AI Reading for Tuesday August 4

iSue: OpenAI published emails and texts rebutting Apple's suit, showing its lawyers mis-sent a message and admitting a claimed call with OpenAI's counsel never happened. - OpenAI

OpenAI publicly disputes Apple's trade secrets lawsuit, calling it an "oddly personal" fight, as the device partnership turns adversarial. - Financial Times

Apple asked a federal judge for a preliminary injunction barring two ex-employees and OpenAI from using its alleged trade secrets, and sought expedited discovery - Reuters

OpenAI's new GPT-Live voice system removes turn detection for full-duplex speech, cutting latency to match the old system's median - OpenAI

OpenAI's first influencer brand trip to a luxury resort draws backlash over its data centers' rising electricity and water use - Mashable

Fields Medal winner Jacob Tsimerman leaves mathematics to join OpenAI's AI safety team, citing risks of AI takeover and war. - New Scientist

AI ate all your math government funding. - newsletter.ofthebrave.org

AI never trolls Trump. - Terence Tao

The dark night of mathematics? - The Potential Surface

Chrome Sweet Chrome: Google's Gemini Spark now browses Chrome with saved logins to research and book flights or apartment viewings, stopping short of completing payments - Engadget

not necessarily better than asking Claude with Google back-end and browser integration but will bring agents to a large market; seems like a tempting target for prompt injection

Google is financing a $200 billion Wall Street funding structure for Anthropic using private credit, chip leases, and data centre guarantees. - Financial Times

Alibaba released Qwen3.8-Max as an open-weight model for the first time, while DeepSeek's cheaper V4-Flash undercuts OpenAI and Anthropic on price. - theregister

Alibaba's Qwen3.8-Max and Moonshot AI's Kimi K3 are narrowing China's AI gap with the US, creating a 'death zone' for weaker rivals. - Bloomberg

Hugging Face CEO Clément Delangue says China could reach AI frontier parity within one to two years, crediting open collaboration over closed-lab silos. - CNBC

ArtificialAnalysis benchmarks don’t rate Qwen 3.8 Max as high as Kimi K3, on a par with Sonnet 5. Arena.ai human-as-judge rates it the highest open-weights model though. DeepSeek forced OpenAI to cut prices dramatically.

The OpenAI models are also very fast, should see enterprise adoption and stave off China at the the mid- lower tiers. Claude has gotten slow with the massive influx of Claude Code and Cowork users, and Opus 5 is sometimes a little weird.

OpenAI, Anthropic, Google and Meta meet the White House Tuesday to review a voluntary framework for testing frontier models' cybersecurity risks - CNBC

White House officials considered sanctions and trade blacklists on Chinese open-source AI models before pivoting to promoting competitive US alternatives. - The New York Times

Hugging Face CEO Clément Delangue warns that heavy-handed AI regulation risks concentrating power, following last month's Hugging Face hack - Bloomberg

New EU AI Act enforcement powers let regulators evaluate and restrict AI models, with Anthropic and OpenAI already facing scrutiny - CNBC

The FTC banned imports of foreign advanced robots including Chinese-made Unitree hardware used in 90% of recent US robotics research - MIT Technology Review

RAND says falling AI reproduction costs strain export controls, urging safeguards to buy a short window to build countermeasures before the enforcement window closes - RAND Corporation

Navigating AI's risks requires shifting from debate to dialogue, as with the 1986 Challenger launch decision - Fast Company

Mutually assured panic: Beijing officials are anxious that Anthropic's Mythos model could be used as an offensive cyber weapon against China ahead of US talks. - Bloomberg

Huawei's top chip scientist says Nvidia will soon hit physical scaling limits (which Huawei can’t approach yet), and touts Huawei's alternative 'Tau Scaling Law' design approach. - Bloomberg

Loitering with intent: Ukraine's Shrike kamikaze drones now use Auterion's AI to track targets independently, with 50,000 units planned and Russian tanks and helicopters already destroyed - Ars Technica

CVE you later: AI now finds bugs faster than humans can fix them, and its own patches introduce nearly nine times more new vulnerabilities - 4sysops

The Vulnerability Singularity: Kaspersky says AI-driven cyberattacks are outpacing detection, with over half of severe breaches in 2026 going unnoticed for 90+ days - SoyaCincau

Apple's bug bounty inbox got so clogged with AI-generated hallucinated reports that a genuine $200,000 macOS security flaw went unreported - The Decoder

Need better AI to triage the vulnerability reports.

Texas Gov. Abbott paused new data center grid approvals to audit a queue of projects exceeding 474 gigawatts, over five times peak demand. - The Texas Tribune

The logs must flow: Enterprise AI telemetry has tripled in a year and could grow tenfold in two, turning retained logs into ungoverned memory - TechRadar

Asana's new Agentic Work Management gives AI teammates shared, persistent memory across the company while blocking confidential project details from leaking - VentureBeat

Entry level extinction event: Morgan Stanley economists say AI has added 15 basis points (0.15 percentage points) to headline US unemployment, hurting young workers most since ChatGPT's 2023 launch. - MarketWatch

"The details change, but the ending rarely does…tools that dazzled in the demo and stalled the moment they hit the mess of a real, decades-old business." - The New York Times

AI is lowering the cost of starting a business, driving a surge in one-to-nine-employee firms even as skilled-worker shortages persist. - Yahoo

Legacy infrastructure, not AI itself, is enterprise adoption's biggest obstacle, with tech debt costing companies $370 million a year, McKinsey finds. - TechRadar

Enterprises don't yet formally track AI spending, a Norwest VC says, as corporate adoption turns pragmatic and focused on cost and ROI. - The Wall Street Journal

Mistral's revenue reportedly grew twentyfold this year as governments and firms turn to its open-weight models after the US restricted rival AI models. - WIRED

Table Stakes: OpenTable mandates AI agent adoption and usage goals for every employee, aiming to serve over 70,000 restaurants without adding staff. - The Wall Street Journal

poster boy for death by AI and disintermediation:

Front end disrupted by telling your Chrome Spark agent, "find me a French restaurant near Penn Station for Tuesday at 6"

Someone can vibe-code a better restaurant UX / system integration for $20/month

Restaurants have thin margins, feel gouged by an OpenTable monopoly

OFC no one knows anything, they have a lot of data, could also come up with some sick AI features no one else has and somehow use AI to build a moat.

A tracking analysis of the 'AI 2027' takeoff scenario finds governance and security milestones arriving ahead of schedule, even as capability gains lag slightly. - Next Big Future

Honor (Claude) Code: An Alcorn State professor hid the word "Madagascar" in an exam prompt, catching 32 of 35 students who blindly copied AI answers. - SBTech

A Vanderbilt course teaching students how to write better AI prompts raised proficiency from 5% to 48%. - Nature

AI tutors that flatter rather than correct hurt learning: unguarded GPT-4 tutoring cut students' math scores by 17%. - The 74 Million

AI surveillance across 765,962 Danish surgeries found infections in 1.8% of cases; on-time antibiotics cut the risk overall, but not for joint replacements - Nature