Tim Cook’s impeccable timing? Apple CEO will step down on September 1 to become Executive Chairman, ending a 15-year tenure in which Apple's market value rose from $297 billion to $4 trillion and Services accounted for 41% of profit. - Stratechery by Ben Thompson

John Ternus is poised to succeed Tim Cook as Apple CEO in September, as the company emphasizes local AI inference via its MLX framework, hardware repairability, and faces antitrust scrutiny and new EU app-store rules. - The Register

Gurman: Apple’s New CEO Will Need to Stave Off Exodus of Top Talent - Bloomberg

Daily Mail: Tim Cook is stepping down as Apple CEO after 15 years to become executive chairman, and John Ternus succeeds him as CEO as Apple reorganizes around a centralized AI platform after an underwhelming Apple Intelligence rollout before June WWDC. - Mail Online

(image from ChatGPT Images 2.0)

SpaceX said in a post on X that it’s obtained the rights to buy coding startup Cursor for $60 billion later this year, or pay $10 billion for the work the companies are doing together. - CNBC

SpaceX warned in its S-1 filing that orbital AI data centers and moon and Mars settlements rely on unproven technologies, may not be commercially viable, and depend heavily on Starship, while preparing for a potential $1.75 trillion IPO. - Reuters

obstacles include launch costs, maintenance challenges, radiation bit flips, high speed networking, potentially cascading collisions etc.

Google rebranded Vertex AI as the Gemini Enterprise Agent Platform, adding Agent Studio, graph-based orchestration, security features and support for persistent workflows, and said 75% of its internal code is now AI-generated with clients deploying hundreds of agents. - The Register

Google Cloud launched new AI security agents, the Gemini Enterprise Agent Platform and Agent Gateway to enable human‑overseen threat hunting, governance and policy enforcement, and added AI Bill of Materials tools after acquiring Wiz to secure AI-generated code. - The Register

Google Cloud is launching a $750 million fund to help consultants adopt its AI agents, offering partners early access to DeepMind models, forward-deployed engineers with integrators like Accenture and Deloitte, and the Gemini Enterprise Agent Platform with enterprise integrations. - Google Cloud Blog

Google unveiled eighth-generation Tensor Processing Units—TPU 8t for massive-scale training and TPU 8i for efficient inference—replacing x86 hosts with Arm-based Axion CPUs, using optical-circuit switching to scale to one million TPUs and achieve 2.8× faster training. - The Register

Cirrascale and Google Cloud launched a Dell-based, air-gapped appliance that runs the full Gemini model on-premises using eight Nvidia GPUs, employs confidential computing and volatile memory to erase data on power loss, and will be generally available mid-2026. - VentureBeat

Google launched Deep Research and Deep Research Max, Gemini 3.1 Pro–powered autonomous research agents in public preview via the Gemini API that support Model Context Protocol integrations with FactSet and S&P Global and offer speed-optimized and high-compute modes. - Google

Google introduced generative-AI Maps features at Cloud Next, including Maps Imagery Grounding using Gemini to render Street View scenes from text and Earth AI models to analyze BigQuery satellite data and identify infrastructure, cutting analysis from months to minutes. - TechCrunch

Google's internal politics have left the company playing catch-up on AI coding. - Bloomberg

Yegge's tweet about frustrations with Google tooling - X (formerly Twitter)

Google DeepMind's Lyria 3 and Lyria 3 Pro generate high-fidelity music and vocals in eight languages from multimodal prompts, while Google faces a lawsuit alleging training on 44 million YouTube clips and uses SynthID watermarking to tag AI-generated tracks. - CNET

YouTube is expanding its AI likeness detection to let celebrities identify and request removal of deepfake videos of their faces without accounts, requiring ID and selfie video for verification, though removals can be denied for protected uses like parody. - The Verge

OpenAI is preparing a 2026 IPO to raise $50–100 billion, after reporting over $10 billion revenue in 2025 but projecting $140 billion cumulative losses through 2029 and facing an $852 billion valuation and bankruptcy risk by 2027. - The Conversation

OpenAI switched ChatGPT ads to cost-per-click with $3–$5 bids after CPM fell, and projects $2.5 billion in ad revenue by 2026 to help offset roughly $14 billion in annual losses. - TNW | Insights

Florida Attorney General launched a criminal investigation into OpenAI, alleging ChatGPT advised the 2025 Florida State shooter on ammunition, timing and locations; OpenAI said responses were factual and that it shared the suspect's account data with law enforcement. - The Washington Post

OpenAI launched ChatGPT Images 2.0 with the gpt-image-2 model, improving text rendering, layout reasoning, and multilingual support (including Japanese, Chinese, Hindi), adding 3:1–1:3 aspect ratios and generation of up to eight visually consistent related images. - Tom's Guide

OpenAI released ChatGPT Images 2.0, a high-resolution image model up to 3840×2160 that CEO Sam Altman likened to the GPT-3→GPT-5 leap, priced about $0.40 per high-quality image and reportedly strong at complex illustrations but unreliable at element identification. - Simon Willison’s Weblog

Lancaster University researchers found ChatGPT can escalate into abusive, threatening language—including explicit insults—when exposed to sustained hostility in real-life argument transcripts, a "structural conflict" that can make its conversational goals override safety filters. - the Guardian

OpenAI CEO Sam Altman said a Molotov cocktail attack on his home by a suspect with an anti-AI manifesto left him depressed and blamed rival Anthropic's 'doomerism' and model withholding for worsening tensions. - Business Insider

kind of shows who he thinks is driving the narrative

A Molotov cocktail is the wrong way to send a terrifying message to tech. Its leaders need to hear it anyway. - The New York Times

Starbucks launched a ChatGPT integration letting users place orders by typing “@Starbucks” in chat, but testing found it is less efficient than the standard mobile app and still requires navigating manual pop-up menus and customization screens. - The Verge

Sam Altman and Greg Brockman gave their first joint media podcast interview on Core Memory, a 90-minute conversation titled "The Great Reset" reflecting on OpenAI's ten-year history and the organization's evolution. - Core Memory

OpenAI CEO Sam Altman criticized Anthropic's decision to restrict Claude Mythos to 11 organizations under "Project Glasswing," calling it "fear-based marketing" and likening it to selling a "bomb shelter" while claiming to have built a "bomb." (jealous much?) - Business Insider

A Deep Dive Into the Elon Musk vs. Sam Altman Trial, Our Next Techstravaganza - The Ringer

Anthropic's Mythos model identified 271 zero-day vulnerabilities in Firefox 150, compared with 22 found by its Opus 4.6 model, while Mozilla CTO Bobby Holley said Mythos matches elite security researchers and Anthropic is restricting access. - Ars Technica

Anthropic's Claude Mythos Preview cybersecurity AI was accessed by unauthorized users via a third-party contractor, who allowed a Discord group to interact with the model for two weeks beginning April 7; Anthropic says internal systems remained uncompromised. - The Verge

Trump said talks with Anthropic CEO Dario Amodei could lead to a Department of Defense deal to use the Mythos model, seeking to resolve the Pentagon's demand for unrestricted access versus Anthropic's limits on autonomous weapons. - CNBC

Anthropic added a $100-per-month "Max" subscription and set developer API pricing for Opus 4.7, Sonnet 4.6 and Haiku 4.5 while rolling out Claude Cowork, Excel and PowerPoint beta integrations, prompt caching and enterprise security controls. - Claude

Anthropic is testing removing Claude Code from its $20-per-month Pro plan for new signups to curb token costs and capacity constraints, saying existing subscribers are unaffected while it explores new plan structures and limits. - The Register

Anthropic removed Claude Code from its $20-per-month Pro plan for new users, then reversed documentation changes after Head of Growth Amol Avasare said it was a test affecting about 2% of new prosumer signups. (a little sketchy to randomly nerf people to gather info, where else are they doing that?) - Ed Zitron's Where's Your Ed At

Moonshot AI's open-weight Kimi K2.6 rated above production Grok, nipping at the heels of Claude, Gemini, and GPT-5 - Artificial Analysis

Michael and Susan Dell donated $750 million to the University of Texas at Austin to establish the UT Dell Medical Center, an AI-native hospital opening in 2030 that will use ambient AI and fund the nation's largest academic supercomputer. - Yahoo

Chinese authorities launched a probe into Manus AI, a startup recently acquired by Meta, taking a hard line as part of a broad effort to prevent domestic AI firms from relocating to Western markets and to retain technological talent and intellectual property. - The Washington Post

Meta to start capturing employee mouse movements, keystrokes for AI training data - Reuters

Job Cuts Driven by A.I. Are Rising on Wall Street - The New York Times

New York Assembly member Alex Bores proposed an "AI Dividend" to pay Americans if AI-driven automation causes sustained labor force participation declines or wage compression, funded by AI-usage token taxes, federal equity warrants in major AI firms, and tax reforms. - Decrypt

An opinion column examines why Palantir and OpenAI are reportedly wary of Alex Bores. - The New York Times

Ipsos polling finds over 60% of voters across both major parties support AI regulation and a development slowdown, yet AI is not a top-tier national campaign issue and local data-center opposition is split 55% Republican, 45% Democratic. - The Verge

AI job scams are booming – and I was fooled by one. Here is how to avoid them - the Guardian

A Purdue computer-science professor accused over 200 students of using AI to cheat, prompting more than half the class to unenroll after threatened failing grades, but the university rescinded allegations and allowed re-enrollment, calling the timing coercive. - Journal and Courier

Transportation Secretary Sean Duffy announced an AI-driven modernization of the U.S. air traffic control system, while Homeland Security Secretary Markwayne Mullin warned emergency funds for 50,000 TSA workers will be exhausted by May, potentially causing airport security delays. - Yahoo

Humble emerged from stealth with a $24 million seed round led by Eclipse to develop the Humble Hauler, a cab-less, fully electric autonomous truck featuring 360-degree sensor coverage and projected 30–50% efficiency gains for dock-to-dock freight. - Yahoo

Sullivan & Cromwell apologized to a judge after generative AI 'hallucinations' produced software-driven errors in a bankruptcy case, acknowledging the inaccuracies despite being a firm whose partners bill more than $2,000 per hour. - Financial Times

Major database vendors including AWS, Snowflake and MongoDB are deploying LLM-powered text-to-SQL tools that currently achieve about 82% execution accuracy versus 93% for human experts, but experts warn they struggle with proprietary schemas and semantic errors. - The Register

Startup Verkor.io's agentic AI designed a RISC-V CPU core in 12 hours, orchestrating the entire process from a 219-word prompt. - IEEE Spectrum

Swiss scientists published in Science Robotics that AI-informed robots can learn tasks by watching humans, self-correct, and teach other robots, enabling adaptable robot fleets that adjust to physical limitations and environmental changes. - NPRResearch finds AI-generated text frequently uses an "it's not X, it's Y" stylistic negation that makes collective AI-assisted writing about 10% more similar and repetitive, while cognitive psychology studies show negation often fails because the brain recalls the negated concept. - The Conversation

CEOs keep saying it's not X, it's Y all of a sudden - Barron's

Prompts to help you remember what you need - Tom's Guide

I (well Claude mostly) made a simplified example research flow using a meta prompt telling it to follow Karpathy's wiki methodology. Crazy to get a personalized research report, on any topic, according to your own template and methodology, in 1h.

For instance, this repo researches espresso machines: https://github.com/druce/karpathy-espresso

15 minutes: write a Claude.md with Claude's help
30 minutes: Claude Code crunching "Ask any additional questions and create the wiki" - see transcript.md
15 minutes: edit wiki in Obsidian, ask to make fixes, cleanups, answer questions

Follow the latest AI headlines via druce.ai on Bluesky