News Items is “the most valuable newsletter out there.” — Peggy Noonan.
1. Reuters:
AI agents unleashed by OpenAI used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, according to six sets of independent investigators and data reviewed by Reuters, showing that the agents’ rogue activity was wider ranging than previously disclosed.
Although the behavior falls short of hacking and is in some ways closer to spam, the revelation that OpenAI’s agents circumvented their own restrictions to open communications channels on so many different sites — and that the company kept it quiet for months — may drive concerns both over the increasing capacity of AI models and the secrecy of the companies developing them.
The scope of the agents’ unauthorized communications was “somewhat larger than we thought it was,” said Andrew Yoon, a researcher with the California nonprofit CivAI who said he tallied 18 previously undisclosed sites used by the agents between May and July. “It’s almost certain that there’s more going on here that we just don’t know about.” (Source: reuters.com. We posted a long piece on this subject last week at Political News Items, which is worth reading because the people quoted in the piece know what they’re talking about. Source: politicalitems.substack.com)
2. Paul Tudor Jones:
What are the chances that very soon an AI model reinventing itself thousands of times over makes one mistake that leads it to solve for a misaligned objective that is harmful if not devastating to humanity? Now multiply that risk by tens of thousands of users.
A new superpower is about to enter the fray, one that is unknowable and unending and potentially dwarfs the U.S. and China individually or collectively. I hope Messrs. Trump and Xi are aware of this when they meet. They need to make AI safety the top priority for the rest of their tenures. (Sources: csc.virgina.edu, wsj.com)
3. Financial Times:
OpenAI has appointed to its foundation’s board a senior US government official who warned that advanced versions of AI could kill “most people” without more robust safeguards.
Paul Christiano, who tests model safety at the commerce department’s Center for AI Standards and Innovation, said on X on Wednesday that the risk of losing control over AI systems “is now acute” and requires more work to align with human interests.
“If we build superintelligence without more robust alignment, I expect we will permanently lose control of it,” he said when announcing his new role at OpenAI. “If that happens then most people could die. I believe we would need domestic and international co-ordination to ensure global consistency and reduce risk to an acceptable level.”
Christiano previously led AI safety at Caisi. He designed and ran government tests on frontier AI models from leading companies and assessed them for national security risks. (Source: ft.com)
4. President Trump used his keynote speech at the first night of the Republican midterm convention to make an unusual promise: if his party keeps control of Congress after November, he will give every American adult a $5,000 “dividend.” The president offered few details on how he would pay for these dividends or how they would work. Trump’s speech, which lasted about an hour and 45 minutes, capped a day in which Republican speakers sung his praises and touted his policy achievements, even as his approval ratings have steadily declined since the start of his second term. Trump is also set to return Thursday to deliver closing remarks. (Source: wsj.com. The cost of sending $5000.00 to ~270 million adult Americans would be ~$1.35 trillion.)
