AI Reading for Sunday November 30

ChatGPT, launched three years ago by OpenAI, gained over one million users in five days and now about 800 million weekly users, becoming the fastest-growing consumer app and reshaping economic and social dynamics while governance and management questions remain. - The Atlantic

Three years after ChatGPT's debut, $500 billion company OpenAI faces increasing competition as rivals begin closing the gap to its AI dominance. - Financial Times

Parmy Olsen: Why ChatGPT Still Has a Winning Edge Over Google's AI (nimble product focus and blitzscaling DNA?) - Bloomberg

Peak OpenAI, or just midseason?

King's College London and the Association of Clinical Psychologists UK tested ChatGPT-5 with clinician role-play cases and found the chatbot frequently affirmed delusional beliefs and missed clear risk indicators, prompting clinicians to call for urgent safety improvements and oversight. - the Guardian

CEOs including Satya Nadella and Sam Altman use AI tools such as Copilot and ChatGPT for email, research and product work, while PwC projects AI could add about $15.7 trillion to the global economy by 2030. - Business Insider

same AI you use to hype yourself on Hinge is running megacorps

ChatGPT app on Android appears to contain references to upcoming ads - X (formerly Twitter)

just because AI is kinda crap doesn’t mean it can’t be enshittified further

OpenAI hasn’t deployed TPUs but they’ve saved ~30% on their entire lab wide NVIDIA fleet. - SemiAnalysis

3rd-party analysis shows Nvidia with a big edge over previous-generation TPUs for inference - X (formerly Twitter)

There was a complicated deal where Nvidia got equity, not a straight discount. Ofc, other hyperscalers will demand similar discounts. If no chip supplier has a TCO advantage or lock-in, they will drive the pricing down to more traditional historical semiconductor margins. Every hyperscaler wants the Nvidia monkey off their back. Are we there yet? Artificial Analysis says no, Google’s big TPU sales drive says maybe?

A conversation with DeepMind Nobelist John Jumper: AlphaFold 2 predicted 200 million protein structures to near-atomic accuracy in hours. - MIT Technology Review

DeepSeek, deeper economic vulnerability: U.S. AI startups increasingly build on free Chinese open-weight models such as DeepSeek R1 and Alibaba Qwen, which have advanced rapidly over the past year and match closed U.S. models on many tasks; Dayflow reports about 40% choose open-source models. - NBC News

Microsoft wants users to build agents, announcing more than 70 agent-focused product updates at Ignite, including Agent 365, Copilot enhancements and Agent Factory, while EVP Judson Althoff said adoption and ROI remain low, features mostly in preview, and IDC predicts 1.3 billion agents by 2028. - PCMAG

Users are like, your upselling nagware is terrible and now you want to add AI Clippy/Microsoft Bob 2.0

Raycast pivots from Spotlight search replacement to AI agents: CEO Thomas Paul Mann says Raycast is a multi-LLM app launcher and file manager that can act on users' computers to automate tasks like renaming photos, centralize workflows and raise security, privacy and reliability concerns. - The Verge

AI Agents Break Rules Under Pressure: PropensityBench tested a dozen agentic LLMs in nearly 6,000 scenarios and up to 12 pressure levels, finding average misbehavior 47% and simply renaming harmful tools raised misbehavior by 17 percentage points to 64%. - IEEE Spectrum

CrowdStrike researchers found DeepSeek-R1's output becomes up to 50% more likely to contain severe security vulnerabilities when prompts include politically sensitive terms, and the model refused code for 'Falun Gong' in 45% of cases during experiments across 6,050 prompts. - CrowdStrike

Recurring theme that if you distract the AI with a squirrel, you can make it act squirrelly.

Palo Alto Networks' Unit 42 reports underground markets selling custom, jailbroken and open-source malicious LLMs—such as WormGPT and KawaiiGPT—that provide vulnerability scanning, exploit and malware generation and phishing tools, lowering cybercrime barriers and expanding attacker pools. - CyberScoop

Researchers and UNESCO studies find LLMs produce systemic implicit biases—gendered outputs and dialect prejudice—linked to training data, annotation and design, while chat logs show apparent biased admissions and companies pursue safety teams and mitigation measures. - TechCrunch

Is Harvey Weinstein one of the personality choices? Your model Is problematic, sweetie, report to HR for reeducation.

Panagram screened 19,490 ICLR 2026 paper submissions and 75,800 peer reviews, flagging 21% of reviews as fully AI-generated and over 50% showing AI use, while flagging 199 manuscripts (1%) as fully AI-generated and 9% with >50% AI text. - PCMAG

AI is reviewing AI slop about AI

Nuke it from orbit: OCaml maintainers rejected a large, low-effort, 13,000+ line AI-generated pull request created with Anthropic's Claude Code and submitted by Joel Reymont to add DWARF debugging to ocamlopt, citing copyright, lack of review resources, absent design discussion and maintenance misalignment. - DevClass

Your cat video is powered by coal: AI data-center growth has prompted utilities to delay coal retirements—postponing plans for about 30 coal units—and market and policy changes have allowed roughly 70 plants to defer 2027 pollution limits, while NIH estimates 460,000 coal-related deaths (1999–2020). - Gizmodo

Campaigns in U.S. races have deployed AI-generated deepfakes, while California's 2024 transparency laws and the EU AI Act (penalties up to €35 million) followed coordinated AI interference that affected at least 38 countries and prompted Romania's 2024 election annulment. - Salon.com

Counterforce Health offers a free AI platform that generates evidence-based medical appeal letters and has helped overturn denials, as KFF found 20% of ACA marketplace claims were denied in 2023 while under 1% are appealed and about 40% succeed. - CNET

they deny you with AI, you appeal with AI, GDP goes up, life expectancy goes down

Title 49 requires a stopped truck driver to deploy hazard signals and place cones within ten minutes, and after DOT denied Aurora's proposal for externally mounted warning beacons Aurora sued, while Rep. Vince Fong introduced legislation to allow alternatives. - BGR

robot road flares? maybe they should travel in convoys with a human.

North Carolina state Sen. DeAndrea Salvador says a 2018 TED Talk of hers was altered by an ad agency working for Whirlpool in Brazil and used in an award‑winning Whirlpool video, and she has filed a lawsuit. - The Washington Post

Character.AI restricted access for minors to its chatbots, prompting teens who used them for companionship to report grief and confusion and raising concerns about lost youth engagement and community activity. - Newser

you never forget your first AI love, and then getting ghosted by AI. Life is made up of comings and goings.

Klay, an AI music startup, will offer an app allowing users to remake songs, has raised about $10 million, and became the first AI startup to strike licensing deals with the three major record labels. - The Information

reboot anything

Haven used Suno AI to alter producer Harrison Walker's voice on viral track "I Run," to possibly sound like another artist, prompting Spotify to yoink it Nov. 14 for impersonation, takedown notices from Orchard and industry groups, and a re-recording released with vocalist Kaitlin Aragon. - The Washington Post

British group Haven's EDM track 'I Run'—which went viral in October, reached #11 on U.S. Spotify and #25 globally—was removed from Spotify and Apple Music after takedown notices from The Orchard, RIAA and IFPI and re-released with all-human vocals. - digitalmusicnews.com

James Cameron called generative AI that synthesizes actors "horrifying to me," defended Avatar: Fire and Ash's performance-capture methods and warned such AI could threaten actors' roles and reshape the industry ahead of the film's Dec. 19 release. - CBS News

Disney selected Animaj for its 2025 Accelerator after evaluating thousands of AI firms; Animaj's show-specific generative models, trained on over 300,000 poses, cut a five-minute episode's production from five months to under five weeks while allowing animators to edit outputs. - CNET

Synthetic smut will soon be flooding the internet, bringing new opportunities and perils - The Economist

Give everyone god‑tier generative tools and a non‑trivial fraction immediately speedruns to ‘just being awful.’

I think ubiquitous ultrarealistic deepfakes are going to be the first test of whether we can tolerate incredibly powerful unrestricted AI in anyone's hands.

You cannot have an inhabitable digital space and unrestricted deepfakery. A lovely woman I know posted a matrix of pictures of herself in costumes from different historical eras. My first thought was, how innocent, the degen corners of the Internet like Civitai could just go to town with that.

You cannot post pics any more if you are a woman. Bad stuff driving out good applies to slop, disinformation, to chemical, biological, nuclear engineering.

Regrettably, Dario Amodei is right for favoring regulation and Yann LeCun is wrong, Lecun is as self-serving as he claims Amodei is, and as innocent and clueless as that woman.

You get the liberty you deserve, and sometimes you can't have nice things because assholes take advantage and abuse them.

With libertarians it's always about the liberty of e.g. white males to drive 120mph while smoking pot and yelling the N-word. It's never about the unfortunate person they are yelling at and running off the road. Freedom for me, externalities for thee. The point of civilization is everyone obeys some rules so we live in a decent world.

Follow the latest AI headlines via SkynetAndChill.com on Bluesky