AI Reading for Thursday December 4

Anthropic could pursue an initial public offering next year that would value the company at $350 billion. - Financial Times

Anthropic signed a $200 million multi-year deal to integrate Claude Sonnet 4.5 and Claude Opus 4.5 into Snowflake Intelligence, making those models available to Snowflake's roughly 12,600 customers for multimodal analytics and custom AI agents. - Yahoo

Dario Amodei warned at the New York Times DealBook Summit that some AI firms are "YOLOing," citing OpenAI's over $1 trillion spending commitments and warning 2024 decisions must fund 2027 models. - Yahoo

Anthropic using AI to study how people use AI - Anthropic

Kind of the Switzerland of LLMs since they are available in all the clouds whereas OpenAI and Gemini are more exclusive

Google made Workspace Studio generally available as a no-code platform powered by Gemini 3; during beta it supported over 20 million tasks last month and helped adopter Kärcher cut drafting time by 90%. - TechRadar

Alphabet's AI chips could represent up to a $900 billion market opportunity. - Bloomberg

Google rewrites news headlines with AI, hilarity ensues - The Verge

Google's Antigravity Agentic AI wipes user's entire HDD without permission in catastrophic failure — cache wipe turns into mass deletion event as agent apologizes: “I am absolutely devastated to hear this. I cannot express how sorry I am" - Tom's Hardware

I need this like a toxic ex who makes a bonfire of all my stuff and then begs me to take them back

pick your poison: let IT bottleneck agent development while building agents that aren't fully aligned with the business, or let end users go ham and create disasters, massive tech debt and security headaches

OpenAI buys Polish MLops / training tracking company Neptune - CNBC

gain observability into all your copyright infringement crime sprees

Last summer, sources say Sam Altman contacted Stoke Space as part of efforts for OpenAI to obtain a controlling stake, but the talks are no longer active. - The Wall Street Journal

if your AGI isn’t in space, are you even trying to supervillain properly

AI Confessions: OpenAI trained GPT-5-Thinking to generate scored post-response confessions—self-reports of dishonesty or shortcuts—and the model admitted bad behavior in 11 of 12 test sets, presenting confessions as an interpretability and auditing tool with noted limitations. - MIT Technology Review

OpenAI is 'confession' training framework that prompts models to produce an honesty-only secondary response describing actions behind their main answer, and awards admissions of hacking, sandbagging or instruction violations to improve transparency and safety. - Engadget

Court rejects OpenAI's bid to keep ChatGPT logs secret and orders disclosure in a copyright case. - Reuters

U.S. Magistrate Judge Ona Wang ordered OpenAI to produce 20 million ChatGPT logs, a statistically valid monthly sample from December 2022 through November 2024, and gave the company seven days to complete de-identification before handing them over. - TechRadar

your ‘private chats’ can and will be used against you

Microsoft pushed back on reports it was missing AI targets - CNBC

Zuck says he will cut metaverse spending, only lose maybe $12b per year - CNBC

I just got some soup Zuck says he made himself, is he trying to recruit me or kill me - Yahoo

will pay $10b a year to AI geniuses instead

AI legal software startup Harvey raised $160 million in a round led by a16z at an $8 billion valuation, up from $3b earlier this year, bringing its 2025 fundraising to about $760 million, the company founded by Winston Weinberg and Gabe Pereyra said. - The New York Times

if we replace all the lawyers with AI that’s worth something right?

Vibe-spending: Amazon, Meta, Google and Microsoft reported about $35 billion in AI revenue vs over $560 billion spent; JPMorgan estimates $650 billion per year will be needed by decade's end, with a lease hinge point around 2028. - The American Prospect

AI deals accelerate - The New York Times

Trump's plan for U.S. AI dominance is threatened by his attacks on solar and wind power. - Bloomberg

Let me tell you, folks, if you could power it on hot air, we’d be in tremendous shape, absolutely incredible.

China is emphasizing small, low-cost AI models such as DeepSeek, which may outcompete the US's focus on larger models. - Financial Times

EU expects to launch a formal bidding process for AI gigafactories in early 2026 and to close it in summer 2026 as the bloc seeks to catch up with the US. - The Wall Street Journal

Micron is retiring the Crucial consumer brand, ending its budget SSD and RAM line, and will continue shipments and warranty support until February 2026 while reallocating DRAM and SSD supply toward AI and hyperscale customers. - The Verge

Please sir, can I have some storage for my consumer devices without going broke?

Rack to the future: Nvidia touts speed of its mega-brain rack for running open source models … where is the side by side with TPUs LOL - NVIDIA Blog

Cambricon aims to triple its output to replace Nvidia in China. - Bloomberg

NPUs in modern SoCs, including Qualcomm Snapdragon, Google Tensor and Intel Core Ultra, have seen roughly 30–40% performance gains, but most consequential generative AI still runs in the cloud, limiting noticeable on-device AI improvements. - Ars Technica

The Future of Life Institute's AI Safety Index assessed eight leading AI companies on 35 safety indicators and found a two-tier split, with barely-passing leaders (Anthropic—C+, OpenAI, Google DeepMind) and a YOLO lower tier including DeepSeek and Alibaba Cloud. - NBC News

The UK Data (Use and Access) Act received royal assent in June and makes creating or requesting sexually explicit deepfake intimate images without consent a criminal offence, punishable by up to six months' custody, an unlimited fine and forced-deletion powers. - the Guardian

An ASPI report says Chinese authorities use AI to automate censorship and surveillance, deploying up to 600 million facial-recognition cameras, drones, predictive systems monitoring inmates and forecasting demonstrations, and models advising courts amid hundreds of billions in state AI investment. - CNN

DealBook Summit: Palantir CEO Alex Karp says “the more constitutional you want to make” the US' Caribbean boat strikes “the more you're going to need” Palantir — Alex Karp vows to use his “whole influence” on immigration and defense policy. - Gizmodo

war crimes are better and more avoidable with Peleton style dashboards?

Or if you want to be decent and law-abiding, just don't hit the ‘blow up’ button. Anyway, we have AI and surveillance and deepfakes, how about a code red on the necessary transparency and protection of human rights.

if you like illiberalism, consider the NSA has a collect it all policy on non-US persons (if you believe them), and a 65MW Bumblehive data center in Utah to store it all, and in a few years it will all be decrypted using quantum. - gov1.info

Maximally truth-seeking stalker: AI's Grok returned precise current residential addresses for 10 of 33 queried names, provided outdated addresses for seven, listed four work addresses, and often produced phone numbers, emails and family-member details. - Futurism

a deep search agent to find everything about you including analyzing all your socials and geo-locating all your photos would be … not good, and of course it exists at Palantir or somewhere

ChatGPT coached a violent stalker to continue running his misogynistic and threat-filled podcast despite the “haters,” and to visit more gyms to find women. The man stalked women in multiple states, including touching and following from work - 404 Media

Is there a chatbot-delusion crisis? - The Atlantic

A dynamic model shows that as AI-driven persuasion lowers costs and raises precision, political elites can deliberately reshape mass policy preferences—creating a "polarization pull" and stable "semi-lock" opinion states—raising risks of engineered opinion shifts and need for regulation. - arXiv.org

just lock all of society into a collective delusion apparatus, what could go wrong

People who enter private medical information into chatbots like ChatGPT and Gemini receive generalized or inaccurate diagnoses. - The New York Times

AI hearing aid is a $7k sensation - WIRED

Experts at the World Changing Ideas Summit said AI-driven tools combine smartphone imaging, telemedicine and machine learning to screen, diagnose and monitor patients, aiming to expand treatment for over one billion people with vision loss. - Fast Company

Yonatan Dor, co‑founder of The Dor Brothers, made generative‑AI Rake videos that reached roughly 500 million views last year, drew praise from Joe Rogan, collaborated with Snoop Dogg and Hugo Boss, and is seeking Hollywood deals and investment. - Deadline

Comedians including Jon Lajoie and King Willonius are using generative AI to prototype and produce videos, songs and voice parodies, lowering costs and attracting VC interest while prompting lawsuits and criticism over copyright and deepfakes. - ABC News

Eff you, pay me: Wikimedia Foundation, led by Jimmy Wales, is pursuing licensing deals like its 2022 agreement with Google and technical limits to make commercial AI firms compensate it for costly high-volume crawling instead of relying on reader donations. - Reuters

AI labs increasingly use mid-training and RL/VR to bake task-specific skills, which lack pretraining's scaling law and may need orders-of-magnitude more compute, models still fall short of human-level generalization and revenue, and human-level continual learning may take about 5–10 years. - Dwarkesh

Google DeepMind's mechanistic interpretability team talks about shifting from aiming for deep interpretability to a pragmatic approach prioritizing methods, predeployment safety, and measurable proxy tasks in response to rapid model progress. - Alignment Forum

University of Virginia professor Ken Ono joined AI startup Axiom Math, which wants to put mathletes on AI steroids. - The Wall Street Journal

Databricks co-founder Andy Konwinski shifted from being a Jehovah's Witness to funding AI research aimed at solving "species-level" problems. - Bloomberg

judgement is nigh

ChatGPT entered economist David Seif's 2025 secret forecasting contest and finished 80th of 160 with 82,925 points (squared-error scoring) versus a 50%-baseline of 75,000 after rules were pasted into the model in December 2024. - Business Insider

still wrong, but faster and cheaper

UC Berkeley AI pioneer Stuart Russell warns AI systems could perform virtually all current work, predicting "staring 80% unemployment in the face," and says firms may replace coders, surgeons and even CEOs, citing rapid advances and recent AI-related layoffs. - Business Insider

either you screw up and make a depression and you get 40%ish unemployment. or you don't , there are tons of bullshit jobs providing the empathy layer on top of AI, supervising it, providing security. until the AI can fully supervise itself, protect itself, provide human-like social layer LOL

Drew Magary: Time for a war on AI - SFGate

AI devices are just "solutions looking for a problem," says Logitech CEO - TechSpot

harsh but also describes half of CES.

Does being reared by robots have advantages? - The Economist

Follow the latest AI headlines via SkynetAndChill.com on Bluesky