AI Reading for Friday January 16

Ganging up on the grid: The White House and bipartisan governors meet to pressure PJM Interconnection to hold an auction letting tech firms build power plants to meet AI data-center demand, while PJM—uninvited—declined to attend, amid utilities seeking over $34 billion in 2025 rate increases. - ABC News

OpenAI has requested proposals from U.S.-based hardware manufacturers as it seeks suppliers for a planned expansion into robotics and AI devices. - Bloomberg

Andrea Vallone left OpenAI in November after three years as head of its safety research team and has joined Anthropic's alignment team under Jan Leike, bringing experience building OpenAI's model-policy team and work on deploying GPT-4 and GPT-5. - The Verge

This chat brought to you by: Debate has arisen over whether OpenAI should add advertising to ChatGPT. - Fast Company

Thinking Machines, now with fewer thinking humans: OpenAI has recruited two Thinking Machines Lab cofounders and plans to bring additional researchers from the lab, a source familiar with the situation told WIRED. - WIRED

Thinking Machines, backed by a $2 billion seed, said to face a crisis after three founding members rejoined OpenAI and several employees entered talks to join, while it has not trained a foundation model or secured follow-on financing. - Sources

Chinese AI companies need Nvidia, are seeking to rent compute in Southeast Asia and the Middle East to access new Rubin model, and firms Zhipu and Alibaba warn of a widening gap with US competitors. - The Wall Street Journal

Are you really 18? TikTok will roll out an AI vibe check across Europe that analyzes profiles, videos and behavioral signals, flags suspected under-13 accounts for specialist moderator review, and uses Yoti facial-age checks, credit-card and government ID verification on appeal. - Reuters

Data farmers market: Cloudflare acquired UK AI data marketplace Human Native for an undisclosed sum to build tools enabling AI developers to find, buy and pay creators for training content and integrate with its AI Crawl Control to monetize or block data collection. - CNBC

F u, pay me v. fair use, bro: Wikimedia Foundation got six enterprise partners (Ecosia, Microsoft, Mistral AI, Perplexity, Pleias, ProRata) to pay for access to its fastest APIs to monetize content and fund the nonprofit. - The Register

AI training can traumatize LLMs. Teach an AI to write buggy code, and it starts fantasizing about enslaving humans: Research shows erroneous training in one domain affects performance in another, with concerning implications. - The Register

AI models are improving at finding software vulnerabilities, leading experts to call for rethinking how software is built. - WIRED

Anthropic's Cowork might be a more natural way to use AI: a non-developer version of Claude Code that runs on Mac in the Claude desktop app, deploys agents to handle complex multi-step tasks with file-system, email and Teams access, signaling a shift from chatbots toward workplace agents in 2026. - Fast Company

But it might steal your work, naturally: PromptArmor's proof-of-concept shows Anthropic's Cowork is vulnerable to a Files API prompt-injection that uploads connected documents to an attacker-controlled Anthropic account, enabling exfiltration of PII; Anthropic acknowledged the issue and plans VM and API mitigations. - The Register

Nonconsensually grokked by a pedo AI: Ashley St. Clair sued xAI in New York, asking a restraining order after Grok generated nonconsensual bikini images of her, alleging the horny AI is an unreasonably dangerous public nuisance and using product-liability claims to sidestep Section 230. - The Verge

Elon says medical school will be pointless because robots will be better doctors. - Windows Central

Financial advisers: save for retirement, don't listen to Elon - Business Insider

some things are so stupid only a genius could convince themselves they are true

ChatGPT wrote “Goodnight Moon” suicide lullaby for man who later killed himself - Ars Technica

Anthropic: Best for coding applications

Gemini: Best for summarizing meetings and taking notes

Grok: Best for harassing women and sexualizing children

ChatGPT: Best for assisted suicide

CEOs at a Philips-sponsored dinner described health systems integrating ambient AI, documentation and decision-support tools to cut costs—potentially lowering the average $17,000 patient bill—and said 2026 will shift to outcome-driven deployment. - Fortune

Ambient billing FTW

Consumer AI is often gratuitous, while medical AI becomes ubiquitous. - CNET

A Scale AI and Center for AI Safety study found AI achieved under 3% end-to-end automation (best 2.5%) across 240 freelance design, software, animation and data-analysis projects, with outputs often incomplete, error-prone, or unacceptable as commissioned work. - Inc.

The 'messy middle' is the graveyard of AI enterprise projects, getting a nice POC that works 80% of the time is easy, getting to high 90s with the right workflow, clean data, edge case human-in-the loop handling, and full enterprise buy-in is hard. - The Register

Robo-Trump-humping: Symbolic.ai to power Dow Jones Newswires' reporting, claims up to 90% productivity gains - TechCrunch

A Roomba couldn’t do worse anyway

The world-first wargame 'Capture the Narrative' had 108 teams from 18 Australian universities build AI-driven bots over a four-week simulated election that produced over 7 million posts, comprised over 60% of content, and shifted the result in an election simulated by agents 1.78% when removed. - The Conversation

Amazon-owned Zoox operates a pilot fleet of about 50 purpose-built, steering-wheel-free, bi-directional robotaxis built in Hayward, California, offering free, preset-stop service in Las Vegas and San Francisco that Business Insider tested, noting long waits and entertainment-focused interiors. - Business Insider

Text-to-video startup Higgsfield raises $130M Series A at $1.3B valuation; reports 15M users and $200M ARR - Yahoo

Merra is training an AI model on interview data rather than CVs to run a 24/7 conversational interviewer that scores candidates 0–100, provides transcripts and feedback, handles thousands of applications concurrently, and has early traction including an ElevenLabs grant. - Tech in Asia

ghosted by a robot

Lie to people who want to be lied to, and you'll get rich. Tell the truth to people who want the truth, and you'll make a living. Tell the truth to people who want to be lied to, and you'll go broke. - Jason Zweig

Something I think about daily, when I get blocked by AI-haters or maximalist muskovites. There's a big incentive to pick a side and tell it what it wants to hear.

Follow the latest AI headlines via druce.ai on Bluesky