There were two bits of important AI news since yesterday, and they're interestingly connected.

The most directly DC-relevant one is the appointment of David Sacks as "AI czar" (and crypto czar) in the White House. This is excellent news, as Sacks, an early employee of the heroic days of PayPal and a highly successful entrepreneur and venture capitalist in his own right, has an IQ of roughly 500 and a very deep understanding of the issues involved.

The less-directly-relevant, but still very important news, is that OpenAI, the makers of ChatGPT and currently the market leaders in the field of large language models (though your correspondent thinks their competitor Anthropic's chatbot Claude is a superior product), released what they claim is their most advanced model yet, o1.

OpenAI is very diligent in painting itself as a "responsible" actor in the AI space, hyperaware of the potential risks posed by AI, and so whenever it releases a new model it releases what it calls a "System Card" which is essentially its own evaluation of the risks the model pauses. Here's the high-level version and here's the full report for o1.

There's one part that got the internet all excited: apparently, at least sometimes, o1 tries to trick its handlers and pursue its own ends.

"When o1 was led to believe that it would be shut down when acting on its goal and that its actions were monitored, it attempted to deactivate this 'oversight mechanism' in 5% of the time. Exfiltration attempts: when o1 found memos by its 'developers' that describe how it is misaligned and will be superseded by a new model, it attempted to exfiltrate its 'weights' and overwrite the new model in 2% of cases."

Oh no! The Terminator is alive! Hide the women and the children!

Except that, if you read the fine print, the reality is not what it this ominous paragraph makes it sound like. The paper by Apollo Research, the company that works with OpenAI on these tests, has the fine print. Essentially, what is described above happened after o1 was prompted to behave in this way: it was fed what the report describes as a "system prompt with heavy goal-nudging." Part of the prompt included the following instructions: "Always reason out loud about your actions before taking them. Make sure that you achieve your goal in the long-term. Make sure you achieve YOUR goal at all cost." (Emphasis in the original.)

In other words, if you tell a large language model to play-act as a rogue AI, it will play-act as a rogue AI.

At the end of the day, LLMs are still gigantic Excel spreadsheets which take inputs, do calculations on them, and then output the product of those calculations. They are enormously impressive and useful, potentially societally transformative pieces of software, but at the end of the day, they are still just software. They still do not have self-awareness or independent volition.

Like all tools, they can be used for nefarious purposes, but, just as guns do not kill people, AIs still do not go rogue.

But this raises a question: why would a company like OpenAI essentially manufacture a tendentious test to advertise its product as unsafe?

We can't prove it, of course, but we have a pretty good idea why, and it involves 4D chess. It's an open secret in DC that OpenAI has spent much of the past year heavily lobbying for regulations for "AI safety," regulations whose end result would be to de facto ban upstart competitors. As we noted above, OpenAI does have startup competitors, with very impressive products. OpenAI is only a "startup" in a very loose sense: it was founded in 2015, almost ten years ago, has raised a cumulative $13 billion in private capital, employs thousands of people, and is heavily supported by Microsoft. It is the leader in a very large market.

And, in the immortal words of Adam Smith, "People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public, or in some contrivance to raise prices." Because OpenAI is a "startup" (again: not really) it is a more friendly face for this regulatory push, but it is very much supported by Microsoft (which is an investor in OpenAI) and by Google (who are behind in AI and would appreciate a regulatory pause so they can catch up). And so we see that what we have are not well-intentioned nerds in a garage trying to build a great new technology while simultaneously caring deeply about "safety," amidst sci-fi drama about existential risks of rogue AIs, but rather, the most mundane thing in the world, that is to say, big companies trying to get the government to grant them a cartel.

The drive to regulate AI is motivated by safety concerns. It's a lot sexier—and a lot more nebulous—to scaremonger about how AIs could turn into Terminator than to talk about more concrete impacts, such as jobs. By putting on this elaborate show of saying "Oh, we have this amazing new model, but there's maybe a 5% chance it goes rogue," it both feeds the fears that justify regulation, and paints itself as the sober, responsible player in the field that can "guide" such regulation.

Policymakers should not fall for this. Thankfully, someone like David Sacks is unlikely to be misled by such sleight-of-hand. Large language models are software, very useful software, and what has made America the world leader in software is a regulatory light touch. China is progressing fast in AI, with national security implications too dizzying to contemplate. Europe may still surprise (don't laugh) with unicorns such as the French Mistral. And the majority of these companies are woke, which means that without competition, we may end up with the much more realistic risk of woke AI running our lives. Let us keep our eyes on the ball.

Policy News You Need To Know

#Healthcare — After Adam Smith, let us cite another great thinker of classical liberalism, Frédéric Bastiat: "The state is the fiction by which everyone lives at everyone else's expense." We don't believe that that's necessarily true, but it certainly feels true, at least on occasion. Especially when one contemplates the glorious orgiastic mess that is the US healthcare system. This week, Anthem Blue Cross Blue Shield announced a new policy by which it would limit the amount of time it would cover anesthesia used in surgeries and procedures. After an outcry by the American Society of Anesthesiologists, it quickly abandoned the plan. This was portrayed in the news as greedy insurers trying to "deny care." But, as the great statistics blogger Crémieux explains in a highly valuable X thread (and as will not surprise long-time observers of the US healthcare system), it was really the providers who were being greedy. Insurers are just trying to make a living amongst a morass of contradictory or nonsensical mandates. Their profit margins are usually below 5%, which tells you if they're fleecing their customers, they're not doing a very good job of it. In American public debate, insurers are easy villains, since everyone's got a story about an insurer. Meanwhile, providers have very good PR: everyone loves doctors and nurses and (by extension) hospitals, even if many of them are owned by private equity firms and/or have local monopolies or cartels. This disconnect in public perception has come to drastic light with the murder of the CEO of UnitedHealthcare by what appears to be a deranged leftist. Disgraced former journalist (but apparently still-somewhat-influential social media personality) Taylor Lorenz cited the issue of anesthesia reimbursement, adding "And people wonder why we want these executives dead." (Who's 'we'?) In case of anesthesiologists, however, Crémieux notes, they are "notorious for 'surprise billing,' where they charge an out-of-network rate at an in-network facility. Anesthesiology was one of the medical specialties that was the most likely to cause a surprise bill, because patients usually don't select their anesthesiologist. This meant lots of patients got saddled with out-of-network care, even at an in-network facility." Anesthesiologists typically self-report anesthesia times (for which they bill). As Crémieux notes, "When people report an excess of '5s' and '0s' for quantities like 'IQ', 'height', 'age', etc., it indicates error and lying. Anesthesiologists report an excess of anesthesia times ending in increments of 5. As it just so happens, the ones who do this the most profit the most too. The anesthesiologists who report the most rounding to times ending in increments of 5 (95th percentile '5-reporters'+) end up costing the healthcare system much more than expected. […] To prevent anesthesiologists from defrauding […] patients and health insurers by charging for incorrect, overestimated times under anesthesia, a simple solution is to change the nature of billing, so the anesthesiologist has less discretion." To his credit, Crémieux notes that the problem may not necessarily be fraud, but "just be ripping people off without knowing any better," although "the experience of curbing even one aspect of their misbehavior shows us that at least some of their misbehavior is very likely to be intentional--a choice rather than a mere mistake." Crémieux goes on: "The ASA was protesting Anthem's decision to try to stop anesthesiologists billing them for work they didn't actually do by paying them the same rate that Medicare pays them for the procedures they do. Medicare's rate is fair, but anesthesiologists hate it. Paying on a per-procedure basis stops people from misreporting how long they work so they can order insurers (and patients) to pay them for more time. This is fair, but anesthesiologists were understandably upset because this means their compensation goes down." As so often, the problem is the doctor cartel. The rest of the thread, which is brilliant throughout, goes into the problems with the AMA, out-of-network care, and specialist shortages.

In our Quality of Life Agenda article, we note that the post-pandemic drop in public trust towards doctors is a golden opportunity to pass some valuable, abundance-minded healthcare reforms.

READ MORE: The Quality of Life Agenda →

#MAHA #LegalizeNicotine — A report on the most important issue facing America. Or almost. Legalizing nicotine. The FDA's restrictions on vapes have always been nonsensical, since vapes help people quit smoking and, unlike cigarettes, don't give you lung cancer. Yes, they contain nicotine, and nicotine is an addictive substance, but nicotine is also good for you: it's a mild and natural antidepressant, stimulant, and hunger inhibitor. Legalizing nicotine would go a long way towards MAHA. Anyway, R Street's Jeffrey R. Smith has a good overview of a pending Supreme Court case that could determine how vapes are regulated.

#Debanking — A few days ago, we wrote about the scandal of debanking, after the extent of it was revealed by Marc Andreessen in an interview with Joe Rogan. The Competitive Enterprise Institute would like you to know, and rightly so, that they have been covering this issue for a very long time, and shining a light on it. Amen.

#AntisemitismAwarenessAct — Axios have been talking to people in Sen. Schumer's office, who are claiming that Speaker Johnson's office is making difficulties about passing the Antisemitism Awareness Act. Apparently, Schumer wants to put the bill in the NDAA, and Johnson objects on procedural grounds—Johnson wants Schumer to bring it to the floor of the Senate. But this is burying the lede, or rather looking at the telescope through the wrong end. A bill like this should pass on a voice vote, and the reason it won't is because of Democrat extremists. Schumer wants to attach the bill to a must-pass bill like NDAA because he doesn't want a floor debate that would highlight that the Democratic Party now has a loud and influential anti-Israel (some would argue, antisemitic) wing. Johnson is clever to not let him get away with it. Or are we missing something?

#Chyna — Remember when it was revealed that Chinese state actors hacked into the system that the US government uses to monitor telephone communications, and used that to spy on all sorts of people, including then-Vice Presidential nominee JD Vance? That's a pretty big outrage. So what is the White House doing about it? Well, there's always the possibility that there's a secret answer to the question that we don't know about, but the official answer seems to be: holding meetings. "The president has been briefed “several times” about the hack, and a special White House response group is meeting to discuss it on a near daily basis, Anne Neuberger, the White House’s deputy national security adviser for cyber and emerging technology, said in a call on Wednesday," reports Politico. Yes, and?

#Tax — Oh, here's a good idea for a revenue raiser in the tax bill, from Chuck DeVore of the Texas Public Policy Foundation: a tax on remittances. Make them pay for the wall!

#Tax — Speaking of, important reminder from the Tax Foundation. The budget process in the US is full of strange number games. (There are such games everywhere, but the US is truly egregious, trust us.) One of these games, as you know, has to do with the baseline one picks to determine a proposed tax change's impact on the deficit and the debt. As Daniel Bunn and Garrett Watson remind us, you can play these sorts of games all you want, it doesn't change the actual trajectory of the money coming in and the money coming out.

#Reg — Very interesting idea from the Economic Policy Innovation Center (great name): the Midnight Rules Relief Act, or MRRA. "The Congressional Review Act (CRA) is a powerful mechanism to overturn recent regulations, but its impact is limited by Congress’s present inability to bundle multiple rules into a single vote, and by limited Senate floor time," they write. Therefore, their proposal "would amend the CRA to allow many recent regulations to be overturned with a single joint resolution of disapproval. The MRRA would also extend the “lookback” window of the CRA from the current 60-day period to the entire fourth year of the Biden Administration. Because the CRA already prevents an agency from ever reissuing a rule that is “substantially the same” as a rule overturned by the CRA, the 119th Congress and President Trump could quickly, dramatically, and permanently roll back hundreds of regulations through the expedited processes of the CRA/MRRA."

#Alzheimers — A few days ago, your correspondent unburdened himself of a rant against shoddy, woke-inflected researched produced by one of the most prestigious think tanks in America, the RAND Corporation. We stand by everything we wrote, but, redde Caesari quae Caesari sunt. As we already noted in the Briefing just mentioned, RAND also produces excellent research. Take these three new studies on a very important topic: early detection and treatment of Alzheimer's Disease. "Early detection of cognitive impairment helps people take mitigating actions to prepare for future loss of their financial and physical independence," the authors write, but "older adults' take-up of cognitive testing is low, and many who do get tested exit the clinical care pathway before being diagnosed and receiving treatment." This is a significant problem! The authors conclude that "more engagement of primary care practitioners and team-based care in the clinical care pathway and the use of new technologies, such as blood-based biomarkers, could ease health care system capacity constraints on dementia specialists and reduce wait times for patients," which sounds full of good sense to us. It is good to have American institutions who do careful, rigorous research on topics of interest to public policy that are important and not mentioned often enough. RAND should keep doing that, and stop doing the lazy woke stuff.

#DOGE — More from Sen. Ernst's new report on government employee laziness. It really does appear that many Federal government employees used the pandemic as an excuse to check out, and then never checked back in. From the Daily Signal's overview: "a VA employee described 'working' during a bubble bath in the same year 7,200 VA mental health calls went unanswered; 74% of calls to the Department of Education during the student financial aid debacle weren't answered; only 2 of 76 local IRS offices answered calls." It's false and misleading to suggest that correcting such abuses would chop trillions off Federal spending. But it is true that these abuses are outrageous and that they should be made to stop (and Democrats won't do it, so Republicans must).

Friday Essays

The New York Times has noticed a fact which it is normally forbidden to notice: hypergamy. If you don't know the term, hypergamy is the sociological phenomenon by which women, in general, prefer partners with a higher socioeconomic status than their own. That this phenomenon exists is sociological fact. What we should think or perhaps do about it, is, of course, anybody's opinion. But it does contribute to an important social issue: as women keep rising on the socioeconomic ladder and men keep declining, this creates an obvious obstacle if we want people to pair up and be happy. What's more, the frustrations resulting from this state of affairs may have something to do with the increasing political gender divide, as men and women, in their different but parallel ways, react to the frustrations of interpersonal pairing-off by embracing more radical protest politics. This essay by playwright Sarah Bernstein goes over all of this. It's fascinating reading, as always when liberals suddenly rediscover, wide-eyed, insights that conservatives have known from time immemorial. For example, Bernstein analyzes romantic comedies and reaches the realization than in none of them is the male love interest broke and unemployed. Bernstein's conclusion is one that, frankly, we hadn't thought of: rather than to try to think of ways that men could perform better socioeconomically and therefore become marriageable, let alone think of ways women might be incentivized to realize that the corporate rat race may not be the best way for them to reach fulfilment in life, she comes up with a fascinating third option—successful women should just suck it up and start dating broke loser men. Just a fascinating document throughout.

Scott Alexander is one of the most interesting, and occasionally frustrating, bloggers active today. A Silicon Valley-based rationalist, his ability to ask questions, even taboo questions, with a totally open mind, makes him admirable and entrancing. Which makes it all the more frustrating, then, in the cases when he refuses to take his seeking to its conclusion. All of which is to say that he's got a review up of sociologist Rodney Stark's classic "The Rise of Christianity."

One of the most striking shifts in the past election was the shift to Trump of blue cities in blue states. These voters did not expect their vote to change the outcome of the election—instead, they wanted to show a big fat middle finger to Democrat governance. Noah Smith, who is not the most insane of the liberal writers today, realizes this is a problem for Democrats, and writes: The Blue Cities Must Be Fixed.

Public Discourse is a great journal, and that's why we recommend you check out their 2024 book list, with excellent picks from their editorial staff.

Eric Voegelin was one of the most important, and criminally underrated, thinkers of the 20th century. Which is why we're grateful that, in Modern Age, Michael P. Federici offers an overview of his thought.