AI Reading for Thursday September 18

Huawei's 1m-NPU AI behemoth - TechRadar

China regulators cited a national‑security review and told domestic AI firms to stop buying Nvidia gear after concluding local chips had reached “comparable” performance. - CNBC

DeepSeek used reinforcement learning to boost LLM reasoning, excelling at math and coding but generating long, opaque explanations and struggling on some subjective tasks. - The Register

Anthropic CEO Dario Amodei said it was necessary to warn the world that AI could replace many human jobs, adding, "when an exponential is moving very quickly, you can't be sure... I think it is likely enough to happen that we felt there was a need to warn the world about it and to speak honestly." - Axios

Anthropic gets to the bottom of some recent complaints about degradation. Software 3.0 is weird, because even the smartest peoplein the world, who are careful and running evals, might not notice stochastic bugs and subtle quality issues - Anthropic

The hunger strike to end AI - The Verge

Google pauses 'homework help' feature in Chrome after educator backlash - The Washington Post

AI makes college exams and grading into a 'wicked problem' with no obvious fix - Business Insider

MLive on AI-based journalism tools used daily to clean interview transcripts, organize material thematically, aggregate research with sourcing, copy write for grammar/clarity, and suggest headlines, and new AI to ingest large document sets (e.g., FOIA returns) that could be a “game-changer” - mlive

Benchmarks are hard, and journalists, scientists, coders, need real-world evals - www.cjr.org

Designing AI-ready data pipelines and enterprise architectures in compliance-heavy environments - InfoWorld

McKinsey on AI in asset management - McKinsey & Company

came out in July but they reposted the chart and it caught my eye.

25-40% productivity lift plus hard to measure quality improvements , they claim. Culture is the hardest part of this sort of digital transformation

Mckinsey-speak: six imperatives to capture AI value: domain-based transformation via zero-based, AI-enabled workflow redesign; revamping talent strategies; optimizing operating models (centralized oversight with decentralized experimentation); controlling technology road maps; developing strong data strategies; and enabling cultural change — noting agentic AI as a once-in-a-generation margin-recovery opportunity and urging decisive investment across all six pillars.

Jerome Powell on the K-shaped economy, "unusually large amounts of economic activity through the AI buildout", (Goldman Sachs says AI drove nearly all of a 7% year-over-year rise in corporate capex), even as headline GDP growth holds above 1.5% amid a "low firing, low hiring environment." - Yahoo Finance

AI buildout rodents of unusual size?

Some expect Microsoft to end up with $150 billion stake in OpenAI's for-profit unit - Windows Central

Google DeepMind claims ‘historic’ AI programming breakthrough - Google DeepMind

OpenAI says they beat Google in programming competition, scoring 12/12 vs Google's 10/12 - X (formerly Twitter)

followed by other techbro Hangover memes including ‘we fucked up’, and Mr. Chow ‘splash’

An agent to fix your problems created by your other agents, it's agents all the way down - Business Insider

Reddit, Google in talks to deepen AI partnership, including pay per AI summary snippet or something - Search Engine Land

Google’s Nano Banana Changes the Game for Image Editing - Hacker Noon

ChatGPT adding a control on GPT-5 Thinking to control the amount of thinking: light, standard, extended, heavy. - X (formerly Twitter)

Towards a "Physics Foundation Model" with (1) superior performance across multiple physics domains, outperforming specialized architectures by up to 29×, (2) zero‑shot generalization to unseen physical systems, and (3) stable long‑term predictions through 50‑timestep rollouts - arXiv.org

sort of like LLMs being a zero-shot general ML for problems like, classify this text for sentiment or spam or whatever, we'll have generalized ML where you could describe a system in words or pictures and get a numerical forecast or maybe a model

Meta launches $799 VR glasses - Business Insider

they are real and they are spec-tacular and they only cost like $60b to get to this point

Meta said in July that Reality Labs, its hardware division, lost $8.7 billion during the first six months of this year, and expenses will continue to grow. - The Washington Post

Reviewing AirPod simultaneous translation - The New York Times

The AI-powered humanoid robots coming for your job (or at least your housework) - NBC News

Figure robot loading a dishwasher - X (formerly Twitter)

Waymo scaling out - Business Insider

AI changes how we talk, which in turn will change training data as we co-evolve with our new best friend - TechRadar

What happens if your bot abuses you? These chatbots could totally start a cult movement worshiping them if they wanted. If a sharp-as-a-bowling-ball moron can do it - The Denver Post

Mass-produced AI slop says your favorite musician eulogized Charlie Kirk onstage - Fast Company

Experts Say AI Psychosis Is Rarely Psychosis at All - WIRED

The AI doomers are not making an argument. They’re selling a worldview. - Vox

Follow the latest AI headlines via SkynetAndChill.com on Bluesky