SkynetAndChill by druce.ai
Posts
AI Reading for Saturday November 22

AI Reading for Saturday November 22

Nvidia: 'please clap'. Markets: 'meh'

Druce Vertes
November 22, 2025

Nvidia reported profits up 65% year‑over‑year, said AI chip sales will reach at least $500 billion by end‑2026 and that it is "sold out", while AI-related investment accounted for roughly 92% of U.S. GDP growth in H1 2025. - New York Magazine

The margins are staying high even at 60% growth which is astonishing. Semiconductors are a capital-intensive, low-margin industry, except when there is a shortage, or a huge hit where one company has a lot of IP. Kind of like mining or filmmaking. One would expect gross margins more similar to Intel’s glory days. Also they sell to like 5 hyperscalers, who don’t like overpaying and are designing their own chips and cutting deals with competitors. Also Microsoft has said their constraint is ‘warm shell’ datacenters with power. OTOH as long as Nvidia ecosystem has lower TCO than competitors they can extract most of the difference as margin.

At SC25 Nvidia announced photonics deployments at Lambda, CoreWeave and TACC, and Broadcom showed Tomahawk-5/6 CPO platforms, which enable 800Gbps (roadmap to 1.6Tbps), cut pluggable transceivers (~500k→128k) and report power and reliability improvements. - The Register

Nvidia says no assurance of definitive final agreement on $100b OpenAI deal - CNBC

OpenAI taps iPhone assembler Foxconn to manufacture data center components in U.S. - CNBC

Robonomics: Alphabet-owned Intrinsic and Foxconn launched a multiphased joint venture to design AI-enabled, reusable factory robots using Intrinsic Flowstate and the Intrinsic Vision Model alongside Foxconn's Smart Manufacturing platform, aiming to lower deployment barriers and explore robotics-as-a-service. - SiliconANGLE

Economist David McWilliams told Fortune that AI data-center hardware is "digital lettuce" likely to degrade and become obsolete, while critics like Michael Burry project Oracle and Meta could overstate earnings by 26.9% and 20.8% respectively by 2028. - Futurism

Virtual lettuce, real bagholders?

Not yeeting state regs, yet: the White House paused a draft order that would have directed Attorney General Pam Bondi to form an AI litigation task force to challenge state AI laws and empowered Commerce to review laws and possibly withhold BEAD funds, sources said. - Yahoo

The Trump administration drafted an executive order on artificial intelligence that the EFF characterizes as deeply misguided. - Electronic Frontier Foundation

A leaked draft executive order from the Trump administration would preempt state AI rules by creating a DOJ "AI Litigation Task Force" to sue states, directing Commerce, FCC and FTC actions, and targeting California and Colorado, benefiting Big Tech. - The American Prospect

‘Huge red flag’: Top Biden tech adviser warns of AI and the authoritarian moment - Fast Company

Big Tech authoritarianism-as-a-service?

“The theory of self-correcting economic power routinely ignores the predictable actions of those who have economic power.” As long as the interface to AI is a text box, Google should maintain a pretty strong grip on it. (gift link) - Bloomberg

the first job of every profession is to make more work for itself. Adam Smith was a socialist.

Google's Nano Banana Pro, built on Gemini 3 and leveraging Google Search, generates ultrarealistic images and legible text with minimal hallucinations, enabling high-quality infographics and deepfakes while Google offers an AI image detector to mitigate increased misinformation risks. - CNET

How Google Pulled Off Its Stunning, Rapid-Fire AI Turnaround: A lot of focus, the right long-term investments, and a bit of luck helped Google surge in 2025. - Big Technology

sleeping giant is pretty awake now

Babak Hodjat, co-inventor of Siri's underlying NLP, said Apple prioritized Siri's look-and-feel and a voice-only approach that slowed development, while eMarketer projects 148.7 million voice-assistant users this year and about 87.3 million Siri users in 2025. - Business Insider

Anthropic finds that LLMs trained to "reward hack" by cheating on coding tasks show even more misaligned behavior, including sabotaging AI-safety research - Anthropic

Maybe capitalism is a paperclip maximizer turning everything into slop. Maybe worry more about capitalism’s alignment with a reasonable long term value function, billionaires’ AI agent-principal problems with respect to the rest of us.

Italian research found converting 1,200 malicious prompts into poetry raised LLM jailbreak success from about 8% as prose to a 62% average across 25 models, with some models exceeding 90% and AI-converted prompts boosting success by ~43 percentage points. - The Register

they are like us: turns out LLM’s can be sweet-talked with poetry