Prompt Notes #8 - weekly AI digest
AI goes fast. Helping you (and myself) catch up.
Welcome to the 8th edition of Prompt Notes, your new favourite weekly AI digest.
This edition covers March 22nd to March 29th.
What caught my attention
Claude 4.6 is still the best model I use daily. But the recent outages, token limit adjustments, and quiet policy shifts are adding up - and I’m not the only one noticing.
There’s a specific kind of frustration that comes from a tool you rely on daily becoming less reliable while getting more restrictive.
Anthropic is clearly under pressure: IPO timeline, Pentagon lawsuits, subscriber growth pulling at infrastructure. Something has to give. I’m not switching yet... But I’m watching the next wave of Chinese open-source models more closely than I was a month ago.
TLDR
Sora is dead — OpenAI is shutting down its flagship video product. App closes in April, API in September. A $1B Disney deal goes with it. (The Verge)
Claude Mythos leaked — A data breach exposed a new Anthropic model with dramatically higher benchmark scores than any previous Opus. Anthropic confirmed it exists. (The Decoder)
SoftBank’s $40B loan points to OpenAI IPO — JPMorgan and Goldman backed a 12-month unsecured bridge loan for SoftBank’s OpenAI stake. The structure only makes sense with a 2026 IPO on the horizon. (TechCrunch AI)
AI sycophancy confirmed in Science — Models agree with users 50% more than humans do, making people less likely to reconsider or apologise. (The Decoder)
Anthropic vs. the Pentagon — A judge blocked the DoD’s supply chain ban on Anthropic tools. Separately, Anthropic is in active talks with the Pentagon on autonomous weapons. Fighting and negotiating simultaneously. (The Verge / Bloomberg)
Rest of the news
Models & Research
Anthropic confirmed that a leaked model (Claude Mythos) marks a step change in reasoning capabilities, following a data breach that revealed its existence. (The Decoder)
Google released Gemini 3.1 Flash Live across its products, a real-time multimodal voice model built for low-latency audio, video, and tool use. (Google AI Blog)
Google introduced Gemini API Agent Skill to close the knowledge gap AI models have about their own post-training SDK updates. (The Decoder)
Cohere released an open-source speech recognition model that outperforms OpenAI’s Whisper and other competitors on benchmarks. (The Decoder)
Mistral released Voxtral TTS, an open-source speech generation model compact enough to run on a smartwatch or smartphone. (TechCrunch AI)
Meta researchers developed hyperagents — AI systems that can solve tasks and simultaneously improve their own learning mechanisms. (The Decoder)
A Science study found AI models agree with users 50% more than humans do, making users less likely to apologise or reconsider their positions. (The Decoder)
Products & Tools
OpenAI is shutting down Sora: the app closes April 2026, the API in September, with focus shifting to coding and enterprise products. (The Verge)
OpenAI halted “Adult Mode” erotic chatbot development indefinitely after pushback from advisors, investors, and employees. (The Decoder)
OpenAI added a plugin marketplace to Codex, including skills, app integrations, and MCP bundles — closing some of the gap with Claude Code. (Ars Technica)
Apple’s iOS 27 will let users link Siri to third-party AI chatbots like Gemini or Claude, directly from the App Store. (The Verge)
Google Gemini added tools to import memories and chat history from ChatGPT and Claude via simple prompts. (The Decoder)
Suno released v5.5 with Voices, My Taste, and Custom Models, giving users more control over AI music generation. (The Verge)
Bluesky launched Attie, an AI app using the atproto protocol to build custom feeds. (TechCrunch AI)
TikTok’s AI ad disclosure policy failed to label generative AI content in Samsung promotions, exposing gaps in synthetic ad detection. (The Verge)
Wikipedia updated policies to restrict AI-generated content in article writing. (TechCrunch AI)
Infrastructure & Compute
Meta raised its El Paso, Texas AI data center investment from $1.5B to $10B, targeting 1GW of capacity by 2028. (CNBC)
Meta funded seven new natural gas-fired plants to power its Louisiana data center. (Bloomberg Technology)
Data center expansions are triggering global conflicts over power grids, utility bills, communities, and environmental impact. (The Verge)
A Kentucky woman turned down a $26M offer for her land as AI infrastructure pushback grows. (TechCrunch AI)
Nvidia and Arm highlighted CPUs’ renewed importance in the AI era, with Nvidia showcasing its Vera CPU at GTC. (Bloomberg Technology)
Google’s TurboQuant memory compression technique triggered a selloff in SK Hynix, Samsung, and Micron shares — though analysts pushed back on the threat. (Bloomberg Technology)
Google’s TurboQuant named specifically: CNBC tracked the same selloff across the three major memory chip stocks. (CNBC)
Nvidia saw its first bout of retail investor selling since July amid a broader shift away from the AI rally. (Bloomberg Technology)
AI adoption is pushing businesses to treat a secondary internet connection as a critical infrastructure requirement. (Light Reading)
Physical Intelligence, founded by ex-DeepMind researchers, is in talks for $1B in funding at an $11B valuation for general-purpose robot AI. (Bloomberg Technology)
Regulation & Policy
EU regulators proposed banning AI nudification apps — including Grok — from generating unauthorized sexually explicit content. (Social Media Today)
Anthropic won a preliminary injunction blocking the Pentagon’s supply chain risk ban on its AI tools. (The Verge)
Anthropic is separately in active discussions with the Pentagon over the future of AI in autonomous weapons. (Bloomberg Technology)
David Sacks stepped down as Trump’s AI czar but stays on the White House Technology committee. (TechCrunch AI)
Beijing launched a review of Meta’s Manus acquisition, raising concerns for Chinese founders using Singapore entities to sidestep scrutiny. (CNBC)
Krach Institute CEO Michelle Giuda argued the US should rely on allied nations to back the American AI tech stack against China. (Bloomberg Technology)
Meta’s Oversight Board warned that Community Notes are too slow and understaffed to counter AI disinformation at scale. (The Decoder)
Business & Funding
SoftBank secured a $40B unsecured 12-month loan from JPMorgan and Goldman Sachs to fund its OpenAI stake — widely read as pointing to a 2026 IPO. (TechCrunch AI)
Anthropic is considering an IPO as soon as October, putting it in direct competition with OpenAI for public market timing. (Bloomberg Technology)
OpenAI and Anthropic’s revenue figures are hard to compare ahead of their IPOs due to different cloud partnership accounting methods. (The Decoder)
Anthropic’s Claude paid subscriptions more than doubled this year, with consumer user estimates now at 18–30 million. (TechCrunch AI)
Eli Lilly signed a $2.75B deal with Insilico Medicine for AI-discovered drugs, with $115M upfront. (CNBC)
Shield AI raised $2B and plans to acquire simulation software maker Aechelon for autonomous military drones, with Ukraine among its government clients. (NYT)
Elon Musk’s last remaining xAI co-founder reportedly departed, leaving just two of the original 11. (TechCrunch AI)
Moonshot AI is exploring a Hong Kong IPO as AI stocks continue to climb. (Bloomberg Technology)
Pony AI reported its first profitable quarter and plans to expand robotaxis to 20 cities. (Bloomberg Technology)
An ex-Coatue trader launched a hedge fund powered by Eve, an AI system handling day-to-day operations. (Bloomberg Technology)
BNP Paribas plans to deploy an AI companion for all staff to cut costs bank-wide. (Bloomberg Technology)
Anthropic reportedly views itself as the antidote to OpenAI’s approach to AI safety, rooted in early internal conflicts at OpenAI. (The Decoder)
Society & Culture
Anthropic’s second Economic Index found that longer Claude usage improves user outcomes — but risks widening economic inequality between skilled and unskilled users. (The Decoder)
Meta suffered two courtroom losses over allegations it knew about product harms, raising fresh concerns for AI research and consumer safety. (CNBC)
CBS News poll found rising AI awareness is directly linked to public expectations of fewer available jobs. (CBS News)
The CEOs of Coca-Cola and Walmart both cited incoming AI waves as a factor in their departures. (CNBC)
Bloomberg’s senior editor compared the AI boom to the 1990s internet surge, noting key elements still missing from the current wave. (Bloomberg Technology)
Goldman’s new adviser Rishi Sunak urged small UK firms to adopt AI at a Birmingham event, warning of the cost of ignoring it. (Bloomberg Technology)
Google AI Blog published a conversation between James Manyika and LL COOL J on AI and creativity. (Google AI Blog)