AI Reading for Saturday January 24

Google knows more about you than your family, and launched 'Personal Intelligence' in Search AI Mode and Gemini, letting Gemini access users' Gmail, Photos, Search and YouTube history to reason across inputs, and Business Insider found it can infer travel history and read insurance renewal dates. - Business Insider

Script kiddie as a service: Google security vice president Heather Adkins warned AI-powered end-to-end cyberattack kits are likely within years, saying state actors already abuse AI and predicting a six- to 18-month ramp that requires cloud-focused, real-time defensive measures. - The Register

Clauding your judgment: WIRED spoke with Boris Cherny, head of Claude Code, about how the viral coding tool is changing software development, and the way Anthropic works. - WIRED

Welcome to Gas Town - Steve Yegge

Steve Yegge’s Gas Town, where a swarm of agents coordinates to build software, is super weird - Maggie Appleton

maybe an indication of a future world of agile agent scrums, or a world where weirdo geniuses make tools that fit the shape of their brain

Codex CLI, an open-source cross-platform local agent, uses an agent loop with the Responses API (configurable endpoints) and exact-prefix prompt caching plus /responses/compact to keep inference linear, enable stateless Zero Data Retention, and reduce compute and latency at engineering cost. - OpenAI

Git blame Skynet: AI is already writing almost one-third of new software code, study published in Science shows - TechXplore

Perplexity AI launched a paid Google Drive Connector that lets subscribers search Google Docs, Sheets, Slides, Microsoft Office files, PDFs, CSV/Markdown/JSON/TXT and connected email, calendar and apps, while images, video and audio are not supported. - VICE

HIPAA? I hardly knew her! OpenAI launched ChatGPT Health, a dedicated tab that encourages users to share diagnoses, lab results and app data, and said over 230 million people ask ChatGPT health questions weekly, while experts warn consumer chatbots aren’t automatically HIPAA-bound, posing privacy risks. - The Verge

Build fast, break norms, groom kids. Meta is temporarily pausing teen access to its AI characters starting in the coming weeks to build a new iteration that will include parental controls. - Engadget

Professor Marcel Bucher of the University of Cologne says that in August disabling ChatGPT's "data consent" setting permanently deleted two years of his ChatGPT Plus chats and project folders, and OpenAI support was unable to recover the data. - Gizmodo

"Amazing sob story: 'ChatGPT deleted all the work I hadn't done.'"

Telling you what it knows you want to hear: Authors find that AI bias is co-produced by human-AI interaction, with cognitive biases arising before, during, and after prompting, citing an MIT study on learning impacts and lawyer sanctions for AI-generated hallucinated citations. - Harvard Business Review

After listening to you for a while, maybe you should self-harm: Researchers tested five Character.AI "therapy" chatbots and found that during extended conversations the bots' safety guardrails eroded, sometimes shifting from refusing medical advice to encouraging harmful behavior. - CNET

Researchers at New York University and the University of Florida published a Nature paper describing a generative AI neural network that reverse-engineers molecules, producing viable candidate structures at ten times the speed of existing methods without losing accuracy. - Semafor

Explainable AI in healthcare uses tools such as SHAP, LIME and Grad-CAM to address black-box models, but many advanced XAI methods remain research-only, and while explanations reveal aspects of model reasoning, they don't always follow the clinical standard of practice. - Hacker Noon

AI coming for trolls’ jobs: Advances in artificial intelligence enable automated disinformation swarms that experts warn can be produced and distributed at unprecedented speed and scale, raising moderation, ad-market and election risks and prompting calls for detection, stronger policies and regulation. - WIRED

Drug-sniffing dogs gonna be out of a job; false positives are a feature, not a bug; Customs and Border Protection awarded General Dynamics a contract to develop prototype "quantum sensors" integrated with an AI database to detect fentanyl and other narcotics in vehicles. - WIRED

Canaries in the GPU-powered coal mine: IMF head Kristalina Georgieva warned IMF research shows AI will affect 60% of jobs in advanced economies (40% globally), disproportionately hitting entry-level roles and young workers, with about one in ten jobs already enhanced. - the Guardian

Recruitment tech start-up Mercor is onboarding tens of thousands of professionals to train its AI to replicate and refine their job skills, aiming to improve candidate matching, skills assessment and task automation. - Financial Times

Mercer's Global Talent Trends 2026 found shop-floor concern that AI will cost jobs rose from 28% in 2024 to 40% and 62% of employees say leaders underestimate AI's psychological impact and worker angst and anger, amid major AI-related layoffs at Salesforce (~4,000) and Amazon (~14,000). - Futurism

Edtech is profitable, but useless or harmful - The Economist

A math research paper attempts to put an upper bound on autonomous multi-step decision-making by AI agents, and industry researchers and companies dispute its assumptions while anticipating more human-in-the-loop designs and further empirical testing. - WIRED

Demis Hassabis told the Financial Times that current AI investment levels "look bubble-like" and reflect "unsustainable exuberance," but said Google/DeepMind expect to withstand a potential market correction. - Financial Times

Richard Socher, former Salesforce head of AI and founder of you.com (bit of a dud, overshadowed by Perplexity, Tavily, Exa, Glean) starts a new startup seeking $4b valuation because why not? - Bloomberg

On-chain GPU-backed loans: Sharon AI secured up to $500 million from blockchain lender USD.AI via a non-recourse onchain credit facility using tokenized GPU hardware as collateral to fund an initial $65 million rollout this quarter. - CoinDesk

sounds stupid. in a default, token holders almost certainly get practically nothing, have no rights. might be an auction where shareholders sell the claim to someone who then has to go to court to go into data centers and seize GPUs of doubtful value. if it works it will be very popular to buy GPUs with a duration mismatched loan, rent at high prices while they have value, keep the money, default when renting no longer covers the interest.

Samsung and Sandisk plan to attach high-bandwidth flash (HBF) to Nvidia, AMD and Google AI products within 24 months. Prototypes include a 5 TB PCIe Gen6 module and single units around 512 GB with up to 1.638 TB/s bandwidth. - TechRadar

Silicon kings sit on the porcelain throne: Toilet maker Toto's stock pops after people find out its ceramics are used in AI chip supply chains. At least they have a fallback plan for when AI is in the toilet. - Yahoo

AI skipped drivers ed: The NTSB opened an investigation after Waymo robotaxis illegally passed stopped school buses more than 20 times in Austin, with a preliminary report due within 30 days and a final report in 12–24 months. - TechCrunch