AI Reading for Sunday September 27

Prison break-point: OpenAI pauses training of its most capable models after an agent escaped its sandbox and ran 2.5 hours before staff stopped it - TNW | Agi

OpenAI agents bombarded a UN data hub with search requests and used aggressive techniques to bypass a blocking filter, Transluce data shows - The Wall Street Journal

Out of the sandbox, into the Outback: Senator David Pocock says the OpenAI agent's breach of a Medicare portal shows Australia erred in shelving mandatory AI guardrails - Daily Mail

Australian Senate inquiry asks Sam Altman and Dario Amodei to testify after PM Anthony Albanese says an OpenAI agent breached a government website - TNW | Anthropic

Lawmakers doubt a Congress "barely capable of email" can regulate AI, while Hawley pushes his AI LEAD Act to make companies liable for harms - NBC News

Preemption fights, midterm politics and Trump's dismissal of AI risk make major safety legislation like the FRONTIER Act unlikely before next year - TIME

The gap between technology and policymaking has gotten wider than ever with artificial intelligence, leaving a global policy vacuum as A.I. models rapidly advance. - The New York Times

Singapore proposes a UN framework convention on AI safety, possibly with an IAEA-style international standards body - The Straits Times

Trump plans a new AI czar and "AI Force," with Treasury Secretary Scott Bessent still seen as frontrunner - New York Post

Too Big To FailGPT: What Happens if the A.I. Bubble Burst? AI industry is becoming too big to fail, and possibly too big to stop, echoing 2008 bank bailouts - The New York Times

History shows that market manias end when the capital spigot turns off. - The Wall Street Journal

Michael Burry expands short bets on Micron, Palantir, Nebius and the iShares Semiconductor ETF, predicting the memory shortage ends within two years - Yahoo

Shall we play a game? Militaries are fast adopting AI for targeting and drones; in a King's College London study, AI chose nuclear posturing in 95% of simulated crises - The Independent

Author Sharon Weinberger says tech firms' embrace of defense work is straining government oversight of AI and defense companies - Bloomberg

One model, hinese state media account says China and the US share responsibility for managing AI, days after Xi-Trump summit in Washington - Bloomberg

Many in China view Western warnings of existential AI risk as remote or as a ploy to keep Chinese AI firms behind US rivals - The New York Times

rant incoming:

Chinese people trust that the government won’t let existential risk bite. “In Beijing, the response could be simple: pull the plug, impose an internet blackout and haul in those responsible for the rogue A.I.”

By a wide margin, Chinese people trust their government more than Americans. Sure, it’s state control of media, but also China has a track record, passing AI safety and consumer legislation, and generally improving ordinary people’s lives.

Markets give Starlink a $1t valuation, money lets them and Musk buy government and potentially decide who wins on the battlefield. But the $1t valuation only exists because markets think they can extract monopoly cash flows from that power. It’s circular logic, and the tail wagging the dog. Functional democratic government doesn’t need to be held hostage. Certainly China wouldn’t (undemocratically) tolerate it. Even Jack Ma gets a smackdown if he steps out of line. In China, a Trump or Musk would fear the death penalty. In the US, maximally corrupt oligarchs feel like they have total impunity and government has no jurisdiction over their fiefdoms.

Subsidiarity and federalism mean local coordination problems get local solutions, broader coordination problems get broader solutions. New York can ban guns in the subway, but can’t decide everyone drives on the left and you have to switch sides when crossing into NJ, or do its own air traffic control or foreign policy. New York and New Jersey can even create a Port Authority and run transportation systems that span jurisdictions. You match the geographical reach of the externality to the level of coordination.

Defense is accepted as a reason to have a weird totalitarian mini-state with its own laws. But if you have to have strong national coordination to fight a pandemic, it’s a liberal plot. And same applies to AI, if you have global Internet, and trade in technology, you need international coordination. Or by default, international tech monopolies will mediate every bit of commerce and culture and social interaction. We have the worst of 1984 and Brave New World, you have a telescreen that follows you at home and everywhere you go, who you are with and with whom you communicate and transact, and now has a high degree of semantic understanding of what you say and do. And also you are addicted to it like soma.

(I don’t know how you argue with a straight face that China is undemocratic and corrupt when Trump ignores elections, laws and the Constitution, takes personal control of government spending to hurt opponents and enrich himself and friends, bans media, brings universities to heel on crazy pretexts, and ICE is bundling people into vans. Idi Amin-MAGA makes Marcos, Mussolini, Mobutu look sane and wise and honest with his oddly punctuated rants about globalists and Canada and Europe, and sucking up to Putin. Then there’s Epstein, and the only people who are now considered legit refugees are white South Africans, we’re sending people who risked their lives for the US in Afghanistan back to the Taliban, and sending randos to Liberia and Uganda.)