Prompt Notes #10 - weekly AI digest
AI goes fast. Helping you (and myself) catch up.
Welcome to the 10th edition of Prompt Notes, your favourite weekly AI digest.
This edition covers 18 May to 24 May 2026.
What caught my attention
Google shipped 100 announcements at I/O this week. Omni, Spark, Antigravity 2.0, Gemini 3.5, XR glasses, world models wired to Street View. A genuinely staggering amount of working product. The main labs should be worried, if you ask me.
Speaking of which… the two labs everyone treats as the main characters spent the week showing their books and fighting in court over who gets to steal a non-profit. We learned that OpenAI burned $1.22 for every dollar it earned. xAI lost $6.4 billion last year and is now buying $2.8 billion in gas turbines to keep the lights on. A jury just sided with OpenAI in Musk’s suit over technicalities.
TLDR
The 5 stories that mattered this week:
Alibaba’s Qwen3.7-Max ran autonomously for 35 hours to optimize code for its own custom chip and now matches Claude Opus 4.6 on benchmarks. (The Decoder)
DeepSeek is raising ~$10 billion at a $45 billion valuation while telling investors AGI research comes before short-term profit. (The Decoder)
Google dropped 100 announcements at I/O 2026, including Gemini Omni, Spark, Antigravity 2.0, Gemini 3.5, and Android XR glasses. (Google AI Blog)
An OpenAI reasoning model produced the first AI-generated proof deemed suitable for a top math journal, disproving Erdős’s 1946 unit-distance conjecture. (The Decoder)
Trump withdrew a planned AI safety executive order after last-minute calls from Musk, Zuckerberg, and Sacks. (The Decoder)
Rest of the news
Models & Research
Alibaba released Qwen3.7-Max, which ran autonomously for 35 hours to optimize code for its custom chip and matches Claude Opus 4.6 on benchmarks. (The Decoder)
Google demonstrated Gemini Omni, an anything-to-anything model capable of generating and editing video from images, audio, and text inputs. (The Verge AI)
Cohere open-sourced its Command A+ model under the Apache 2.0 license. (The Decoder)
Stability AI released Stability Audio 3.0, capable of generating up to six-minute songs with a smaller on-device variant limited to two minutes. (TechCrunch AI)
Gemini 3.5 Flash costs 5.5 times more to run than its predecessor and exceeds Gemini 3.1 Pro costs on agent tasks. (The Decoder)
An OpenAI reasoning model produced the first AI-generated proof deemed suitable for a top mathematics journal by disproving Erdős’s 1946 unit-distance conjecture using algebraic number theory. (The Decoder)
Claude Mythos Preview identified over 10,000 critical vulnerabilities across partner systems in Project Glasswing, outpacing patching capacity. (The Decoder)
Google DeepMind combined the Genie 3 world model with Street View data to produce explorable AI environments from real locations. (The Decoder)
Microsoft Copilot produced stereotypical results on identical datasets with altered country labels during data analysis tests, while thinking models avoided the error when selected. (The Decoder)
Demis Hassabis stated humanity is in the foothills of the singularity, while Yann LeCun said current AI systems are not genuinely intelligent and Oriol Vinyals noted models lack experience-based learning. (The Decoder)
Products & Tools
Google claimed its Antigravity 2.0 agents built a complete operating system from a single prompt at a reported cost of $916 in API fees. (AI Snake Oil)
Google introduced the Gemini Spark AI agent for autonomous task execution, addressing limitations seen in prior personal assistant tools. (The Verge AI)
OpenAI launched a ChatGPT beta plugin for PowerPoint that generates and edits slides from notes or images, with a warning to save work beforehand. (The Decoder)
Figma added an AI assistant that accepts natural language prompts to generate, edit, and iterate designs on the collaborative canvas. (TechCrunch AI)
Google demonstrated prototype Android XR glasses that overlay Gemini-powered translation and navigation in the user’s field of view. (TechCrunch AI)
Google Search now uses Gemini to generate custom product explainers and integrate ads with AI Overviews. (The Verge AI)
Google added voice-based conversational search to Gmail allowing users to query inbox contents via Gemini. (TechCrunch AI)
Google added an experimental “Agentic Browsing” category to Lighthouse that checks websites for llms.txt files. (The Decoder)
Google’s AI Overviews inserted unrelated chatbot-style responses when users searched the term “disregard.” (The Verge AI)
YouTube introduced an AI remix tool for Shorts that generates alternate video versions using Omni models. (Social Media Today)
Spotify released an ElevenLabs-powered tool that generates audiobooks without requiring exclusive distribution rights. (TechCrunch AI)
Spotify launched the Studio AI app that generates personalized daily briefings and podcasts from user data. (The Verge AI)
A Wired reporter generated lifelike video clones of themselves using Gemini’s AI avatar tool. (Wired AI)
SAP integrated Mistral AI models to assist customers migrating legacy systems to S/4HANA. (The Decoder)
Tesla activated Full Self-Driving (Supervised) in China after extended regulatory delays. (CNBC Tech)
OpenAI was named a Leader in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents. (OpenAI Blog)
Infrastructure & Compute
xAI has shifted focus to natural gas power while SpaceX pursues orbital data centers, moving away from earlier solar power emphasis. They will purchase $2.8 billion in natural gas turbines over three years, according to SpaceX's filing disclosures. (TechCrunch AI)
Nvidia CEO Jensen Huang identified CPUs for AI agents as a new $200 billion market opportunity. (TechCrunch AI)
Google and Blackstone formed a joint venture to deploy TPU-based AI cloud capacity targeted for 2027 availability. (Capacity Global)
Regulation & Policy
President Trump withdrew a planned AI safety executive order following calls from Elon Musk, Mark Zuckerberg, and David Sacks. (The Decoder)
Anthropic is expected to continue supplying Claude to the NSA despite a Pentagon supply chain risk flag, with the deal excluding the prior “any lawful use” clause. (The Decoder)
The Pentagon began testing rival AI models to identify replacements for Anthropic’s Claude among 25 power users. (Bloomberg Technology)
Singapore signed AI partnership agreements with Google and OpenAI, including a $234 million OpenAI commitment. (CNBC Tech)
OpenAI launched the multi-year “OpenAI for Singapore” partnership for talent development and public sector deployment. (OpenAI Blog)
Hackers are shifting from simple prompt exploits to targeting chatbot personalities in newer AI systems. (The Verge AI)
Business & Funding
Anthropic is set to close a funding round exceeding $30 billion at a valuation above $900 billion. (Bloomberg Technology)
Anthropic is projected to report its first profitable quarter with $559 million operating profit on $10.9 billion Q2 revenue. (The Decoder)
DeepSeek is raising approximately $10 billion at a $45 billion valuation while prioritizing AGI research over short-term commercialization. (The Decoder)
DeepSeek is assembling a Beijing team to build “Deepseek Code,” an agent competing with Claude Code and Codex. (The Decoder)
Chinese labs are delivering frontier-level model performance at significantly lower inference costs than U.S. counterparts, raising questions over OpenAI and Anthropic IPO economics. (CNBC Tech)
OpenAI reported $5.7 billion in Q1 2026 revenue with an adjusted operating margin of minus 122 percent. (The Decoder)
OpenAI is targeting a possible September IPO following resolution of the Musk lawsuit. (TechCrunch AI)
A jury rejected Elon Musk’s claims against OpenAI partly due to the timing of the filing. (TechCrunch AI)
SpaceX plans a June 12 IPO that analysts suggest could represent the largest public float in history. (CNBC Tech)
xAI recorded a $6.4 billion loss in 2025 according to SpaceX’s IPO filing. (TechCrunch AI)
SoftBank insiders expressed concerns over Masayoshi Son’s $60 billion commitment to OpenAI. (Bloomberg Technology)
OpenAI appointed Chris Lehane as global affairs chief to manage regulatory and reputational matters. (Wired AI)
Luma and Wonder Project formed Innovative Dreams to produce longer-form AI video content beyond short clips. (The Verge AI)
GitHub lost ground in AI coding tools to newer competitors despite user growth under Microsoft ownership. (CNBC Tech)
Events & Industry
Google listed 100 announcements from I/O 2026 including Gemini Omni, Antigravity 2.0, and Universal Cart. (Google AI Blog)
Stratechery analyzed Google I/O 2026 announcements and DeepMind’s alignment with Google’s commercial priorities. (Stratechery)
Anthropic hosted the Code with Claude developer event showcasing enterprise coding workflows. (MIT Tech Review)
Society & Culture
Benedict Evans argues that assigning precise AI exposure scores to jobs, companies, and industries is mostly impossible due to unpredictable job changes and measurement challenges. (Benedict Evans)
A Commonwealth Short Story Prize regional winner published by Granta magazine showed characteristics of AI-generated text. (The Verge AI)
Gartner research found 70 percent of B2B buyers prefer digital self-service yet still seek human validation over AI outputs. (MarTech)