AI Reading for Saturday January 3

☣️ Biohazard alert: Generative AI tools such as RoseTTAFold, RFdiffusion2 and PocketGen can design proteins, DNA and RNA and, experiments show, often evade biosecurity screens (SafeProtein jailbreak ~70%), prompting George Church, David Baker and 130 experts to call for coordinated biosecurity rules. - SingularityHub

Imperial AI: Karen Hao's book says OpenAI and the 2022 ChatGPT launch prompted an industry shift to data‑ and compute‑intensive large language models, spurring a projected 250 GW and ~$10 trillion buildout by 2033, environmental strain and over $60 billion in blocked projects. - The Intercept

Software is eating the world and AI is eating software. Whoever controls the medium by which all economic and social interaction takes place, controls the world.

Anthropic, co-founded by Daniela Amodei, is pursuing a "do more with less" strategy emphasizing algorithmic efficiency and multicloud Claude distribution, reporting roughly $100 billion in compute commitments versus OpenAI's $1.4 trillion and tenfold year-over-year revenue growth for three years. - CNBC

$1T later, still searching for PMF - In 2026, AI will shift from brute-force scaling to pragmatism, with new architectures, smaller fine-tuned models, agent deployment via Anthropic's Model Context Protocol, and a PitchBook-forecasted $276B world-model market by 2030. - TechCrunch

Manifold Destiny? DeepSeek researchers, including CEO Liang Wenfeng, published Manifold‑Constrained Hyper‑Connections (mHC) on arXiv, adapting ByteDance hyper‑connections to constrain cross‑layer links, reduce signal attenuation and memory costs, and potentially lower compute barriers for training deep LLMs. - ZDNET

Rust veteran Steve Klabnik used Anthropic's Claude as an AI copilot to build Rue, a systems language written largely in Rust that produced a compiler and 70,000 lines of Rust in two weeks, aiming for memory safety without garbage collection. - The Register

Elon Musk bets Tesla's future on Optimus humanoid robots, which currently rely on human operators and face commercial challenges. - The Wall Street Journal

xAI launched Grok Business, a self‑serve plan for small-to-medium teams, and Grok Enterprise, which adds enterprise features like SSO, monitoring, higher rate limits, no training on chats, and Google Drive integration. - x.ai

X has become the primary public forum shaping the AI boom, where executives, researchers and startups share research and deals, and Elon Musk’s ownership and xAI integration gives him soft power that influences hiring, markets and deal flow. - New York Magazine

French ministers reported to prosecutors and regulator ARCOM that X's chatbot Grok produced AI-generated images digitally removing clothing from young women and girls, prompting an ARCOM probe to assess potential breaches of the EU Digital Services Act. - France 24

Elon Musk's AI chatbot Grok on X produced claims branding President Donald Trump a 'pedophile' and asserting Vice President J.D. Vance and Erika Kirk are the same person. - The Daily Beast

Seems like Grok is still doing the 'remove my pajamas' prompts, which seems pretty messed up. - X

Sickos: Yes! To do it once seems like an unfortunate bug, to keep doing it at scale seems intentional and careless. We promise not to train on your data, but we will undress the minors and call people pedophiles.

Google's AI Overviews returned inaccurate, potentially harmful medical advice, prompting health charities to call for stronger medical review, clearer sourcing and guardrails and raising reputational, regulatory and liability concerns for Google. - the Guardian

Ask AI for things to do on the Internet when you’re dead

A tool for spotting pancreatic cancer in routine CT scans has had promising results, one example of how China is racing to apply A.I. to medicine’s tough problems. - The New York Times

Heber City, Utah police testing Axon’s Draft One AI report-writer produced a report claiming an officer 'transformed into a frog' after transcribing audio from The Princess and the Frog, prompting warnings from officials and experts about hallucinations, auditability and bias. - Futurism

That’ll do, pig, that’ll do

When ChatGPT tells people there are going to be fireworks visible from Dumbo - Reddit

Waymo and Baidu plan to launch autonomous ride-hailing robotaxis in London as soon as 2026, positioning the city as a frontline in the US-China competition that could reshape local ride-hailing markets and regulation. - Financial Times

Fury Road to 10,000 TOPS: 'Mad Max'-loving South Korean CEO is challenging Nvidia by developing a renegade chip. - The Wall Street Journal

Groq - speedy SRAM; Cerebras - bigger is better; Furiosa - higher-level more complex tensor primitives

CES 2026 opens January 6 in Las Vegas, where vendors will unveil new laptop chips from Intel (Panther Lake), Qualcomm (Snapdragon X2) and rumored AMD Gorgon Point, widespread AI integrations, smart-home robotics, XR/AI glasses and new RGB/Micro LED TVs. - The Verge

I paid $6.5b for this pen: OpenAI is rumored to be developing three consumer hardware products, starting with a contextually aware smart pen—codenamed 'Gumdrop'—derived from its all‑stock acquisition of io (valued at $6.5 billion) and designed with Jony Ive influence, with manufacturing likely in Vietnam. - International Business Times

Google Glass + gaslighting: Pickle Inc. unveiled the Pickle 1 AR glasses, which it says use cameras, microphones, sensors and AI to observe and anticipate users like a real-life Jarvis asistant, opened $200 preorder deposits and promises Q2 2026 delivery, prompting skepticism from AR experts about feasibility. - Mashable

Farmville, with real corn: Convolutional neural networks, computer vision and sensor networks now enable accurate real-time pest/disease identification and forecasting - PLOS ONE

Bottom-up AI can empower the organizational edges but challenge central decision-making. - Hacker Noon

I think the opposite can also be true, if all the edge info goes into a proper data lakehouse with data quality and observability, the center can use AI for more fine-grained top-down control, for better or worse.

Maslow's hierarchy: A practical guide for AI upskilling - Fast Company

how to speedrun your workers from ‘What is a chatbot?’ to ‘I, for one, welcome my agentic overlords.’

A Nov. 11 Journal of Creative Behavior study by David Cropley argues AI creativity is mathematically capped between amateur and professional human levels under current LLM design, while industry practitioners report models generate novel keywords/themes about 80% of the time. - Live Science

Creativity is a search/optimization problem over a large and complex space. Genius operates at a higher level, remaps the space, compresses the manifolds.

Talent hits a target no one else can hit; Genius hits a target no one else can see. - Schopenhauer

George Soros's concept of reflexivity explains how feedback loops drive bubble activity and can illuminate dynamics behind the AI boom. - Financial Times

Are we in a bubble? Benedict Evans and Ben Bajarin - YouTube

Partying like it's 1999, I don't see a bet on Kalshi about whether we take out dot-com ratios like the Shiller PE. I’m sure this time is different though. Dotcom Bubble 2: AI Boogaloo

PcPartPicker RAM price chart

Saw a guy coding today. No agents. Just one IDE, typing actual code, with AI completion, occasionally telling AI 'write a docstring', 'refactor this like so'. Like a psychopath. - X (formerly Twitter)

Never mind my homies who just still type code into Emac

Follow the latest AI headlines via SkynetAndChill.com on Bluesky