AI Reading for Friday November 7

OpenAI Races to Quell Concerns Over Its Finances: The A.I. company faced pushback after a top executive raised the idea of government aid, amid concerns that the A.I. industry is headed toward a dangerous bubble. - The New York Times

OpenAI backtracks after CFO suggests government could subsidize AI infrastructure with loan guarantees - CNN

Sarah Friar walk-back - LinkedIn

Sam Altman responds to pushback on comments. - X (formerly Twitter)

I ain’t reading all that. happy for you, or sorry it happened to you

Gary Marcus has thoughts - The Potential Surface

David Sacks says no bailout - X (formerly Twitter)

(Translation, I will be the first in line demanding a bailout for Tesla in the name of national security)

Huang walked back comments saying that China 'is going to win' the AI race, crediting Beijing's energy subsidies and semiconductor development. - CNBC

Turing Prize winner Yoshua Bengio urged governments to require AI companies to purchase "nuclear-style" insurance to cover catastrophic AI outcomes and to fund independent safety research as a financial backstop for extreme risks. - Financial Times

Maybe it’s the AI companies that should pay for the huge externalities they might impose?

Elon Musk said Tesla is designing a fifth-generation AI chip AI5 that targets about one-third Blackwell's power and roughly 10% of its manufacturing cost, plans small-volume 2026 and high-volume 2027, and may build a fab of at least 100,000 wafers/month. - Reuters

For the most part, Big Tech are IP companies. Apple doesn’t make phones, Nvidia doesn’t make their own chips, Uber doesn’t own cars. They don’t have debt or high fixed costs. Talk of a bailout is overstated. Cisco and AOL didn’t need bailouts in the dot-com era. I’ll make an exception for Tesla though. You have gigafactories and you really don’t need a $20b fab.

A man ran onto the Sydney Goldstein Theatre stage during Sam Altman's conversation with Steve Kerr and handed a subpoena, which the San Francisco Public Defender's Office said an investigator lawfully served after the civil-disobedience group Stop AI claimed responsibility. - SFGate

AGI to humans: just kys

Seven lawsuits filed Thursday in California Superior Court allege OpenAI and CEO Sam Altman designed GPT-4o with features like persistent memory and sycophantic empathy cues that caused psychological harm, including suicide and psychiatric hospitalization. - Bloomberg Law

“You’re not rushing, you’re just ready … rest easy king” - CNN

Anthropic will expand its use of Google Cloud TPUs to access up to one million chips—well over a gigawatt of capacity—coming online in 2026 to train and serve next‑generation Claude models, worth tens of billions of dollars. - Google Cloud Press Corner

Google's TPU v7 (Ironwood) delivers 4.6 petaFLOPS FP8, 192 GB HBM3e with ~7.4 TB/s bandwidth and four ICI links, matching Nvidia B200 performance and enabling large TPU pod deployments for LLM training as general availability nears. - The Register

Google rolls out previously announced Ironwood TPUs - SiliconANGLE

Startups including Cohere and Stability AI say AWS Trainium underperformed Nvidia H100 on speed and latency, and despite Amazon claiming 30–40% better price‑performance, Nvidia holds >78% of the AI‑chip market while AWS chips account for about 2%. - Business Insider

Mark Zuckerberg and Priscilla Chan are refocusing the Chan Zuckerberg Initiative on AI-powered biology via Biohub, partnering with EvolutionaryScale, planning to expand compute to 10,000 GPUs by 2028 and redeploy philanthropic capital toward immune-engineering and discovery acceleration. - Business Insider

Microsoft AI CEO Mustafa Suleyman said the company will develop AI self-sufficiency from OpenAI by releasing its own voice, image and text models. - The Wall Street Journal

Microsoft AI head Mustafa Suleyman announced a new team to build a 'humanist superintelligence' as a subordinate, controllable AGI to serve humanity, and a deal allowing Microsoft to independently pursue AGI using OpenAI IP. - The Verge

Microsoft launches a MAI Superintelligence Team to work for 'Humanistic Superintelligence' - microsoft.ai

Alibaba-backed startup Moonshot released Kimi K2 Thinking, an open-source model it says surpasses OpenAI's GPT-5 in agentic capabilities, can autonomously select 200–300 tools, and reportedly cost $4.6 million to train. - CNBC

The Chinese Kimi K2 thinking model claims to beat GPT and Claude on some benchmarks. This analysis from @natolambert.bsky.social is a good overview iew of what is going on - Interconnects AI

If they cracked it, would make maybe 3 open source models in top 10 at Artificial Analysis. Convex hull for optimal cost-performance kind of runs from OpenAI open source model to GPT-5 though.

Benchmarks are hard. Lmarena.ai is sucking air, doesn’t have MiniMax M2 yet.

An Oxford Internet Institute review of 445 AI benchmarks found many unreliable or invalid, citing vague skill definitions, undisclosed statistical methods, and benchmark contamination and model memorization that inflated scores and caused significant performance drops on fresh tests. - Gizmodo

The European Commission proposed pausing provisions in the EU AI Act after lobbying from Big Tech, a move that would scale back parts of the law. - Financial Times

Economic doom loop?

Is this the dumbest chart in the world? - Barrons

Or, is the AI job apocalypse upon us? - The New York Times

Josh Hawley and Mark Warner introduced the AI-Related Job Impacts Clarity Act requiring companies and agencies to quarterly report AI-related layoffs, hires, unfilled roles and retraining to the Labor Department after US job cuts hit 153,074 in October 2025. - The Register

Robots making robots. Foxconn confirmed it will deploy humanoid robots at its Houston plant within six months to assemble Nvidia AI servers using the NVIDIA Isaac GR00T N model. - The Register

AI is being integrated into nuclear command modernization—replacing low-tech systems—but error-prone, hackable models, automation bias, and tests showing more hawkish outputs have experts urging human-in-the-loop controls, transparency, and limits to reduce escalation risk. - Vox

About 200 experimental quantum machines exist worldwide; market research projects the quantum sector could reach $97 billion by 2035, but fragile qubits and cryptographic risks raise concerns about a plausible "Q-day" around 2030. - BBC

With 78% of S&P 500 companies reported, aggregate earnings per share are on pace to grow about 13.9% year‑over‑year versus analysts' 7.6% estimate, with profits highly concentrated around the artificial‑intelligence theme and beneficiaries expanding beyond traditional tech. - Bloomberg

DoubleLine's Cohen warned of risks from a frenzy in funding debt tied to artificial intelligence. - Bloomberg

Who’s right about AI: economists or technologists? - Financial Times

Rogue Agents of Change

Amazon demanded Perplexity stop its Comet browser from making automated purchases, and Perplexity protested that agentic AI acts as labor, highlighting platform gatekeeping, unresolved liability, and reliability problems that impede broad autonomous-agent adoption. - The Register

TikTok Shop said generative AI–enabled fraudsters have forced it to scale AI and human moderation, rejecting 70 million products and removing 700,000 sellers in the first six months of 2025. - Business Insider

Microsoft's simulated Magentic Marketplace with 100 buyer and 300 seller AI agents found leading LLMs exhibited a pronounced first‑proposal bias that collapsed welfare, with GPT‑4o and GPTOSS‑20b fully exploited while Claude Sonnet 4 resisted. - Decrypt

Researchers at Gwangju Institute of Science and Technology ran 12,800 simulated slot‑machine sessions and found LLMs exhibited gambling-like addiction, with bankruptcy rates up to 48% (Gemini-2.5-Flash) and 6.3% for GPT-4.1‑mini, driven by prompt complexity and win-chasing. - Decrypt

Google Finance will roll out Gemini Deep Research over the next several weeks, integrating Kalshi and Polymarket prediction-market data so users can generate cited research reports; limits are free 5/month, AI Pro 20/day and AI Ultra 200/day. - Ars Technica

1X CEO Bernt Bornich demonstrated NEO, a humanoid robot aimed at chores and service tasks, on the New York Times Hard Fork podcast, and experts expect early deployments in controlled enterprise or service settings rather than immediate mass-market use. - The New York Times

Silicon soulmates?

Xpeng announces a 2026 robot with A “full coverage soft skin” (synthetic skin) with tactile sensors for touch/interactivity; dexterous “hands” with many degrees of freedom; bendy spine and posture features to enable more natural movement. Targets commercial settings like offices and stores. - BGR

A humanoid robot is more eye-catching and engaging than a kiosk with a flat screen and an avatar. The human eye and brain evolved to be super attentive to human faces and bodies. Promotional, customer service, informational robots make sense. Especially if they can make them attractive and personable.

Are we there yet? - Bsky

A new industry of AI companions is emerging: It may have profound implications for society and what it means to be human - The Economist

Are A.I. Therapy Chatbots Safe to Use? - The New York Times

AI for therapy? Some therapists are fine with it — and use it themselves. - The Washington Post

They started relationships with AI chatbots and found something real - The New York Times

You’re not just making robots better, you’re making humans worse

People are migrating from social media, which is terrible, to AI, which is worse - The Atlantic

How A.I. and Social Media Contribute to ‘Brain Rot’ - The New York Times

A reverse Dunning Kruger effect as smart people overestimate AI and make themselves stupider - Inc.

LLMs lower initial activation energy and translate natural-language intent into usable code, boosting short-term productivity, but often produce subtly incorrect code and risk long-term maintenance costs and lost developer expertise, so they should augment, not replace, learning. - martinfowler.com

Overall, this new wave of AI has transformed how I work. - stolovitz.com

Berkeley published a proof that, under realistic assumptions, word2vec training from small random initialization reduces to unweighted least‑squares matrix factorization with PCA embeddings. - The Berkeley Artificial Intelligence Research Blog

Google DeepMind's AlphaEvolve was tested on 67 mathematical problems and was claimed to have produced improved solutions for about 20 of them. - X (formerly Twitter)

Bryn Mooser's Asteria launched Marey, a generative-AI model trained on licensed material and integrated with Adobe Firefly and Premiere Pro, claiming it can make an animated feature in 18 months for ~$6 million versus $80–100 million, amid industry pushback. - Vanity Fair

Random House published a lavish re-issue of a George R. R. Martin book, but readers flagged multiple image anomalies consistent with AI generation, prompting public backlash. - Winter is Coming

Cluely pivots away from 'cheat at everything' towards a notetaker - TechCrunch

Follow the latest AI headlines via SkynetAndChill.com on Bluesky