Prompt Notes #11 - weekly AI digest

AI goes fast. Helping you (and myself) catch up.

Welcome to the 11th edition of Prompt Notes, your favourite weekly AI digest.

This edition covers 1 June to 7 June 2026.

Note: I forgot to publish last week’s newsletter… Apologies :)

What caught my attention

Anthropic confidentially filed for IPO this week. A five-year-old at a $965 billion valuation, with $47 billion annualised revenue. In terms of funding: $132 billion raised across 18 rounds - the last one a $65 billion Series H closed six weeks ago.

SpaceX and OpenAI are also heading to public markets in the same window, creating a historic concentration of capital. The big question is whether public markets know how to price companies with no comparable.

We’ll see this Friday with SpaceXAIStarNeuralink, or whatever the name is nowadays :).

One thing about SpaceX: the financial engineering and rule-twisting around this IPO is quite something. I highly recommend a listen to this.

TLDR

The 5 stories that mattered this week:

• Google pays SpaceX $920M/month for xAI compute - 32-month deal worth roughly $29.4B total; Google buying capacity from a direct competitor’s infrastructure is a sign of how broken the compute supply chain still is. (CNBC Tech)

• xAI trained on Claude outputs after access was revoked - not a technical story, an ethics and IP story; raises real questions about what “access revoked” actually means in practice. (The Decoder)

• Anthropic’s Mythos powering NSA offensive cyber ops - the safety-first AI company’s most capable model reportedly running offensive operations against China and Iran; the gap between the mission statement and the deployment reality is now public. (The Decoder)

• Microsoft announces seven in-house models at Build 2026 - a direct shot at OpenAI dependency; the partner relationship is quietly becoming a competitive one. (CNBC Tech)

• OpenAI and the Trump administration are negotiating a government equity stake - a sovereign wealth fund taking a stake in a private AI lab is structurally unprecedented; it reframes what “national AI strategy” actually means. (The Decoder)

Rest of the news

Models & Research

• OpenAI plans to rebuild ChatGPT as a full agent-based superapp integrating coding tools and third-party services. (The Decoder)

• OpenAI’s ChatGPT memory upgrade now learns user habits automatically without explicit commands. (The Neuron)

• ChatGPT’s “Dreaming” memory system stores narrative user dossiers organized by work, hobbies, and travel. (The Decoder)

• Google DeepMind’s Gemma 4 12B runs multimodal inference on laptops with only 16 GB RAM under Apache 2.0. (The Decoder)

• Alibaba released Qwen3.7-Plus, a multimodal agent model that autonomously built a 10,000-line vocabulary app. (The Decoder)

• xAI released Grok Imagine 1.5, an image-to-video model supporting 720p cinematic output. (The Decoder)

• Ideogram 4.0 launched as an open-weight text-to-image model with native 2K resolution and improved typography. (The Decoder)

• A new open-source voice model processes continuous audio and decides every 0.4 seconds whether to respond. (The Decoder)

• Sakana AI launched a research lab focused on recursive self-improvement to reduce reliance on compute scaling. (The Decoder)

• Nvidia released Nemotron 3.5 Content Safety, a customizable multimodal safety model for enterprise use. (Hugging Face Blog)

• Sebastian Raschka published a curated list of notable LLM research papers from January to May 2026. (Ahead of AI)

• SISTRIX data showed ChatGPT citation patterns shifted after the GPT-5.5 release. (SEJ Generative AI)

• Ethan Mollick reflected on the transition from co-intelligence to co-existence with AI systems. (One Useful Thing)

Products & Tools

• ChatGPT added Lockdown Mode that disables web access, Deep Research, and Agent Mode to reduce prompt-injection risks. (The Decoder)

• Perplexity introduced “Search as Code,” allowing models to generate custom Python search pipelines instead of fixed APIs. (The Decoder)

• Perplexity introduced a hybrid orchestrator that routes tasks between local and cloud models. (The Decoder)

• OpenAI expanded Codex with role-specific plugins for non-developers in data analysis, sales, and banking. (The Decoder)

• Microsoft launched Scout, an OpenClaw-inspired personal assistant integrated with Microsoft 365. (TechCrunch AI)

• Microsoft introduced a portable policy specification allowing developers to control AI agent behavior. (TechCrunch AI)

• Microsoft announced seven in-house generative models at Build 2026, outperforming Google in image generation while trailing on reasoning. (The Decoder)

• Meta launched an AI-generated “For You” feed of clickbait-style stories inside its standalone AI app. (The Verge AI)

• Meta launched an AI creator assistant on Facebook that answers performance and comment questions. (TechCrunch AI)

• Meta will offer the Meta Business Agent through its paid Meta One subscription. (CNBC Tech)

• Meta’s AI agent for WhatsApp Business became available globally with token-based pricing. (TechCrunch AI)

• Nous Research released Hermes Desktop, an open-source AI agent application under the MIT license. (The Decoder)

• Apple approved Poke as the first AI agent on its Messages for Business platform. (TechCrunch AI)

• Morgan Stanley will open its wealth-management platform to external AI agents. (CNBC Tech)

• GitHub outlined its roadmap for supporting the growing ecosystem of AI coding agents. (Latent Space)

• Snowflake introduced AI tools that let marketers apply governed intelligence across the full customer journey. (MarTech)

• The Verge examined Google’s Gemini Spark agent and noted its effectiveness alongside privacy and cost concerns. (The Verge AI)

• Nvidia introduced Cosmos 3-powered systems for physical AI research and autonomous agent workflows. (AI Business)

• Reddit users developed Claude-based tools to bypass inflated FIFA World Cup ticket prices. (Wired AI)

Infrastructure & Compute

• Apollo and Blackstone closed a $35 billion financing package for Anthropic’s AI infrastructure. (Bloomberg Technology)

• Google agreed to pay SpaceX $920 million monthly for 32 months to access xAI compute capacity. (CNBC Tech)

• Oracle and OpenAI began construction of the Stargate data-center campus in Michigan. (Data Center Dynamics)

• Megaport secured AU$458.9 million in AI-related contracts and plans to raise funds for an inference cloud. (Data Center Dynamics)

• Microsoft and Meta executives described the data-center industry as reaching a “real inflection point.” (Capacity Global)

• Stratechery analyzed Nvidia’s AI PC strategy versus Microsoft’s agent-centric vision at Build 2026. (Stratechery)

• MIT Technology Review discussed agentic AI’s potential to rehumanize strained global healthcare systems. (MIT Tech Review)

Business & Funding

• Anthropic’s annualized revenue reached $47 billion in May 2026 ahead of its planned IPO. (TechCrunch AI)

• Anthropic hired OpenAI’s second hardware engineer, Clive Chan, amid both firms’ IPO preparations. (The Decoder)

• Anthropic reported Claude now generates over 90 percent of its production code and advocated for a global AI pause mechanism. (The Decoder)

• DeepSeek became Ramp’s top trending software vendor in June 2026 as US firms adopt cheaper Chinese models. (The Decoder)

• Enterprises are adopting model routing to reduce AI spend, putting pressure on frontier lab revenue models. (CNBC Tech)

• xAI reportedly trained coding models on Claude outputs for months after access was revoked. (The Decoder)

• Investors continue to back both OpenAI and Anthropic despite their competitive positioning. (Wired AI)

• Uber imposed usage caps on AI coding tools to manage rising costs. (Bloomberg Technology)

• Airbnb CEO Brian Chesky announced plans to launch an internal AI research lab. (TechCrunch AI)

• SoftBank CEO Masayoshi Son described OpenAI’s next model as a sign of superintelligence arriving sooner than 2036. (CNBC Tech)

• Reid Hoffman announced he will leave Microsoft’s board of directors. (NYT Technology)

• Sam Altman described the next phase of AI as “proactive” systems that act continuously in the background. (The Decoder)

• Satya Nadella discussed Microsoft’s core competencies, OpenAI relationship, and agentic platform strategy. (Stratechery)

• Perplexity CEO Aravind Srinivas stated “value per watt per user” will determine AI race winners. (CNBC Tech)

• Google co-founder Sergey Brin stated AI is on a path to AGI but could not describe what follows. (SEJ Generative AI)

• Satya Nadella publicly rejected an internal proposal to make Microsoft’s Scout agent deliberately addictive. (The Decoder)

• Microsoft positioned itself as a direct AI competitor to OpenAI at Build 2026. (The Verge AI)

• OpenAI began a retro-themed advertising campaign for ChatGPT. (NYT Technology)

• French President Macron invited Sam Altman to the G7 summit. (CNBC Tech)

• Meta Chief AI Officer Alexandr Wang stated future models will differentiate via consumer health capabilities. (Bloomberg Technology)

Regulation & Policy

• President Trump signed an executive order inviting voluntary early government access to frontier models. (CNBC Tech)

• Sam Altman met with US lawmakers and Trump administration officials regarding AI policy. (CNBC Tech)

• OpenAI and the Trump administration are discussing a government equity stake via a proposed Public Wealth Fund. (The Decoder)

• OpenAI published a blueprint for federal governance of frontier AI covering safety and national security. (OpenAI Blog)

• OpenAI released its public-policy agenda addressing safety, youth protection, and workforce transition. (OpenAI Blog)

• Florida sued OpenAI and Sam Altman personally, treating ChatGPT as a defective product and public nuisance. (The Decoder)

• Bernie Sanders called for the US public to receive a 50 percent stake in OpenAI and Anthropic. (Data Center Dynamics)

• White House AI policy adviser Sriram Krishnan is stepping down from his role. (Bloomberg Technology)

Society & Culture

• Anthropic’s Mythos model is reportedly supporting NSA offensive cyber operations against China and Iran. (The Decoder)

• SKT joined Anthropic’s Project Glasswing to scan codebases for security vulnerabilities. (Light Reading)

• Anthropic expanded Project Glasswing to 150 organizations across 15 countries for critical-infrastructure cybersecurity. (Capacity Global)

• Proliferation of AI agents is creating new security challenges for telecommunications operators. (Fierce Network)

• Hackers continue to exploit Meta AI agents to gain control of Instagram accounts. (Social Media Today)

• xAI reportedly trained coding models on Claude outputs for months after access was revoked. (The Decoder)