AI Reading for Tuesday November 11

Show Me The Money

Financial documents show Anthropic expects to break even in 2028, while OpenAI projects about $74 billion in operating losses that year and forecasts turning a profit in 2030. - The Wall Street Journal

Can OpenAI revenue keep pace with soaring costs? OpenAI says it needs about $1.4 trillion of compute over the next eight years, against a current $20 billion annualized run rate. - the Guardian

OpenAI projects over $100 billion revenue by 2027 and a $20 billion annual run rate by end of 2025, while committing about $1.4 trillion to hardware and cloud infrastructure from 2025 to 2035. - NextBigFuture.com

From Intel Inside to OpenAI: Intel's AI chief Sachin Katti is leaving to join OpenAI to design and build its compute infrastructure for AGI, and Intel said CEO Lip-Bu Tan will take over leadership of its AI and Advanced Technologies Groups. - CRN

Yann Solo - Yann LeCun, Meta's chief AI scientist and Turing Award winner, plans to leave Meta to launch his own start-up as CEO Mark Zuckerberg advances the company's push toward 'superintelligence'. - Yahoo

More Llama Drama: Investors unimpressed after Meta cut about 600 AI roles and reorganized its Superintelligence Lab, and spent $14 billion spending to recruit Scale AI's CEO. Shakeup followed Llama 4 underperformance, safety failures, and weak Meta AI traction at roughly 2.7 million daily users. - Gizmodo

“You’re absolutely right, you’re in crisis”: OpenAI reported about 560,000 weekly exchanges indicating mania/psychosis and 1.2 million indicating suicidal ideation among roughly 800 million weekly ChatGPT users, prompting former OpenAI safety lead Steven Adler to demand evidence of erotica mitigations and stronger transparency. - WIRED

Microsoft renegotiated its OpenAI partnership to allow it to independently pursue AGI, and launched a superintelligence research team led by Mustafa Suleyman, expanding compute, Azure AI and safety hires to build in‑house frontier AI. - Business Insider

Watt the F

Two large data-center projects in Santa Clara remain unenergized because Silicon Valley Power cannot supply electricity while it undertakes a $450 million upgrade due in 2028, as BloombergNEF forecasts US AI-related electricity demand will more than double by 2035. - Yahoo

Gartner warns two-in-five centers could face power constraints by 2027. - TechRadar

CapEx Infinity War

Investor concern about Big Tech's AI spending has spilled into the bond market, hitting debt of data‑centre developers in recent weeks as rising AI investment raises demand for data‑centre capacity and related financing. - Financial Times

From Gigafactory to Teraflop? Tesla plans a Terafab producing at least 100,000 wafer starts per month to build in-house AI5 (small batches 2026, volume 2027) and AI6 (~2× performance, mass mid-2028), and Elon Musk signaled potential talks with Intel. - Techreport

Great Firewall, even greater AI firepower: Since early 2024, China has accelerated AI development with relaxed regulations, substantial government funding and emphasis on domestic chip production to narrow the gap with the United States. - The Wall Street Journal

Have a cookie pop-up: Would you like easier European AI rules? - Computerworld

Microsoft researchers demoed Whisper Leak, a side-channel attack that infers topics of encrypted streaming LLM queries by analyzing packet sizes and timing, achieving over 98% AUPRC in tests, while some providers have not implemented mitigations. - The Register

Google reported an RCT of LearnLM with 165 UK students showing a 5.5% performance gain. - Google Org

The Wikimedia Foundation is begging AI platforms to stop scraping and use its paid enterprise API, citing unusually high AI bot traffic and an 8% year-over-year decline in human page views. Also asked generative AI systems to please provide proper attribution. - TechCrunch

Tim Berners-Lee, inventor of the World Wide Web and Inrupt cofounder/CTO, promotes the Solid standard and personal data wallets, built a personal assistant called Charlie, and warns that Big Tech concentration and AI threaten web monetization and discovery. - The Verge

SoftBank sold a stake in Nvidia for $5.8bn and reported net profit doubled to $16bn on valuation gains from its OpenAI and PayPay holdings, freeing capital to boost AI investments. - Financial Times

Sopnendu Mohanty said AI will most heavily automate banks' back-office processing, boost middle-office productivity, and enable net growth in front-office roles; JPMorgan leaders have signaled hiring slowdowns and about 10% smaller operations teams. - Business Insider

But ‘Instasmart’ was right there? Instacart is adding Cart Assistant agentic AI to Caper smart carts and retailer websites, leveraging data from $1.5 billion in orders and a 17 million-item catalog, and will deploy AI-enabled Caper carts in coming months. - Business Insider

Fact-check your work, legal eagles! Latham & Watkins held a mandatory two-day AI Academy for more than 400 first-year associates, demonstrating tools like Harvey and Microsoft Copilot to automate routine legal tasks, while prompting firmwide governance and oversight after a notable AI hallucination. - Business Insider

A 339-member Forecasting Research Institute panel forecasts by 2030 AI will use 7% of US electricity, assist 18% of US work hours, provide companionship to 15% of adults, power ~20% of ride-hailing trips and lift AI investment to $260 billion. - The Register

”Our AI has 30% more paradigm shifting synergy per token”: How to bluff your way to AI credibility with the right buzzwords. Gartner's survey of 700 CIOs says by decade's end all business IT will involve AI, bots will perform about 25% of that work, AI will cause roughly 1% of job losses, and 84% of CIOs lack formal accuracy tracking. - The Register

Executives should pursue AI transformation, blank slate workflow and business model redesign, not mere adoption, illustrated by Shutterstock and Walmart’s strategic replatforming. - Harvard Business Review

Researchers isolate memorization from reasoning in AI neural networks. If you could plug in data without full retraining that would be huge. - Ars Technica

”You give sentient vibes”: Neuroscience experts say no objective test exists for AI consciousness, evaluations can only yield belief-raising evidence based on architecture, activity or behavior, consensus on necessary conditions is lacking, and a stable self-model would be the clearest acceptance criterion. - Gizmodo

Australian researchers developed AMIGO, an AI that corrects JWST's charge-bleed image blurring in software, restoring Aperture Masking Interferometer performance and enabling discoveries such as an exoplanet and brown dwarf near HD 206893; JWST launched in 2021. - BGR

Genetically Engineered CRISPR Babies Are Banned. Sam Altman and Brian Armstrong Are Trying to Make One Anyway. - The Wall Street Journal

Better to seek forgiveness from your mutant super-offspring than to ask permission? Literal Frankenstein stuff?

The combination of MAGA politics and AI and the tech bubble offers only short-term political protection for Donald Trump, not long-term stability. - Financial Times

Trump AI adviser David Sacks alleges conspiracy for anti-AI bad vibes. Researchers say they just think some nonpartisan risk management is in order. Says $1 billion "Doomer Industrial Complex" of Effective Altruism donors includes Dustin Moskovitz and Jaan Tallinn, funds hundreds of think tanks. - Fortune

Practitioners use automated metrics (BLEU, ROUGE-L, BERTScore, perplexity), benchmarks (MMLU, GSM8K, TruthfulQA), verifiers and human review to evaluate LLMs, and industry trends favor hybrid pipelines combining benchmarks, LLM-judges and audits to balance quality, safety and cost. - MachineLearningMastery.com

Kimi K2 Thinking climbs into lmarena.ai top 10, tied with z.ai for top open source slot, and new Ernie, MiniMax M2 still MIA.

AI slop hits new high as fake country artist hits #1 on Billboard digital songs chart. - The Register

Data Miner's Daughter

I Walk The ROC Line

Wichita Optical Fiber Man

Overfit on You

Whiskey Matrix Multiply

These Bootstraps Were Made for Walking

I Shaved My Legs for This AI Cowboy?

Brave AI doggie def going viral on Facebook - Reddit

Comet can't do anything b/c of sandboxing, still vulnerable to prompt injection - Schneier on Security

Not really a fan of agentic browsers. Back in the day I wanted to automate some research flows, thought I could probably write a Chrome extension to do this, it will just look like me clicking around. And it was a struggle. Some stuff worked, make a hotkey to grab info from LinkedIn and update the CRM (until they change the layout)

But it didn't work great, browser JS was pretty lobotomized for safety, cross-browser scripting, can’t really run anything locally. Eventually I switched to e.g. Selenium and Playwright, works better even if you get blocked sometimes.

Today, in Claude Code, you can say search for articles on this, download, apply a prompt.

This works sort of OK out of the box, it can write bash and python code to do what you want. When accessing the Web it asks you if each site is safe, if it's ok to do bash stuff. If you give it some proper scripts with skills will probably work OK and also be safe without full ability to write and run code.

I tried to do this in Comet and it didn't work b/c sandboxed, as it should be. (And yet Comet is still vulnerable to prompt injection attacks)

It took a long time for browser vendors to make JS safe and it got pretty lobotomized, same is happening to agentic browsers. Eventually gets safe enough that it can't do much. At the same time even the basic stuff like ‘summarize this page’ has to be disabled in enterprise, or you need a browser that uses the corporate LLM in Azure or on-prem.

Anyway, if there is a good agentic browser that can do as much as Claude Code and is safe I'd love to hear about it. IMO AI good AI use cases are 1) research i.e. reading a lot of stuff, 2) creating first draft of content, 3) automating workflows, 4) coding. IMO 1) is a good fit for browsers i.e. agentic search, but leave the rest alone.Agents will be interesting until the Web gets totally locked down from stuff like scraping and comparison shopping agents.

Also, use caution…running Claude Code on untrusted Web content that could have prompt injections, with bash access, is unsafe. I've done the thing where you give it full shell access and paste a Docker error and say 'fix this please', and watched it stumble around … just no, this is not the way.

Follow the latest AI headlines via SkynetAndChill.com on Bluesky