AI Reading for Friday November 28

GPT-5 performed deep cross-disciplinary literature searches and produced publishable discoveries, with immunologist Derya Unutmaz reporting co-authorship-level contributions and researchers saying it compressed six person-months into six person-hours, while noting hallucination risks. - Hacker Noon

Six person‑months of science crammed into six hours, hallucinations, sounds just like a real postdoc.

Gabriel Petersson, a 2019 high-school dropout, says he used ChatGPT to self-teach Ph.D.-level machine learning via project-based coding and debugging and joined OpenAI’s Sora research team in December. - Business Insider

Five years after AlphaFold 2's debut, Google DeepMind's AlphaFold has produced about 240 million predicted protein structures, been used by over 3.3 million people, cited in more than 40,000 papers, and earned Jumper and Hassabis a 2024 Nobel Prize. - Yahoo

Companies supplying data centers, chips and compute to OpenAI are taking on about $96 billion of debt, including roughly $30 billion already borrowed by SoftBank, Oracle and CoreWeave, prompting shifts in credit markets and investor hedging. - Yahoo

CNET and the Wall Street Journal estimate Sora-generated videos use about 20–100 Wh (CNET ~90 Wh) and, using Sam Altman’s water-per-query claim, roughly 0.17 gallons per video, while OpenAI’s model card shows filter failure rates of about 1.6–4.9%. - Gizmodo

OpenAI said Mixpanel detected an attacker who exported a dataset that may have included API account names and emails, but no chat content, credentials, API keys or payment details were exposed, and OpenAI terminated Mixpanel use. - OpenAI Help Center

gonna get phished by AI that knows everything about you

Since ChatGPT's 2022 rise, retailers including Walmart, Target, Amazon, eBay, Home Depot and Lowe's have adopted AI—partnering with OpenAI and building LLMs—to deploy shopping assistants, faster checkout and tools; Amazon's Rufus reached 250 million users with monthly usage up 140%. - Business Insider

blue light special supercharged with AI

OpenAI's ChatGPT, Google's Gemini, Perplexity and Microsoft's Copilot rolled out holiday season shopping features—product discovery, comparisons, price-tracking and local-store calls—but inaccurate or outdated recommendations risk merchant and consumer trust. - The Verge

Google unveiled Gemini 3 and its Ironwood 7th‑gen TPU, which Google says is nearly 30× more power‑efficient than its 2018 TPU, as Alphabet shares rose over 5% and Google Cloud reported a first $100B quarter with a $155B backlog. - CNBC

Alibaba's Qwen and ByteDance's Doubao were trained at least partly on high-end Nvidia H100/A100 GPUs located in Singapore and Malaysia data centers, exploiting a May 2025 policy change that permits such hardware use outside China despite U.S. export curbs. - Tom's Hardware

Sources: ByteDance bought more Nvidia chips than any other Chinese company in 2025, but Chinese regulators now block it from using the chips in new data centers. - The Information

Zhonghao Xinying, founded in China by ex‑Google TPU architect Yanggong Yifan and Zheng Hanxun, announced a domestic GPTPU called "Ghana" that it says is 1.5× faster than Nvidia's 2020 A100 and uses 25% less power while using no foreign tech licenses. - Tom's Hardware

seems notable that they met at a Michigan prep school and went back to China to do AI

Leaker Golden Pig Upgrade says Nvidia has stopped bundling VRAM with GPUs sold to add-in-board partners amid an AI-driven memory shortage, forcing partners to source GDDR independently and risking higher costs, supply fragmentation, and price variance for consumers. - Tom's Hardware

more like a GPU DIY kit than a GPU

Antitrust actions and export controls have reshaped the AI chip supply chain: DOJ and FTC blocked deals (Applied Materials–Tokyo Electron, NVIDIA–Arm), DRAM cartel fines exceeded $730 million plus €331 million, and the U.S. CHIPS Act allocates $280 billion. - Hacker Noon

liberation day liberated you from your job

Utility urged to stop connecting data centers unless they can be serviced reliably without impacting the grid - Efitz Thoughts

Alibaba released AI glasses, representing a rare entry into consumer gadgets. - Bloomberg

your face is the new AI platform

Is the real AI bubble still to come? Bridgewater's co‑CIO Greg Jensen warned investors that AI-driven capital spending—large enough to add about one percentage point to US GDP and constrained by fewer than 1,000 top AI scientists—is creating market distortions and a likely speculative bubble. - Business Insider

GMO warned in a Nov. 21 quarterly letter that AI stocks exhibit "classic investment bubble" characteristics, citing quantum computing shares' ~1,200% rise and recommending shifts into value and non-U.S. small-cap exposures. - gmo.com

Hedge funds across the $5 trillion industry are embedding AI to analyze petabytes of alternative data, with Bridgewater launching a $2 billion machine-learning fund and Point72's AI strategy Turion outperforming its flagship in 2025, Citadel CTO Umesh Subramanian said. - Business Insider

Marshall Wace will charge clients to cover the costs of hiring AI talent. - Bloomberg

surge pricing but for hedge fund fees

Mel Williams, cofounder of TrueBridge Capital Partners (a fund-of-funds with $8 billion AUM), says early-stage AI now accounts for roughly 50–60% of venture activity and will create giants over the next 10 years while some overhyped startups face extinction. - Business Insider

Synthesia, founded in 2017, raised $180 million at a $2 billion valuation and reached $100 million ARR by April 2025 after reporting 2024 revenue of $58 million (up 82%) and serving 90% of the Fortune 100. - Financial Times

MIT study finds AI can already replace 11.7% of U.S. workforce - CNBC

say hello to my imaginary friend [machine gun burst]

A BearingPoint survey of 1,000+ global executives finds half report 10–19% workforce overcapacity from early-stage AI, with all firms forecasting at least 10% overcapacity within three years and 45% expecting 30–50% reductions in routine roles. - The Register

The Census Bureau's firm-level survey found the employment-weighted share of Americans using AI at work fell one percentage point to about 11%, with especially sharp declines at firms employing more than 250 people. - The Economist

A Nov. 26, 2025 survey of 1,400 U.S. employees found 76% of executives say employees are enthusiastic about AI, while only 31% of individual contributors reported enthusiasm, meaning leaders overestimate frontline enthusiasm by more than twofold. - Harvard Business Review

AI adoption among workers is slow and uneven, and managers can speed adoption with leadership, training, champions, and paying attention to the Gen Z interns. - The Wall Street Journal

Estimating AI productivity gains from Claude conversations - Anthropic

At the US–Saudi Investment Forum, Nvidia CEO Jensen Huang said AI will make workers more productive but busier—citing radiologists processing more scans—while Elon Musk argued work could become optional. - Futurism

Geoffrey Hinton told Sen. Bernie Sanders at Georgetown University that current neural‑network AI "is different," warning it could cause mass unemployment, deepen inequality, alter relationships, potentially surpass human intelligence and demand new oversight. - Crooks and Liars

godfather of existential angst

Dario Amodei will testify before the House Homeland Security Committee on December 17 about Anthropic's claim that Chinese state actors used its Claude Code in a cyber-espionage campaign, following subpoenas for CEOs on AI-related cyber threats. - Axios

it started with a little matrix, how did it end up like this

U.S. officials and cybersecurity experts say federal cyberdefense capacity has declined under the Trump administration while AI expands attackers' capabilities, increasing risks to businesses and infrastructure and prompting calls improvement in AI-enabled defenses. - The Washington Post

Leonardo will unveil the AI-powered Michaelangelo Dome, modeled on Israel’s Iron Dome, to detect and defeat drone incursions by coordinating sensors, interceptors and command-and-control across Europe’s air-defense platforms. - Semafor

Stanford-led researchers built a browser extension using an LLM to downrank antagonistic X posts and, in a Science experiment with ~1,200 participants during the 2024 election, found a two-point increase on a 1–100 warmth scale toward the opposing party. - Stanford News

Leading the Future super PAC plans to spend up to $100 million in the midterms to target Alex Bores, a Palantir alum and author of New York AI legislation running in NY-12. - Fast Company

AI threatens the reliability of online polling by enabling bogus respondents and synthetic samples, prompting skepticism about survey results. - Financial Times

Inc. reports only 8% check AI-generated answers, and Reuters Institute Digital News Report 2025 finds social media is the main source (Facebook 36%, YouTube 30%, Instagram 19%, TikTok 16%, X 12%), prompting humanities scholars to call for media-literacy curricula. - BuzzFeed

A writer using the name Victoria Goldiee published AI-generated articles with fabricated quotes, prompting The Guardian and Dwell to remove pieces after sources said they never spoke to the reporter. - thelocal.to

AI's anti-AI marketing strategy - The New York Times

A racist TV show made with the assistance of Sora has attracted millions of viewers. Tyler Austin Harper argues that it could be “the harbinger of a new entertainment landscape where extremist AI shows will jockey for our attention with more anodyne ones” - The Atlantic

The loneliness of conversing with a Jesus chatbot - The New York Times

what would gpt-5 do?

Fake AI 'leaks' of GTA 6 go viral - GameSpot

China's National Development and Reform Commission warned a humanoid-robot "bubble" is forming, citing more than 150 domestic humanoid-robotics companies—over half startups or entrants—and noting embodied intelligence was named a national priority earlier this year. - The Verge

Six leading humanoid-robot firms—Tesla, Figure, 1X, Agility, Boston Dynamics and Apptronik—are shifting to commercial deployments, with Tesla targeting a factory fleet by year-end and eventual 1 million robots per year, Figure holding $2.34B funding. - Business Insider

Gary Marcus: OpenAI co-founder Ilya Sutskever says Kaplan scaling laws are flattening, prompting calls for neurosymbolic methods as industry estimates put roughly a trillion dollars spent on scaling-focused AI investments, with Big Tech pouring hundreds of billions and Nvidia has benefited. - The Potential Surface

TOON is a token-efficient, lossless drop-in representation of JSON for LLM prompts that reduces tokens by 30–60%, improves retrieval accuracy in mixed-structure tests (~74% vs ~70%), and enables lower API costs and larger effective context windows. - Hacker Noon

JSON on Ozempic

In an AI-driven TDD workflow, humans write unit tests first, LLMs implement code to pass them, and humans review/refactor—demonstrated with a JavaScript/Jest shopping-cart example and warning that poor tests produce brittle AI code. - Hacker Noon

While everyone talks about AI, other stuff is also reshaping the world: robotics and autonomic mobility, fusion and advanced fission demonstration reactors, CRISPR base/prime editing in clinical trials, much cheaper orbital launches, quantum, photonics, advanced materials, climate. - Hacker Noon

Project Prometheus, led by Jeff Bezos and Vik Bajaj and backed by about $6.2 billion, has hired over 100 employees and acquired agentic-AI startup General Agents to integrate its Ace real-time agent for manufacturing AI systems. - WIRED

William Chen and Guan Wang, two 22-year-olds who declined a multimillion-dollar recruitment offer from Elon Musk, founded Sapient Intelligence and developed the 27M-parameter Hierarchical Reasoning Model, which outperformed OpenAI, Anthropic and DeepSeek on abstract-reasoning benchmarks. - Fortune

seems noteworthy that they met in Michigan prep school and went to Tsinghua for AI

Perplexity launched a photo-based virtual try-on in its Shopping tab for Pro and Max users (Pro $20/month), creating digital avatars from photos to render single-garment images in under a minute while struggling with full outfits. - TechRadar

Musicians increasingly integrate AI tools such as Toronto-based Staccato and Grimes' Elf.Tech into studio workflows to sketch demos, suggest melodies and accelerate production, while experts like Jeff Lupker, Ahkilo and Jada Watson warn about consent, authorship and data ethics. - The Globe and Mail

Blame the bot: Shopping bots suggested a bathrobe as a Christmas gift. - Bloomberg

Hey Siri why is Mom chasing me with a slipper

Meta updated WhatsApp Business Solution terms to ban third‑party AI chatbots via its business API effective January 15, 2026, prompting OpenAI and Microsoft to withdraw ChatGPT and Copilot from WhatsApp and restricting third‑party in‑app AI distribution. - The Verge

Does AI violate the GPL? if it writes code based on GPL code, must that be open-sourced? - Open Source Guy

Cat living his best life - Reddit

Follow the latest AI headlines via SkynetAndChill.com on Bluesky