White House Accord on Superintelligence, Explained

Before the hosts get to policy, they spend a minute on an accident of physics. Elon Musk's Starlink satellites broadcast and receive radio waves, and there are so many of them in orbit that together they behave like a phased array telescope, a radar built from thousands of small receivers acting as one. Pointed at the ground, that array can pick up aircraft that are built to be invisible to ordinary radar (00:19). A system built to sell broadband has become, almost by accident, an instrument for seeing what was meant to stay hidden. That turns out to be the week's real subject: who gets to verify what is true, whether the subject is a fighter jet, a language model, or a headline.

On Tuesday, President Trump brought the leaders of the biggest AI and chip companies into the White House for what cohost David Sacks calls the Bretton Woods of superintelligence, after the 1944 conference that set the postwar financial order (03:30). Many of these executives compete bitterly and some are suing each other, but six frontier model companies left having signed the White House Accord on Superintelligence. Sacks, who helped broker it, says the companies agreed to three things: they accept responsibility for what they build, they submit to external auditors who verify their internal safety controls, and those audit reports go to an independent board committee (04:30). That committee, Sacks says, cannot legally ignore the findings, because doing so would expose the board to losing its directors and officers liability insurance, the policy that protects executives from personal lawsuits (04:54). The agreement is voluntary. The consequences of ignoring it, he argues, are not.

Chamath Palihapitiya, relaying what he heard at the event, says Trump opened his remarks with a single line and let it sit in the room.

Whoever wins superintelligence wins [09:30], Chamath Palihapitiya

The idea, as Chamath frames it, is that there is no referee big enough to split the outcome of this technology fairly among nations, so the only real choice is whether the United States gets there first (09:30). He then flew to Orlando, where his company Eighty Ninety has a standing partnership with the accounting firm Ernst & Young, and found the summit's abstractions turning into a product overnight: an audit infrastructure meant to give boards traceability (what was deployed and why), a map from company policy to actual risk, and a paper trail that can be handed to a regulator, a lawyer, or a customer on demand (11:15).

Not everyone in the room believes the fix is regulation at all. David Friedberg, who sits on a White House science and technology council, argues that open-source, open-weight AI models (versions anyone can download and run without a central company's permission) have already made centralized control close to impossible, because the software exists on local machines in 197 sovereign countries.

I think we will see a massive rise in defense systems. There is no turning off the open source, open weight models or shutting down all the data centers or centralizing command and control [20:18], David Friedberg

His argument is practical rather than ideological: since the technology cannot be switched off, rising cyber risk will force companies and governments to spend more on AI-driven defense, not less, which in turn means the US cannot afford to slow down building the data centers that run it (21:31). Chamath takes that further, framing data centers themselves as a form of national security infrastructure, because the scarce resource is not computing chips but the electricity to run them. Sacks points out that China has been doubling its power grid roughly every decade, while US grid capacity has been essentially flat for 25 years (24:46).

The generation of electrons is actually the critical thing, not necessarily the generation of energy as we've classically thought about it [32:50], Chamath Palihapitiya

As a small case study in markets solving problems without a regulator, Friedberg cites Google's same-day launch of SynthID, a system that embeds an invisible watermark into AI-generated protein sequences and DNA code, so a lab can later verify whether a given biological design was produced by a machine (35:07).

The numbers behind the mood

Sacks then turns the conversation toward the economy, arguing that public pessimism has outrun the data. Second-quarter GDP growth was revised up to 2.2 percent from an initial 1.5 percent, and first-quarter growth was revised up to 2.5 percent; the Atlanta Fed's running estimate for the third quarter sits near 3.7 percent (46:07). August payrolls added 162,000 jobs against a forecast of 55,000, unemployment held at 4.1 percent, and core PCE inflation (the Federal Reserve's preferred inflation gauge, which excludes volatile food and energy prices) came in at 3.0 percent, below the 3.3 percent expected (46:55). He adds a figure meant to rebut the idea that only the wealthy are benefiting: median household income has reached almost $90,000 (47:44).

The poverty rate in the US has fallen to its lowest level in history, ten point two percent [48:17], David Sacks

Friedberg pushes back on the celebration with a warning about interest rates. Short-term Treasury yields have climbed 60 basis points (hundredths of a percentage point) in about a month, the highest level since 2002, which he traces to years of federal deficit spending (49:27). Using ChatGPT to model bank balance sheets, he estimates that of 4,295 FDIC-reporting banks, about 95 are likely to show an equity impairment of more than 20 percent when quarterly reports land on October 30, days before the midterm elections (50:22).

On those midterms, the hosts split from the betting markets. Polymarket, a prediction market where users trade on real-money odds, currently shows about a 64 percent chance Democrats sweep both chambers (61:00). Sacks and Chamath both predict instead that Republicans narrowly hold the Senate while Democrats take a slim House majority, arguing that strong economic data has not yet caught up with sour public sentiment, the same gap Sacks says he misjudged before the 2022 midterms (61:23). Friedberg, meanwhile, widens the lens to the entitlement math ahead: he notes that seizing 100 percent of the net worth of every US billionaire would fund the federal government for only one year, which he takes as proof that no wealth tax alone fixes the deficit (68:44).

A coda on trust

The episode closes on a story that circles back to its opening theme. An Omani co-pilot on a Flydubai flight from Dubai to Tel Aviv stabbed the captain and tried to crash the plane; passengers, including an off-duty pilot, subdued him and helped land the aircraft safely with 174 people aboard (70:07). Friedberg's complaint is about language: the first wire reports and several outlets, he says, called it an altercation or a struggle between pilots, phrasing that implies a mutual dispute rather than a hijacking attempt. The following day's New York Times front page led with investigators seeking a motive and Israeli politicians using the attack for election points, rather than naming it as terrorism (73:17). The paper's communications account later defended its restraint.

Our headlines and live coverage contain all the facts we're able to confirm as we confirm them, unlike activists and partisans, we don't publish speculation or unverified information [76:11], New York Times communications (read aloud by Jason Calacanis)

Sacks and Chamath call the hedging dishonest rather than careful. The thread that ties the episode together is not AI or GDP or an airplane. It is the question of who verifies a claim, and what happens when the verifier flinches: a board that cannot ignore an auditor's report, a watermark baked into a strand of DNA, a newsroom choosing its words the morning after a hijacking. Several of the hosts argue that the AI Accord worked precisely because it built in people and institutions whose job is to check the story against the facts. Whether that same discipline survives contact with a newsroom, or a campaign, is the question they leave open.