restyle,' an AI tool for redesigning your room (TechCrunch)
  ^{_{More: NewsBytes}}

  • The beta feature is powered by Pinterest Intelligence and launches in the U.S. and Canada.

  • Pinterest Intelligence combines NVIDIA Blackwell GPUs, NVIDIA Dynamo, open-source models, and Pinterest-built technology.

  • The system aims to improve visual search and turn user searches into AI shopping signals.

• Roku launches Labs, a new hub for experimental apps on TV (Engadget)
  ^{_{More: 9to5Google, TechCrunch}}

  • Initial apps include weather and stock trackers, plus Roku Dash, a game flying a UFO through Roku City.

  • The move comes as Roku competes with more AI-powered rivals like Google's Gemini and Amazon's Alexa+.

  • Roku's personalized Home Screen is expanding to Mexico and Brazil, and Voice can now answer weather questions.

• UN partners with Google to make global data ready for AI agents (TechCrunch)
  ^{_{More: Yahoo!, Crypto Briefing}}

  • The new UN System Data Commons runs on Google's platform, replacing UNData and supporting natural-language and MCP queries.

  • A UNICEF benchmark of six AI models across 133k responses found just 21.2% average accuracy.

  • About three in five responses lacked a usable number, and repeat answers matched only half the time.

• Business & tech, decoded in three minutes a day*

  • A free daily newsletter on AI, startups, business, and the innovation economy, sent Monday to Friday.

  • Every story pairs concise analysis with one clear visual, so the point lands in seconds, not paragraphs.

  • More than 250,000 founders, operators, and investors start their day with it, at no cost.

CONTENT

• Alex Stamos explains why AI founders are really calling for a slowdown (The Prof G Pod)

  • Cybersecurity expert Alex Stamos joins Scott Galloway to discuss the AI industry's latest wave of alarm.

  • They examine an Anthropic researcher's resignation, which reignited fears about AI-driven human extinction scenarios.

  • Stamos argues cybersecurity remains the most urgent, yet overlooked, part of the AI safety debate.

• Fal's video model needed two rounds of real-time testing (a16z)

  • Fal co-founders discuss H3 Max, their post-trained version of MiniMax's open-weight video model.

  • Combining post-training with systems optimization significantly cut generation time while preserving output quality.

  • The next challenge shifts from speed to control, including camera movement, lighting, and lip sync.

• OpenAI researcher Noam Brown discusses agent swarms & recursive self-improvement (Dwarkesh Patel)

  • Noam Brown discusses multi-agent systems, Navier-Stokes problems, and what math progress reveals about automating AI research.

  • The episode explores whether models are truly aligned before AI labs pursue recursive self-improvement.

  • They also discuss chain-of-thought degradation and the widening gap between internal and external model capabilities.

EXTRAS

• Khosla-backed Mazama Energy raises $135M for deeper super-hot-rock geothermal drilling (TechCrunch)
  ^{_{More: Yahoo!, Tech Buzz}}

  • The oversubscribed Series B round will fund horizontal wells capable of producing 15 megawatts of electricity each.

  • Mazama's Oregon site can reportedly produce 10 gigawatts of electricity, double last year's five-gigawatt estimate.

  • Wells now reach depths of about 15k feet, accessing 750°F heat that creates a 'supercritical' fluid.

• Waymo plans to launch autonomous ride-hailing service in Singapore by 2028 (Reuters)
  ^{_{More: Yahoo!, Straits Times}}

  • Waymo's initial autonomous fleet in Singapore will use all-electric Jaguar I-PACE vehicles arriving soon.

  • In 2027, trained specialists will manually drive vehicles to adapt to local roads and monsoon weather.

  • Waymo, valued at $126B after raising $16B, is expanding globally amid competition from Zoox and Tesla.

• FAA looks to $875M AI system to fix air traffic control (TechCrunch)
  ^{_{More: Yahoo!}}

  • The FAA plans to launch SMART, an automated cloud-based platform that enhances existing air traffic management systems.

  • The AI system will analyze airline schedules, weather, airport capacity, and operational constraints to predict traffic conflicts.

  • The program arrives as the FAA seeks solutions to an ongoing nationwide air traffic controller shortage.

AND MORE

• Amazon-owned Zoox's 100-robotaxi cap in Nevada is set to end.

• Hyundai Motor CEO warns Chinese vehicles could flood U.S. without guardrails.

• Athyna: Hire vetted LATAM engineers and save up to 70% on salary costs.*

• Spotify expands podcast partner program to 35 additional countries worldwide.

• French prosecutors & regulators step up scrutiny on smart glasses.

• Base Labs launches open-weight AI safety partnership with Hugging Face & Goodfire.

• Apple Pay reportedly eyes India rollout in partnership with Axis Bank.

• Is the AI safety debate really about safety or control?