Pass the plate: OpenAI reopens funding round again, says it's now closed at $122b at $852b valuation, also took money from qualified individual investors via banks. -OpenAI

Land-grab race to scale/shock and awe… or get plenty of acorns stowed away for winter?

How to train your copyright dragon: Penguin Random House sued OpenAI's Ireland-based European subsidiary in Munich, alleging ChatGPT generated text and images that unlawfully reproduced Ingo Siegner's Coconut the Little Dragon children's books. - the Guardian

The promptletariat: OpenAI's Project Stagecraft, run through data-labeling startup Handshake AI, engaged roughly 3,000–4,000 freelancers to create occupation-specific prompts to train ChatGPT, paid at least $50/hour, and leaked a 439-row occupations spreadsheet plus contractor emails. - Business Insider

Go figure: Figure CEO Brett Adcock ended collaboration with OpenAI after the company chose to develop humanoid robots in-house - Business Insider

OpenAI's updated ChatGPT product-recommendation feature repeatedly produced factual errors in tests, wrongly attributing and replacing WIRED's hands-on picks (e.g., substituting the TCL QM6K and citing untested AirPods Max 2), risking reader trust and publisher revenue. - WIRED

Shotgun! ChatGPT launches on Apple CarPlay for iOS 26.4 with voice-only interactions and no wake-word support - The Verge

OpenAI ships Claude Code plugin to let you call Codex from Claude Code - OpenAI Developer Community

Yo Dawg, I Heard You Like Coding Assistants, I put my coding assistant inside your coding assistant.

A Claude Code 2.1.88 release accidentally exposed 512,000+ TypeScript lines, revealing unreleased features like a Tamagotchi-style pet and a KAIROS agent; leaked files on GitHub have 50,000+ forks, and Anthropic says it patched the issue with no customer data exposed. - The Verge

Leaked Claude Code client source shows Anthropic collects extensive on-device user data and telemetry, syncs sessions to api.anthropic.com, retains data up to five years if opted into training (30 days default), and includes remote-update and feature-flag mechanisms. - The Register

Claude code leak reveals a 3-layer memory architecture using "context entropy" and a "Self-Healing Memory" system. KAIROS & autoDream allows the agent to perform "memory consolidation" while the user is idle, merging observations and removing logical contradictions autonomously. - VentureBeat

A source-mapped deep dive documents Claude Code's implementation across 1,900+ files and 519K+ lines, cataloging 53+ tools, 95+ slash commands and 50+ tool interactions while detailing agent loops, multi-agent orchestration and architecture. - Claude Code Unpacked

What the Claude Code leak means in context of regulated firms: don't blindly trust code assistants; verify outputs and include your toolchain in risk assessments/governance processes; harden CI with strong pre-deployment verifications. - Systima

AI will worsen security in the short term but outperform humans in the long term. - Stratechery by Ben Thompson

LiteLLM, heavy hearts: Mercor confirmed a security incident tied to a LiteLLM supply-chain compromise after extortion group Lapsus$ posted alleged Mercor data, and the company said it is investigating with third-party forensics while LiteLLM removed malicious code. - TechCrunch

AI-generated code outpacing security tools as ratio of lines of code deployed to security engineers widens - SiliconANGLE

A month of AI-discovered bugs - GitHub

A guide for senior developers outlines context-first prompt-engineering patterns, agentic CI/CD workflows and tests to accelerate boilerplate, documentation, and refactoring, warns that poor context produces technical debt, and recommends review for hallucinations, security, and performance. - Hacker Noon

Real transformers are harder to implement than Pytorch transformers: U.S. data-center construction for AI is being delayed because domestic manufacturers cannot produce enough transformers, switchgear and batteries, forcing reliance on imported Chinese electrical components. - Bloomberg

Nvidia is investing $2 billion in chipmaker Marvell to boost AI networking, while semiconductor companies are partnering on silicon photonics to accelerate data-center systems. - Financial Times

Johnny 5 is alive, need input!

Oracle began mass layoffs on 31 March 2026 that TD Cowen estimates cut 20,000–30,000 roles (about 18% of ~162,000) to fund an estimated $156 billion AI data‑centre buildout, with Oracle disclosing a $2.1 billion restructuring charge. - TNW | Business

Maybe NetSuite told them to do it. Not sure if AI autopilot is what you want in an accounting system. - TechRadar

A University of Cambridge preprint found that large AI datacenters produced heat islands with land-surface temperature rises of 0.3–9.1°C (average 1.5–2.4°C), with effects up to 10 km, ~1°C monthly rise to 4.5 km, potentially affecting up to 343 million people. - The Register

In India, MHA and the NCRB are developing CCTNS 2.0, an AI-driven policing platform for predictive, place- and person-based risk scoring, hotspot mapping, patrol optimization and surveillance (drones, cameras, facial recognition), raising civil-liberty concerns about bias and expanded surveillance. - MEDIANAMA

A Fairplay-led coalition of child-safety organizations sent a letter to Google CEO Sundar Pichai and YouTube CEO Neil Mohan demanding labeled AI content, banning AI-generated "Made for Kids" videos, blocking AI recommendations to under-18s, parental toggles, and halting AI investment. - Mashable

Magilla Gorilla will rot your mind too, just turn off the screen it’s all bad

Baidu luck next time: Robotaxi outage leaves Wuhan passengers stranded on highways. - WIRED

dashcam

In letters to Senator Ed Markey, seven autonomous-vehicle developers including Tesla, Zoox and Nuro disclosed remote-assistance programs, with Tesla saying operators sometimes directly control its vehicles, rarely and at speeds below 10 mph. ('directly' seems open to interpretation) - WIRED

Robotaxi firms refuse to disclose remote operator intervention rates to Senator Markey - TechCrunch

Study warns autonomous AI agents pose unpredictable risks outside intended parameters - The Wall Street Journal

Robot ratings: Wharton researchers propose ethics scorecard to audit AI systems before deployment - Knowledge at Wharton

Startup founder Claire Vo built nine OpenClaw AI agents to automate sales, scheduling, CRM, email and household tasks, replacing roughly 10 hours per week of paid work, after an initial calendar-wiping mishap led her to adopt incremental access controls. - Business Insider

Andrej Karpathy built 'Dobby,' an OpenClaw-based AI agent that scans his local network, reverse-engineers device APIs, and executes natural-language commands to control music, lighting and security, replacing multiple home apps though requiring technical skill to set up. - Business Insider

My precious! Dobby is a free model now, Dobby has no master.

Meet your new AI supervisor: Salesforce released an upgraded Slack centered on Slackbot that analyzes users' work and adds AI features—transcription, note-taking, research, integrations and reusable automations—and bundles Slack for all Salesforce customers with native CRM updates from channels. - Engadget

Meta formed MRS Research within Recommendation Systems during an October reorganization to concentrate elite AI researchers, led by Yang Song (joined November 2025) and hiring talent from TikTok, Amazon, OpenAI and Google to improve recommendations and ad relevance. - Business Insider

never give a sucker an even break

Companies push for wide AI adoption, embedding AI into performance reviews, mandating or tracking tool use and using incentives and training to drive adoption. - Business Insider

For legal reasons, this is all a joke, right? Copilot terms of service says "is for entertainment purposes only. It can make mistakes, and it may not work as intended. Don’t rely on Copilot for important advice. Use Copilot at your own risk." - Microsoft Copilot

just kidding, but sure, rebuild all of enterprise computing on it

Apple says nope to iOS vibecoding apps. - 9to5Mac

One possible concern might be app ecosystems that would let you run stuff outside App Store controls and monetization.

Professors at Columbia, Georgia Tech and other universities built faculty-infused AI apps, like Columbia's Caisey, used by thousands at Columbia and 15 other institutions, that prompt students with questions and counterarguments to scaffold reasoning across varied coursework. - The Washington Post

Super-guessing is all you need. Multimodal models trained to read images lean hard on text clues. A text model trained on a corpus with the actual images stripped out and just readable text performed as well as frontier models, among other interesting experiments. - arXiv

Caltech researchers claim radical compression breakthrough for high-fidelity AI models - The Wall Street Journal

AI seed valuations surge as investors write $10 million checks at $40-45 million post-money valuations - TechCrunch

A book excerpt describes how DeepMind's multi-year 'Project Mario' negotiations with Google after an August 2015 AI-safety meeting—including proposals for a 3-3-3 oversight board, spin-out options and a $1 billion Reid Hoffman pledge—shifted Demis Hassabis toward realism. - Colossus

DeepDelver whistleblower strikes again, says Delve copy-pasted open source project without license or attribution - The Potential Surface

Delve says it's all just a big misunderstanding - X (formerly Twitter)

Mobvoi launched the TicNote Pods, 4G-connected always-on AI note-taking earbuds that transcribe conversations, generate summaries and slides, offer ~5 hours recording (25 with case), 32 GB storage, support 100+ languages, and cost US$299 with subscriptions $119–$299/year for heavy use. - New Atlas

Perplexity AI CEO Aravind Srinivas faces backlash after saying workers should embrace being replaced by AI - New York Post

now get me a Diet Coke and Laura Loomer… - X (formerly Twitter)

Follow the latest AI headlines via druce.ai on Bluesky