AI Reading for Thursday December 18

Pre-profit, post-reality valuation? OpenAI has held preliminary discussions with investors about raising up to $100 billion, at an approximate $750 billion valuation, according to The Information. - Bloomberg

OpenAI signed a $38 billion, seven-year deal with Amazon Web Services for access to "hundreds of thousands" of Nvidia GPUs to train its AI models, with all capacity targeted for deployment by the end of 2026 and expandable into 2027. - The Verge

OpenAI is accepting app submissions for ChatGPT via its Developer Platform, releasing a beta Apps SDK and in-chat app directory, while requiring policy compliance and MCP connectivity and initially limiting monetization to linking out for physical goods. - OpenAI

OpenAI opened ChatGPT to third-party apps across its ~800 million weekly users, enabling Adobe, Target, Canva and Figma integrations via the Model Context Protocol—developed by Anthropic and open-sourced in 2024—with Canva reporting 2.6M users and 11M designs. - Fast Company

OpenAI launched a ChatGPT App Directory and SDK letting developers build interactive apps inside ChatGPT's UI, rebranded connectors as apps, allowed apps to access Memory if enabled, and said it may use opted-in user data to train models. - The Verge

Internal documents and executive interviews show Microsoft is executing a sweeping AI-driven reorganization under CEO Satya Nadella, consolidating AI leadership, creating a weekly AI accelerator, accelerating model release cadence from about six months to six weeks, and promoting Judson Althoff. - Business Insider

Prime Directive: AGI: Amazon announced that Rohit Prasad will leave next year as Peter DeSantis will lead a new division overseeing advanced AI models, chip development and quantum computing, while superstar academic Pieter Abbeel will head the frontier model research team. - The Verge

How Trump is helping China realize its dream of AI dominance - The New York Times

Does China have a robot bubble? - The New York Times

When Tiger Global was the master of the bubbleverse - Rest of World

Yann LeCun, Meta's chief AI skeptic and Turing Award winner, is launching an AI start-up to prove everyone else wrong, targeting a €3bn valuation and is in early talks to raise about €500mn ahead of a planned January launch. - Financial Times

what is he done lately though, last generation of Meta models bombed spectacularly, this is a higher valuation than Dr. Fei Fei who is doing something similar.

Elon Musk’s ‘burn the boatloads of cash’ speech, told xAI staff it must survive the next two to three years to prevail, said AGI could arrive by 2026, and that xAI operates ~200,000 GPUs targeting 1 million and can access $20–30 billion. - Business Insider

This long form video about people trapped inside burning Teslas is phenomenal and awful. - Bloomberg

Your power bill is going up so kids can Ghiblify their avatars: Three Senate Democrats are investigating tech firms, saying data centers' concentrated electricity demand for AI workloads is forcing utilities to spend billions on grid upgrades and may lead to higher consumer rates and increased costs for data-center operators. - The New York Times

Prompt engineering pathogens: UK AI Security Institute's Frontier AI Trends Report found generative models improved at biological tasks: LLM assistance made non-experts almost five times likelier to produce viral-recovery protocols, and self-replication success rose from under 5% to over 60% in two years. - Transformer News

The NY congressional candidate the AI industry is determined to stop. - Bloomberg

Finding Nemotron: Nvidia ships a 30-billion-parameter hybrid model you can probably run on a beefy MacBook, that activates 3B parameters, provides a 1 million-token context window, ships open weights, and claims fourfold throughput and about 60% fewer reasoning tokens versus its predecessor. - Geeky Gadgets

China ASML competitor said to advance, although you have to ask for which chips (none) on what node, at what volume (none) - The Chosun Daily

Better living through AI chemistry: Startups like Lila Sciences and researchers including Rafael Gómez-Bombarelli say AI-driven materials discovery is attracting new funding and automating experiment design, but verification, synthesis, and hype raise scalability, reproducibility and commercialization risks. - MIT Technology Review

Swarmageddon: Researchers at MIT, Google and collaborators ran 180 configurations and derived three empirical scaling laws showing multi-agent swarms often worsen sequential reasoning (39–70% worse), amplify errors 17.2× versus centralized 4.4×, and hit a performance plateau. - Hacker Noon

Claude testing agents using different modes - Hacker Noon

e.g. research, analyze, write/create, build, do agentic stuff

How Companies Are Using AI to Squeeze More From Your Wallet - Barron's

We Let AI Run Our Office Vending Machine. It Kept Losing Money. Not sure if this means it's hard for AI to optimize effectively or you're just doing it wrong. - The Wall Street Journal

Not yet smart enough to run a banana stand?

Is SEO over, or is AI search optimization simply a complement to traditional SEO? - The Wall Street Journal

Hackers breached Andreessen Horowitz–backed startup Doublespeed's AI avatar backend, accessing its phone-farm control panel and extracting over 400 TikTok accounts—about half actively promoting products—that were used to run ads and other deceptive, platform- and FTC-violating campaigns. - Futurism

Large language models are generating fabricated academic citations and nonexistent papers that Andrew Heiss found in student work and published articles, and research librarians report spending up to 15% of their time tracking nonexistent records. - Rolling Stone

A filmmaker created a 'Sam Bot' deepfake of OpenAI CEO Sam Altman as an interview stand-in, released an exclusive trailer for the documentary 'Deepfaking Sam Altman', and said the film will debut in January. - WIRED

Junior lawyers' pay has continued to rise even as AI automates tasks, with firms saying automation is freeing their time for more valuable work. - Financial Times

Don Rickles for your landing page: Roast My Web uses GPT-Vision and other models to audit websites, producing category scores (example: aesthetics 87/100, conversion 59/100), flags competitor issues (site ~40% slower), and the author reports conversion gains after implementing its ROI-ranked fixes. - Hacker Noon

“Take a look at this page, folks. Go ahead. I’ll wait.

No really—I’ll wait, because it’s still loading, numbskull.

What is this, a website or a hostage situation?

I click once and it asks me to accept cookies, reject cookies, manage cookies—

pal, I just wanted information, not a long-term relationship with a bakery.

Look at this font. Who picked this font? This isn’t typography—it’s a cry for help. I’ve seen ransom notes with more brand consistency.

And the hero image—ohhh, the hero image. Nothing says “trust us with your data” like a blurry stock photo of people high-fiving in a conference room.

Scroll down—keep scrolling—no, keep going—

Congratulations, you’ve reached the value proposition. It’s vague, it’s circular, and it says absolutely nothing. I’ve heard better explanations from a kid explaining why the homework ate the dog.

And the call to action: “Get Started.” Started on what, genius? A journey? A lifestyle? A support ticket?

This page doesn’t load fast, doesn’t explain itself, and doesn’t shut up.

But other than that—beautiful job, sweetheart. Really.

Now do us all a favor and add a “Back” button—for your career.”

Follow the latest AI headlines via SkynetAndChill.com on Bluesky