Sam Altman || X Chat

INSIDE NYT’S HOAX FACTORY Five months ago, five New York Times reporters were dispatched to create a story about my supposed conflicts of interest working as the White House AI & Crypto Czar. Through a series of “fact checks” they revealed their accusations, which we debunked in detail. (Not surprisingly the published article included only bits and pieces of our responses.) Their accusations ranged from a fabricated dinner with a leading tech CEO, to nonexistent promises of access to the President, to baseless claims of influencing defense contracts. Every time we would prove an accusation false, NYT pivoted to the next allegation. This is why the story has dragged on for five months. Today they evidently just threw up their hands and published this nothing burger. Anyone who reads the story carefully can see that they strung together a bunch of anecdotes that don’t support the headline. And of course, that was the whole point. At no point in their constant goalpost-shifting was NYT willing to update the premise of their story to accept that I have no conflicts of interest to uncover. As it became clear that NYT wasn’t interested in writing a fair story, I hired the law firm Clare Locke, which specializes in defamation law. I’m attaching Clare Locke’s letter to NYT so readers have full context on our interactions with NYT’s reporters over the past several months. Once you read the letter, it becomes very clear how NYT willfully mischaracterized or ignored the facts to support their bogus narrative.

METR (50% accuracy): GPT-5.1-Codex-Max = 2 hours, 42 minutes This is 25 minutes longer than GPT-5.

GPT-5.1 (Thinking High) is about 300 times cheaper per task than o3-preview (Low) while scoring only a few points lower on ARC-AGI-1. 1 year later intelligence has gotten 300 times cheaper. This is why I can’t stand people who say “wahh the models too expensive” it will become cheaper.

Some thoughts on the whole 'OpenAI loan guarantee" situation. 1. First, for context: this issue began a few days ago when openai CFO Sarah Friar publicly floated the idea of the federal government providing a loan guarantee for the development of ai data centers. 2. I, and many others, objected. I objected because of the political economy/regulatory capture implications. Imagine that the federal government made a loan guarantee to OpenAI. Now, OpenAI's financial health is tied up with the government's balance sheet; if OpenAI goes under, the government has a big bill to pay. But what if a new, better competitor to OpenAI emerges? Abstractly, we, as consumers and society, want this new and better competitor to thrive, even if it is bad for OpenAI's financial health. But the government, now, has an incentive for this new upstart company not to succeed. This is the classic reason to disfavor loan guarantees, government equity stakes, etc. 3. In an entirely separate conversation with Tyler Cowen, Sam Altman suggested that government might provide an insurance backstop for liabilities incurred after a catastrophic AI failure or misuse scenario. Ultimately, all catastrophic risks beyond a certain scale are backstopped by the government, but in some cases we formalize this implicit reality. A good example is the nuclear power industry, which has a federally-backed insurance program to protect against the risk of a plant meltdown. In exchange for strict safety regulations, in essence, the nuclear power industry gets a formal federal backstop for meltdown risks. There are merits and demerits to this idea, but it's not a crazy one to consider for advanced AI. 4. In an, again, entirely separate public interest comment submitted to the White House (downstream of a request for information that, incidentally, I drafted while I was in government) late last month, OpenAI discussed broadly the notion of reducing the cost of capital for manufacturers in the AI data center supply chain. We already do this for semiconductor manufacturing through the CHIPS Act. 5. Lowering the cost of capital for manufacturers of strategic goods is not at all a "loan guarantee." Consider natural gas turbines. That industry has gone through brutal boom and bust cycles in recent decades. If you run a natural gas turbine manufacturer, or are a long-term investor in one, or loan money to such firms, you are going to be weary of too much expansion for fear that the AI bubble will pop. This slows down supply expansion for a good that we really do need to power AI in the near term. So what do you do? 6. Well, one thing you could do is have the federal government serve as buyer of last resort of future turbines. You write a contract that says "if the manufacturer makes X turbines over the next five years, the federal will pay Y price for Z number of turbines if no other private-sector buyer emerges at or above price Y." That way, the manufacturer can go to its investors and lenders and say, "don't worry, we've got a buyer for turbines if we expand." And perhaps the lender is willing to offer the manufacturer a lower rate of interest--a lower cost of capital. I myself advocated for precisely this policy when I worked for the Trump Administration (though it didn't make it into the AI Action Plan, sadly). There are many, similar schemes one could imagine. 7. This idea involves the government taking limited, pre-defined risk. The political economy problems with this are non-zero, but they are far smaller than the regulatory capture that would ensue from the US government guaranteeing untold billions of OpenAI debt. 8. As I read OpenAI's public interest comment, I interpret them to have been talking much more about the kind of thing I describe in item (6) rather than the loan guarantee for OpenAI debt. They are referring them to manufacturer cost of capital in that comment; I don't think OpenAI refers to itself as a "manufacturer." 9. I absolutely do not support open-ended guarantees of frontier AI lab debt. I absolute do support targeted industrial strategy to lower manufacturer cost of capital if it (a) exposes the government only to narrow, pre-defined financial risk and (b) seems likely to yield tangible and durable beneficial assets for the American people (in the case of my example, natural gas turbines to make electricity, which is useful beyond AI and which we need much more of regardless of AI).