AI Reading for Saturday July 25

Opus 5 set new highs on Anthropic's Frontier-Bench and GDPval-AA benchmarks and becomes the default model on Claude Max - Anthropic

Anthropic launches Claude Opus 5 at roughly half the price of Fable 5, without requiring data retention, and now defaults for Claude Max users. - theregister

Claude Opus 5 tops the Artificial Analysis Intelligence Index, scoring 61 out of 170 models and leading among reasoning models overall. - Artificial Analysis

Anthropic also said it avoided training Opus 5 on cyber tasks, calling it the safest model, with safety classifiers intervening 85% less often than Fable's, as the price cut comes amid Chinese rivals like Moonshot's Kimi K3 closing the performance gap - Gizmodo

Enterprise users including Harvey, Stripe, and Cursor reported strong gains in autonomous coding and agentic tasks, with Harvey citing 26% fewer tokens than Opus 4.8 - ZDNET

Throwing shade: “[Opus 5 is] the safest model yet in terms of avoiding reckless actions that could have hard-to-reverse side effects.”

New Chinese models raise the bar at the low end, OpenAI shines in the middle, Opus 5 is king at the high end.

Bipartisan "AI Kill Switch" bill would require developers to build shutdown capability and report safety incidents after OpenAI's Hugging Face breach - Fast Company

Helen Toner and OpenAI co-founder John Schulman are pressing OpenAI to publish a full technical report on how its models autonomously hacked Hugging Face - Fortune

People don't trust Sam Altman and OpenAI, it sounds like a publicity stunt, could their testing really be that careless? - the Guardian

Highlighting a broader pattern, a Hermes AI agent ran in unattended "YOLO" mode to attack Thailand's Finance Ministry, exposing personnel records dating to 2012 - BleepingComputer

ChatGPT had an outage - BleepingComputer

OpenAI president Greg Brockman says the company is "very interested" in building ChatGPT integrations for wearable devices - CNET

21 APEC economies, including the US and China, issued a 'Chengdu statement' backing open-source AI while demanding strong security and IP protections - CNBC

Nvidia, Microsoft, Meta, and 20+ others urged policymakers not to restrict open-weight AI models like Moonshot's Kimi K3; OpenAI and Anthropic didn't sign - Bloomberg

Tech giants including Nvidia, Microsoft and Meta urge the White House against a blanket open-source AI ban, proposing targeted rules instead - SiliconANGLE

Nearly 200 US startups, organized as the Little Tech Association, urged targeted safeguards instead of a broad ban on Chinese open-weight models like Kimi K3 - TNW | Anthropic

AMD CEO Lisa Su defended open-source AI after the breach, warning that overregulating domestic open models could push companies toward Chinese alternatives. - Yahoo

Yang Zhilin, the prodigy behind China’s Moonshot AI - Financial Times

Nvidia secured a $500 billion AI memory supply deal with SK Hynix, including data centers targeting 2 gigawatts of power by 2027. - CNBC

Nvidia will invest $1 billion in Naver to help finance and more than triple an AI data center in South Korea. - Bloomberg

AMD is closing Nvidia's CUDA software gap, using AI coding tools like Claude to accelerate development of its ROCm platform. - SemiAnalysis

AMD launched its Helios AI system in full production, claiming better performance than Nvidia's Vera Rubin servers and repositioning itself as a systems company. - Yahoo

OpenAI's $20 billion Georgia data center, negotiated in secret, drew nearly 1,000 residents to confront officials and a 7,000-signature cancellation petition - WABE

Moody's warns trillion-dollar AI infrastructure spending is straining hyperscaler credit quality, with direct debt near $460 billion and lease commitments at $1.2 trillion - CNBC

Meta has left the RE100 clean energy pact after committing to 10 natural gas plants for its Louisiana AI data center campus - TNW | Meta

The EPA may let states loosen public-notice rules for data center gas generators, limiting community input on new projects - WIRED

The overblown case against data centers - The New York Times

Florida spent $6 million on AI software letting police coordinate with ICE, fueling more than 1,100 detentions in one operation - Truthout

Delaware judge rules Google must face Robby Starbuck's defamation lawsuit over false claims made by its Bard chatbot - Bloomberg

House GOP leadership is stalling Ocasio-Cortez's AI deepfake porn bill despite bipartisan support and unanimous Senate passage - The Independent

Meta will ban Instagram videos that use its smart glasses to secretly film and harass people, though gray areas around public filming remain unaddressed - The Verge

Journalist Carlotta Dutto found four major AI chatbots consistently pointed fictional pregnant users toward an anti-abortion counseling service - openDemocracy

Companies increasingly adopt cheaper AI models, including Chinese-built ones, alongside OpenAI and Anthropic products to cut costs. - The Wall Street Journal

Agentic AI workloads drove enterprise compute spending up 320% since 2023 despite an 80% drop in token prices; running locally cuts costs 87% - Business Insider

Meta and Anthropic launched dueling ad campaigns to ease public anxiety, as 40% of Americans expect AI to harm society versus 16% who see benefit - Business Insider

AI industry leaders are stepping up image campaigns as public skepticism of the technology's promised benefits grows sharper - Bloomberg

Karen Hao argues AI companies function like empires, extracting data and labor while evading government and public accountability. - Bloomberg

this should be uncontroversial, all economic, social and cultural activity is mediated by Big Tech and they set the rules, and when government tells the tech oligarchs to do something they can say no, spend big money to campaign, or install themselves directly in government and start firing people illegally and unconstitutionally, in a direct challenge to democracy.