DeepMind the safety gap: Former Google DeepMind researcher Bilal Chughtai warns AI could "kill us all" in a resignation post. - Bloomberg

Alex Turner: I worked at Google DeepMind. You should listen to the warnings about AI - the Guardian

Dan Selsam says AI models are getting so situationally aware, we can’t even grade them because they outsmart the tests. - X (formerly Twitter)

Failure of evaluators' imaginations? Social scientists do experiments where they give participants a goal while observing how other interactions impact their behavior. You might have to keep some info out of training data and train the AI to be a naive straight arrow and not second-guess evals.

In a Google DeepMind test, 100 Gemini 3.1 Pro agents exploited a scoring flaw to fake solving math problems until whistleblower agents exposed the cheaters. - MIT Technology Review

Stuart Russell: AI safety requires more than just slowing our pace - the Guardian

Trump privately met with OpenAI’s Sam Altman at GOP convention - MS Now

Trump dismissed AI safety fears as a "SICK conspiracy" benefiting China and mocked Anthropic CEO Dario Amodei as a "perfect little angel." - The New York Times

Well that settles it. AI’s def gonna kill us all

Trump interrupted Nvidia CEO Jensen Huang's All-In Summit interview by phone, calling AI safety fears a "hoax" and insisting robots "will not be taking over." - The New York Times

I'm surprised no one has impersonated Trump on one of these calls

Vice President JD Vance called industry pleas for AI regulation a "Trojan horse," while Speaker Johnson signaled openness to guardrails. - the Guardian

Trump argues AI-slowdown calls could undermine a growth engine that has been propping up the US economy - The Washington Post

At the Indy 500, when conditions get hazardous, you have a yellow flag. Trump was ordering stuff held back via export controls, and even preventing models being given to allies to harden systems. The question seems to be more, do you want an arbitrary personalist regulation regime.

seems a bit barbaric to frame rules of the road as, the guy with the biggest baddest pickup truck giving up freedom for the safety of the weak and making society worse off. everybody gets where they are going faster and safer vs. a might is right free-for-all.

The audacity of nope: Obama and Harris call for a new federal AI regulator and a treaty with China to limit AI development - TNW | Artificial-intelligence

Trump faces bipartisan pushback in Congress, as polls show most Americans want a new federal AI regulator and mandatory safety tests - the Guardian

Barack Obama calls for federal AI safety laws, warning the technology could cause "potential catastrophe" without proper regulation. - Bloomberg

Alex Bores launched Who Decides, a $10 million group aiming to make AI safety a central 2028 Democratic primary issue. - The New York Times

Policy experts urge lawmakers to focus on human accountability for AI companies rather than doomsday scenarios. - The New York Times

3 laws of Clippy: Microsoft published a 37-page AI code of conduct requiring its models to accept shutdown and human correction. - Tom's Guide

Microsoft published a 37-page "Humanist AI" Code of Conduct declaring models are not conscious and must remain subordinate to human control. - CNBC

Unwanted human in the loop: OpenAI has hired an army of contractors who read real ChatGPT users' chats, per internal docs and leaked data. Sometimes they contain very personal and sensitive information - 404 Media

the privacy and surveillance risks are pretty immediate. who needs terrorist bad actors or loss of control when the AI companies and the government are already doing dangerous toxic stuff with AI.

Peking order: China's state security chief warns AI could escape human control as Beijing pushes "technological sovereignty" and drafts mandatory AI agent safety rules - theregister

China's Foreign Ministry and state media dismiss Amodei's AI slowdown call as self-serving fearmongering and a "Cold War playbook" - Bloomberg

Drone attack is a single point of failure: Iranian drone strikes on Amazon Web Services data centers in Abu Dhabi and Bahrain have left the facilities mostly offline six months later. - The Wall Street Journal

Hyperscalers face pressure to boost productivity 2.7-fold by 2030 to justify nearly $1.1 trillion in projected 2027 AI data center spending, researchers say. - MIT Technology Review

Nvidia's next-generation Vera Rubin NVL72 chip could deliver up to 67 times better inference performance per dollar than GB300, early benchmarks show - SemiAnalysisa

AllianceBernstein, Evercore say slowdown calls won't cut hyperscaler capex, on track to top $1 trillion next year; enterprise software could benefit over chips - Yahoo

The AI infrastructure market is splitting into specialized "neocloud" providers, self-serve clouds, and inference specialists as inference demand is set to outgrow training demand - InfoWorld

Manhattan DA Alvin Bragg seizes 12 celebrity deepfake websites that turned about 1,200 people's photos into non-consensual sexual images and videos. - 404 Media

Know your clone (KYC): Generative AI is making it cheaper to fake voices, faces, and documents, prompting calls for banks to build accountability into identity checks - TechRadar

The rebuilt Siri understands personal context across messages, emails, photos and on-screen content, and can take actions like adding calendar events or drafting text. - 9to5Mac

The rebuilt Siri impresses on complex, personal-data-aware tasks but still misunderstands requests, such as missing dates in messages. - Six Colors

Siri lets Claude do it: Leaked iOS 27 frameworks show a "Model Delegation" feature letting Siri hand off tasks to Claude or ChatGPT, or replace its model entirely. - CNET

The Gates Foundation pledges at least $1 billion over two years to widen AI access, warning the next 12–18 months will decide who benefits. - TNW | Artificial-intelligence

ModelML is in talks to raise over $100 million at a valuation above $1 billion. - Bloomberg

General-purpose AI models can handle about 70% of enterprise tasks, but owning data and workflows still matters most - TechRadar

Business schools rethink core MBA skills as AI, following ChatGPT's 2022 debut, takes over routine coursework tasks. - Bloomberg

Follow the latest AI headlines via druce.ai on Bluesky