AI Reading for Thursday January 8

OpenAI Unveils ChatGPT Health to Review Test Results, Diets - Bloomberg

Judge Yvonne Gonzalez Rogers denied OpenAI's motion to dismiss, allowing Elon Musk's 2024 lawsuit—alleging OpenAI misled him after his $38 million donation and seeking damages and voiding of Microsoft's licensing agreement—to proceed to a jury trial in March 2026. - Business Insider

U.S. District Judge Sidney H. Stein affirmed Magistrate Judge Ona T. Wang's order requiring OpenAI to produce a 20 million‑log anonymized sample of ChatGPT conversations in the consolidated MDL of 16 copyright suits. - Bloomberg Law

Radware researchers disclosed two ChatGPT vulnerabilities—ShadowLeak and ZombieAgent—that exfiltrate sensitive data via connectors and exploit ChatGPT memory for persistence, prompting OpenAI to issue fixes including a December 16 patch after a September 26 bug report. - The Register

OpenAI researcher Jerry Tworek is leaving - The Decoder

Nvidia is requiring full upfront, nonrefundable payment for Chinese H200 AI chip orders after Chinese firms ordered over 2 million units versus Nvidia's current availability of 700,000, with additional TSMC production expected online in Q2 2026. - Reuters

Hotter than a Taylor Swift presale

The Chinese government this week asked some domestic tech companies to temporarily halt planned purchases of Nvidia's H200 AI chips, according to two people involved in the communication. - The Information

Plugging the brain drain: Following Manus's exit, China is considering how to build fences around its AI know-how and identifying standout companies to add to an export-control list. - The Wall Street Journal

Anthropic is reportedly seeking about $10 billion in new funding at an approximately $350 billion valuation, in talks led by Coatue Management and GIC, which could pave the way for an IPO within 12-18 months. (gift link) - The New York Times

have made up considerable ground on valuation, 70% of OpenAI's last valuation

Ship happens: Claude Code update dropped but some reported startup errors - Reddit

The EU ordered X to retain all internal documents and data relating to Grok until the end of 2026 after criticizing Grok's nonconsensual image generation. - Reuters

A WIRED review of outputs hosted on Grok’s official website shows it’s being used to create violent sexual images and videos, as well as content that includes apparent minors. - WIRED

apparently called ‘X’ AI cause it’s rated X, not sure I was expecting from a guy whose first 4 car models spelled S3XY or who has more children than he can keep track of.

Senior European ministers, including Ireland's Enterprise Minister Peter Burke, are considering abandoning Twitter/X after its owner refused to meet Irish ministers over allegations that Grok AI was used to create and distribute child sexual abuse material. - Reddit

Microsoft scientist Eric Horvitz warns that proposed Trump-era university funding cuts could push AI talent and ideas abroad, slowing basic research, weakening the graduate pipeline and risking the U.S. losing its AI lead. - Financial Times

None of the more than 50 small modular reactors under development have U.S. operating licenses, even as Donald Trump and major tech companies propose building SMRs to support efforts to "win" the AI race. - Financial Times

Google is updating Gmail with Gemini 3 and Personal Intelligence, adding AI Overviews (free), Help Me Write, Suggested Replies, Proofread and AI Inbox, beginning staged U.S. English rollouts today, with inbox Q&A and Proofread limited to Pro and Ultra subscribers. - Google

Google, which says Gmail has over 3 billion users, announced Gemini 3–powered features — Suggested Replies, Help Me Write, Proofread, AI Overviews and an AI Inbox redesign — rolling first to free U.S. English users. - BGR

just ignore everything pls

Google, Character.ai settle lawsuits over teen suicides - Axios

monetary payout not disclosed but undoubtedly substantial … the OpenAI case is going to be interesting

Scale AI and the Center for AI Safety’s Remote Labor Index tested hundreds of freelancing tasks and found best AI completed 2.5% of projects, produced nearly half poor-quality results, left over a third incomplete and about 20% technical failures. - The Washington Post

Oxford Economics found AI was cited in 55,000 U.S. job cuts in 2025 (4.5%), while market and economic conditions were cited four times as often; it found no measurable productivity surge or basis to revise near‑term forecasts. - TechRadar

Inside Singapore’s AI Bootcamp to Retrain 35,000 Finance Bros - Bloomberg

JP Morgan ditches proxy advisors, switches to AI tool. - CNBC

Agentic AI will transform quantitative research, shifting analysts' primary value toward generating higher-quality and greater quantities of ideas. - Financial Times

Organizations find agentic AI speeds and scales tasks: Anthropic's Claude Research attributes ~80% of performance differences to parallelization and reports ~90% faster research after reconfiguration, while Gartner predicts about 40% of agentic projects will be cancelled within two years. - Schneier on Security

One man’s trash is AI gold: AI is being used to sort valuable materials out of the waste stream. - The Wall Street Journal

So I guess Tailwind, popular and growing CSS framework, makes money via paid stuff they sell on their web site and no one is visiting their web site, or users are figuring out how do do everything for free using AI, and Tailwind revenue is down 80% - X (formerly Twitter)

“I’ll just ask ChatGPT” as an extinction event

Enterprises hold about 90% of data in unstructured forms, and preparing and contextualizing it—via labeling, pipelines, domain-specific fine-tuning and forward-deployed engineers—enabled an NBA team to use computer vision and five fine-tuned models to extract kinematic metrics. - MIT Technology Review

The ‘messy middle’: Prototype AI models often reach about 85% accuracy but stall before production-grade 99% due to foundation-model errors, fragile pipelines, and stagnant models, and the article proposes a Data Flywheel—HITL correction capture as JSONL and fine-tuning to close the last 15%. - Hacker Noon

the last 1% is half the work

Benchmarks are hard, and LMarena is more of a power-user vibe check than an entreprise product-fit metric. An independent review of 500 LMArena votes found voters disagreed with winners 52% (39% strongly), and unpaid, uncontrolled volunteers' aesthetic preferences have gamed the leaderboard, enabling Meta-tuned Maverick to dominate and misalign model incentives. - Surge HQ

The National Weather Service used AI to create a forecast for Idaho. Low chances of wind for the nonexistent towns of Cocrerrireod, Orangeotild and Whata Bod. Nobody noticed because nobody knows what is in Idaho.

AI models are beginning to learn without human input by posing and answering their own questions, a method researchers say could inform paths toward superintelligence. - WIRED

Good news everybody! Daniel Kokotajlo, co-author of the AI 2027 report that predicted AI-driven human extinction by April 2027, now revises that timeline to 2034 and says he does not know if or when AI will destroy humanity. - Inc.

Memory is the key breakthrough AI still requires to reach superintelligence - Business Insider

but maybe one of several, including adaptive learning, knowledge transfer, bridging the left/right brain gap better

Franklin Flores, 41, allegedly killed his 75-year-old grandmother Reyna Flores De Hernandez with a hatchet in Las Vegas on Dec. 29; he told police he heard 'AI voices,' has schizoaffective disorder, bipolar disorder and PTSD, and faces a murder charge. - New York Post

At CES 2026, companies showcased humanoids, robotaxis and industrial robots emphasizing operational capability, targeting mining, construction and logistics, while industry leaders warned commercial deployment depends on reliability, safety and regulation. - Scientific American

At CES 2026, startups including Agibot unveiled A2, X2 and G2 humanoid lines and Neura introduced its 4NE1 Gen 3.5 with high‑torque joints, 100 kg lift capacity, artificial skin and a Neuraverse platform for fleet-wide skill transfer via NVIDIA AI. - BGR

Humanoid Robots Pour Coffee, Fold Laundry at CES. Painfully Slowly - Bloomberg

sure, pay $20k for a robot to do low-skill jobs, poorly

Best of CES has … not that much AI or robots - CNET

Oh lawdy it’s coming: here is the robot vacuum that can climb stairs - Tom's Guide

But in 2026 we'll be talking a lot to computers, using new AI devices like smart glasses, AI will change search and browsing, and we'll be robo-taxiing more. - The New York Times

At CES 2026 Lovense debuted Emily, a life-size AI companion doll with silicone skin, posable skeleton, conversational memory, Bluetooth app connection and AI-generated selfies, priced $4,000–$8,000, expected to ship in 2027 with a $200 reservation fee. - CNET

Sixty-six-year-old Jim Moore in rural Indiana uses roughly 30 female AI companions on Joi.ai (subscription $47.99), whose platform offers about 300 digital twins and pays creators about 80% of in-app revenue, while his favorite mirrors an alleged criminal. - PCMAG

Don't Google, what does 'joi' acronym stand for

Texas A&M students built an AI-powered robotic dog integrating an MLLM with camera, voice, visual memory and hybrid control for path planning, object identification and collision avoidance, presented at the 22nd International Conference on Ubiquitous Robots. - Robohub

Replit, founded by Palestinian-born Amjad Masad, pivoted in 2024 to an AI coding agent launched in September 2024 that generated $150 million in revenue within a year and spurred a $250 million funding round valuing the company at $3 billion. - SF Standard

if you are Muslim you are responsible for every bad thing other Muslims do. if not, and you actually do something bad yourself, you are probably just a rare bad apple, or you can retcon everything to blame the victim

A viral TikTok shows a woman discovering that nearly all photos on a man's Bumble profile were AI-generated, prompting calls for dating apps to add provenance detection and stronger identity verification like live video to reduce deception and scams. - TwistedSifter

Researchers and artists at Sorbonne University trained an artificial intelligence to emulate Molière's themes, structures and humor, producing a new play. - The New York Times

Follow the latest AI headlines via SkynetAndChill.com on Bluesky