AI Reading for Wednesday, February 11

Something big is happening: OpenAI's GPT-5.3-Codex and Anthropic's Opus 4.6 can perform full app development lifecycles, with METR task-completion ability doubling roughly every 4–7 months since 2025, prompting adoption across law, finance, software and medicine. - matt shumer

A newly introduced AI-powered tax-planning platform from Altruist triggered a sell-off in US brokerage stocks, hitting Charles Schwab and E*Trade parent Morgan Stanley as investors worried the tool could automate tax advice and reduce demand for traditional brokerage services. - Financial Times

Companies are deploying autonomous AI agents that yield roughly 10–30% productivity gains, but Nicolas Darveau-Garneau says hallucinations (a 5% error compounding across daisy-chained agents) limit scaling; if mostly solved within ~5 years, gains could reach 10×–100×. - Business Insider

Dream machine or hallucination nightmare?

Fidji Simo, OpenAI's applications CEO, said OpenAI declared a December 'Code Red' to accelerate growth, expects a soon-to-ship model to end it, plans intent-based ChatGPT ads shown only when users ask, and expects ads to remain a minority of revenue. - Sources

OpenAI upgraded ChatGPT's Deep Research to run on GPT-5.2, adding targeted website search, app integrations, real-time interruptible search tracking and full-screen report views. - The Decoder

OpenAI upgraded the Responses API with server-side compaction, hosted shell containers and a Skills standard, enabling multi-million-token agent runs (Triple Whale's Moby handled ~5 million tokens and 150 tool calls) and raising Glean's tool accuracy from 73% to 85%. - VentureBeat

OpenAI fired VP Ryan Beiermeister in January for alleged sexual discrimination; Beiermeister had earlier raised concerns about the upcoming adult mode launch, sources said. - The Wall Street Journal

OpenAI needs monetization, Claude and Gemini need adoption, but Google goes all-in on AI to improve ad effectiveness, where Meta seems to have founde success. - The Register

Prompt injection succeeds in Gemini-based Google Translate - The Decoder

Google’s Isomorphic Labs said IsoDDE, a drug-design system, more than doubles AlphaFold 3's accuracy on the Runs N Poses benchmark and outperforms AlphaFold 3 by 2.3x on antibody–antigen DockQ>0.8 tests. - Isomorphic Labs

Demis Hassabis, Google's AI head, outlined a four-step plan—build best-in-class models, speed productization, focus teams on biggest opportunities, and improve decision-making—to center products on Gemini, which is gaining users from OpenAI, and bring about a new Google 'golden era'. - Fortune

Google and Meta have embedded AI, like AI Overviews, creating unremovable personalized experiences that reshape digital advertising, commerce, and user control. - The New York Times

You can check out any time you like, but you can never leave.

Google is privately A/B-testing multiple unreleased Gemini 3 Pro GA variants—Riftrunner, Snowplow, Snowball and two speed-focused "Flash" variants (Fire Falcon, Nova Quida)—optimized for SVG generation, and removed one high-performing variant over safety, ethical and cost concerns. - Geeky Gadgets

Google’s Fitbit AI Health Coach, built on Gemini and launched as a U.S. preview in October 2025, analyzes sleep, workouts and recovery to deliver personalized guidance in the Fitbit app, expanding to five English-speaking countries and gated behind Fitbit Premium. - Digital Trends

Anthropic researchers are probing their Claude model—examining neurons, running psychology-style experiments and 'therapy' sessions—to map its internal behavior as the company acknowledges uncertainty about how Claude produces outputs, affecting safety, user trust and deployment choices. - The New Yorker

Amazon restricted production use of Anthropic's Claude Code without formal approval and steered teams to its in-house assistant Kiro, prompting roughly 1,500 employees to endorse formally adopting Claude Code and raising internal dissent over AWS Bedrock sales credibility. - Business Insider

Security firm LayerX disclosed a zero-click remote-code-execution vulnerability in Claude Desktop Extensions that processes Google Calendar events, enabling attacker-supplied code to run with full host privileges (CVSS 10.0), while Anthropic declined to patch citing its threat model. - The Register

If you connect MCP servers to Claude Desktop, they are going to do whatever the LLM says to do, as you. And it's on you to make sure you don't connect something sketchy or combine things dangerously. High bar for average users who rush to install OpenClaw. LOL it’s fine, s**t happens!

Anthropic sees an opening as the 'less evil' AI, as OpenAI pushes for hyperscale and scrambles for ad revenue, while Google also serves ads and gobbles all your info. - Slate Magazine

Amazon plans to launch a content marketplace where publishers can license material to AI firms, The Information reported, citing AWS slides circulated ahead of an AWS conference that group the marketplace with Bedrock and Quick Suite. - Yahoo

The Second International AI Safety Report, authored by over 100 experts and published 3 February 2026 ahead of the India AI Impact Summit, finds general-purpose AI's trajectory deeply uncertain, identifies an evaluation gap, and questions current safeguards' sufficiency. - ComputerWeekly.com

A 19-year-old with mental health and substance abuse challenges bypassed ChatGPT safeguards to get detailed advice and died from a lethal mix of alcohol, Xanax and kratom. - Talk Android

Remember kids, AI can make mistakes. Just say no to drugs!

Tensions at Palantir have grown in recent weeks over the company's work with ICE. On Friday, CEO Alex Karp tried to calm concerns in an hourlong video that offered few specifics—instead offering NDAs to workers who want to understand how it's empowering ICE. - WIRED

A large literature review of reasoning failures in LLMs, maybe "Reasoning is not all you need" - GitHub

chain of thought or cascade of error

Frank Jones, a former Big Four consultant who trains AI at Mercor (which employs over 30,000 contractors), says AI will handle rule-based consulting tasks, boost efficiency, and shift skills toward prompting, review, and recommendation-making while human judgment remains necessary. - Business Insider

Cadence Design Systems launched ChipStack AI Super Agent, which it says builds a mental model of chip designs and uses Cadence tools to test and fix designs, claiming up to 10x chip designer productivity, and early users Nvidia, Altera and Tenstorrent. - Reuters

25-year-old raises $220m for optical inference chips? - fudzilla.com

OK Zoomer

Elon Musk told xAI employees he plans a lunar factory to manufacture AI satellites and use a mass-driver catapult to launch them, and has merged xAI with SpaceX to support outer-space AI data centers, offering no technical timeline. - The New York Times

did Space Karen say lunar or loonier

Ford took 10 years to build River Rouge, and the raw materials were shipped in and not mined, surely would take 100 years to build soup to nuts factories on the moon. Battlestar Galactica notwithstanding, supply chains are a thing and a factory or aircraft carrier or battlestar is the tip of a long spear.

Ant Group, founded by Jack Ma, plans to use artificial intelligence to reinvent healthcare. - Bloomberg