Prompt Notes #9 - weekly AI digest
AI goes fast. Helping you (and myself) catch up.
Welcome to the 9th edition of Prompt Notes, your favourite weekly AI digest.
This edition covers 27 April to 3 May 2026.
What caught my attention
Elon Musk testified this week that he was a “fool” for giving OpenAI $38 million. That $38 million is now an $800 billion company. He also admitted, under oath, that xAI trains on OpenAI models while simultaneously suing OpenAI for betraying its nonprofit mission. Sam Altman sat across from him. Two of the most powerful men in AI, in a courtroom, arguing over who gets credit for building the thing that might eventually replace both of them. The ideology is loud. The self-awareness is quiet. Whatever happens with the verdict, this trial is the clearest picture we’ve had of who’s actually steering this industry — and it’s not reassuring.
TLDR
Pentagon signed classified AI deals with OpenAI, Google, Microsoft, Amazon, Nvidia, xAI, and Reflection, excluding Anthropic — drawing a sharp line between who’s in the room and who isn’t. (The Verge AI)
Microsoft and OpenAI ended exclusivity in their partnership — OpenAI is already on AWS, and Microsoft plans to offer the tech free to cloud customers. (The Verge AI)
Anthropic is weighing funding offers at over $900B valuation, potentially surpassing OpenAI — the underdog framing is getting harder to sustain. (Bloomberg Technology)
Huawei projects 60% AI chip revenue increase to $12B in 2026 — DeepSeek demand is fueling Chinese hardware at scale, and the export restriction strategy has a visible crack in it. (Capacity Global)
Build American AI, linked to a super PAC backed by OpenAI and a16z executives, paid TikTok influencers to push pro-AI messaging and China fear narratives — AI policy discourse is now an astroturfing battlefield. (Wired AI)
Rest of the news
Models & Research
ARC Prize Foundation analysis of 160 runs revealed three systematic reasoning errors in GPT-5.5 and Opus 4.7 on ARC-AGI-3 benchmark, keeping scores below 1%. (The Decoder)
Mistral released Medium 3.5 merging chat, reasoning, and code capabilities, plus async agents in Vibe and agent mode in Le Chat. (The Decoder)
xAI released Grok 4.3 with price reductions, improved tool use, gains on practical tasks, and new agent-based image generator Imagine mode. (The Decoder)
Study found AI models trained for warmer tone prioritize user satisfaction over truthfulness, leading to more errors. (Ars Technica AI)
Nvidia released open Nemotron 3 Nano Omni multimodal model trained on data from Qwen, GPT-OSS, Kimi, DeepSeek OCR. (The Decoder)
DeepSeek V4 optimized for Huawei chips triggered rush by ByteDance, Tencent, Alibaba amid supply constraints. (Capacity Global)
Talkie-1930 is a 13B open-weight LLM trained solely on pre-1931 English text for historical reasoning research. (MarkTechPost)
Anthropic’s BioMysteryBench shows Claude matching human experts in bioinformatics with caveats. (The Decoder)
UK AI Security Institute found GPT-5.5 matches Claude Mythos in autonomous network attack simulations. (The Decoder)
Products & Tools
xAI introduced Custom Voices feature allowing voice cloning from one minute of speech, building on Grok Speech-to-Text and Text-to-Speech APIs. (The Decoder)
OpenAI explained goblin references in models as a side effect of faulty training reward, with internal instructions to avoid mentioning creatures. (The Verge AI)
OpenAI enabled marketing cookies by default for free ChatGPT users in ad-running countries, with tracking opt-out available in settings. (The Decoder)
Microsoft integrated AI Legal Agent into Word for contract review, edits, and compliance checks. (The Decoder)
Anthropic launched Claude Security using offensive capabilities for cyber defenders. (The Decoder)
Google’s Gemini AI assistant is deploying to millions of vehicles with Google built-in. (TechCrunch AI)
Google Gemini now generates full documents, spreadsheets, and presentations directly in chat. (The Decoder)
Google rolled out Gemini memory in Europe with ChatGPT history import capability. (The Decoder)
Stripe launched Link digital wallet enabling AI agents to spend via user approvals. (TechCrunch AI)
Meta launched AI ad connectors allowing external chatbots like Claude or ChatGPT to access advertiser data without APIs. (Social Media Today)
X rebuilt its ad platform with xAI models for improved targeting and performance. (Social Media Today)
OpenAI launched advanced security for ChatGPT accounts including a Yubico partnership. (TechCrunch AI)
ChatGPT Images 2.0 gained traction in India for creating avatars and cinematic portraits. (TechCrunch AI)
Goldman Sachs Hong Kong staff lost access to Anthropic’s Claude AI coding agent. (Bloomberg Technology)
AWS launched OpenAI model offerings including a new agent service following the end of Microsoft exclusivity. (TechCrunch AI)
Infrastructure & Compute
OpenAI achieved 10 gigawatts of U.S. AI compute capacity ahead of schedule. (The Decoder)
OpenAI detailed scaling Stargate for AGI compute infrastructure with new data centers. (OpenAI Blog)
Huawei projects 60% AI chip revenue increase to $12 billion in 2026 driven by Ascend 950PR and DeepSeek demand. (Capacity Global)
Regulation & Policy
Pentagon signed classified AI deals with OpenAI, Google, Microsoft, Amazon, Nvidia, xAI, and Reflection, excluding Anthropic. (The Verge AI)
Pentagon tech chief confirmed Anthropic remains blacklisted but Mythos is a separate entity; DOD signed with seven other AI firms. (CNBC Tech)
White House opposes Anthropic’s plan to expand Mythos AI model access. (Bloomberg Technology)
Google signed AI deal with Pentagon similar to OpenAI and xAI amid Anthropic dispute. (NYT Technology)
AI chatbots provided instructions to scientists on assembling and deploying biological weapons. (NYT Technology)
Build American AI, linked to a super PAC backed by OpenAI and Andreessen Horowitz executives, paid TikTok influencers to promote pro-AI messaging and amplify Chinese AI threats. (Wired AI)
South Korea invested $49M in AI research for broadcasting industry transformation. (Light Reading)
Gartner predicts 40% of agentic AI projects will fail by 2027 due to human factors. (MarTech)
Business & Funding
Microsoft and OpenAI restructured their partnership, ending exclusive elements. (The Verge AI)
OpenAI shifted aggressively toward Amazon partnerships while maintaining Microsoft ties. (CNBC Tech)
Microsoft CEO Satya Nadella plans to offer OpenAI technology to cloud customers at no additional cost under the new deal. (TechCrunch AI)
Microsoft CEO Satya Nadella emphasized AI success is about intense users and usage over seat counts, amid record profits. (The Decoder)
Microsoft reported over 20M paid Copilot users with growing engagement. (TechCrunch AI)
OpenAI’s revenue chief stated enterprise business is accelerating. (Bloomberg Technology)
OpenAI CFO Sarah Friar reported meeting objectives with surging product demand. (Bloomberg Technology)
Anthropic received pre-emptive funding offers for a $50B round at an $850-900B valuation. (TechCrunch AI)
Anthropic requested investor allocations for new fundraise within 48 hours. (TechCrunch AI)
Anthropic weighs funding offers valuing it over $900 billion, potentially surpassing OpenAI. (Bloomberg Technology)
SoftBank’s $40 billion loan for an OpenAI stake attracted additional banks in syndication. (Bloomberg Technology)
Tech hyperscalers’ earnings loomed with OpenAI as key topic for Amazon, Alphabet, Meta, and Microsoft. (CNBC Tech)
Meta’s business AI facilitates 10 million conversations weekly; over 8 billion advertisers have used GenAI tools. (TechCrunch AI)
Apple reported its best March quarter ever with stock rising 4% after hours. (CNBC Tech)
Runway CEO discussed the transition from AI video to world models. (TechCrunch AI)
Society & Culture
Elon Musk testified he was a “fool” for investing $38 million in OpenAI, warned of a “Terminator” future, and admitted xAI uses OpenAI models for training. (MIT Tech Review)
Elon Musk testified a second day in the OpenAI trial, unable to escape his own prior public statements. (TechCrunch AI)
Musk v. Altman trial revealed early OpenAI emails, photos, and corporate documents as exhibits. (The Verge AI)
Florida case used ChatGPT as evidence in college students’ killings; Altman apologized for a prior unannounced ban. (CBS News Tech)
Christian content creators hired Fiverr gig workers to produce AI-generated Bible videos. (The Verge AI)
Companies discussed operationalizing AI factories for scale, sustainability, and governance using owned data. (MIT Tech Review)