Worldviews

How useful is AI?

Supporters claim that AGI will change the world completely. Detractors think it will have no impact on anything. I think the truth is very much in the middle.

I have no idea what will happen in 5 years, the error bars are simply too wide. I think there is actually some non-negligible chance we hit something approximating AGI. But there is also a non-negligible chance we do not.

AI has made me 100x more productive in dealing with certain annoyances when I am coding. Maybe overall it makes me 5-10x faster, this figure is actually quite believable. (But keep in mind that I am using it very judiciously and do not have it making more than 20-line diffs at a time.) But if you add in the time it takes for engineers to communicate, I would wager that in the best case, you’d only get a 2-3x improvement. It’s big, but if we need 1000x multiples of productivity to justify our worldviews, it is not enough. And I think there are many worlds we don’t get to 1000x that fast.

And where does all this new productivity go? AI has made you or me more efficient at programming, but has it produced things the general public wants faster? Probably not. I think we are quality and taste constrained, rather than programmer constrained. What good is it to the world writ large that people can code faster, if the things they are coding faster are just AI B2B SAAS startups? More generally, most of the world just does not want AI: there is just no demand for AI-generated audio or images (so Suno and Sora were doomed to fail).¹

At the same time, this is an incredible time to be building things. Your competition is probably narrower than you think, and leaner teams can probably get more out of AI than large behemoths (which seem to be getting negative productivity as a result, see “workslop”).

Who would I hire?

I make no promises this is what other people think about when hiring, but this is the kind of thing I would hope other employers look for too in this day and age!

If you see this far in the future, my views do not reflect whoever my employer is, etc.

If I think we are quality and taste constrained, I would be hiring people who care about these things rather than focusing on technical ability. (Obviously I would not be hiring people without technical skills, but I don’t think we are in short supply of people with technical skills. Nor do I think those are too hard to attain.)

As an example, I would be more impressed by a software engineering resume which listed some personal interests and had some personal projects relating to it solving real problems that real people with that hobby face, as opposed to someone with a perfect GPA or the hardest courses. Obviously having both is ideal, but one of these is far more important than the other! Especially with AI being so good, I think the important thing is identifying problems worth solving, because now the process of coming up with solutions is so much easier.

The other thing I would look for is a keen eye for quality and an obsession with doing things the right way. An example of what I mean, from Evan Chen:

I would be worried hiring someone who types $\sin (\theta)$ as $sin (theta)$.

Actually, I agree with every word of Evan’s post: there are lots of folks who are technically gifted, and there is a surprising lack of correlation between technical skill and taste/obsession with quality. It is pretty easy to check for the first, so I would focus on the second.

(Incidentally, I think there are a lot of companies who are converging on this idea. It holds quite a bit of explanatory strength for the new grad hiring I’ve been seeing!)

US Midterms

I think this election was a referendum on integrity.

Obviously the usual issues of the economy, etc are what most voters describe as salient. But I think the undercurrent is Republicans did not do what they were elected to do. We were promised reduced cost of living and no new wars. The Iran war went against both. We were promised the Epstein files, and so on. (Obviously you don’t take a known liar for his words, but many people did and witnessed the consequences.) The only thing they actually stuck with was illegal immigration, and whatever you feel about immigration, no decent person would support having violent police academy rejects beating and kidnapping people at polling stations and schools.

I’m quite optimistic about the new batch of senators that were elected. Whatever you believe about their policy positions, at least they are demonstrably analytical and have cogent reasons for believing what they believe. (Setting the bar a bit lower, at least they can string together a sentence.)

Markets

THIS IS NOT FINANCIAL ADVICE.

Compared to the average tech person, I am far more bearish on OpenAI/Anthropic and more bullish on Google. I think there are enough worlds where one (if not both) of the companies flounder because they have no real profit engine, vs Google who actually has a profit engine. In particular, Google has far more compute than OpenAI/Anthropic, and they have the ability to sustainably fund their own AI investments a la Google Ads/Search revenue.²

Generally I think there are enough cases where OpenAI/Anthropic that make me bearish on these stocks when they list. In particular, I think Anthropic’s IPO will look very similar to SPCX: it will be oversubscribed at the start and then shoot up, but it will nosedive afterwards and fall below the initial offering price. The price of a stock is not just the linear expectation of the net revenue of that company in the future. Risk has a price, so even if you think there are outcomes where Anthropic invents AGI and 1000x’s their revenue, that does not mean their stock price goes up proportionally.

But beyond that, I think even if Anthropic gets to AGI, they will not be able to dominate the market. Particularly as AI capabilities improve overall, the need for frontier model $N$ as opposed to model $N - 1$ is going to decrease. This summer, Fable 5.0 and Astra 5.6 were the first models to be genuinely useful for ideation, as opposed to just being something people used as a glorified autocomplete or as a homework assistant. So for a brief moment, it really became important to have access to frontier models. But will this be true in the future?

In a few months, Deepseek or some other open model will be just as good as Astra 6.0 is right now. Maybe the quant labs and AI labs will need the next frontier model, but will someone working on internal tools at big tech really need a frontier model in a year? In two? Particularly if you believe security is a losing battle, it is pretty likely that AGI will be aped by some Chinese lab in 6 months. Where is the moat?

Think of who even needs frontier models now. Most consumers don’t. Look at the content farms on Instagram: even as models advance, they have stayed entirely identical throughout. And corporations are scaling back on how much AI they use. There are only a select few companies who can justify reckless AI spend per employee: the companies who make a lot of money per employee. This is essentially just quant shops.

There is also very little political appetite for AI. The probability of some regulations substantially setting AI back in the short term are pretty high (e.g. Texas datacenter regulations), although I think training regulations are going to be quite hard to codify and pass. If anything passes that makes it harder for Anthropic to scale, the portion of their stock valuation explained by the probability of AGI being profitable is going to go even lower.

(All this applies even moreso to OpenAI, but they haven’t IPOed yet. I suspect this is part of the reason why.)

However, I do think AI has changed the world a lot already and has the potential to change the world even more. A company like Google that remains well poised to profit even if they don’t hit AGI soon. (There are concerns about search revenue, and I do think “just put ads in Google search AI responses, bro” is a bit of a lazy copout. I think this + Google AI flounders is the bear case, and it doesn’t seem unreasonable to me.)

Firing employees

You may have heard that a lot of tech companies have started firing because of AI. I think this is largely untrue.

Regardless of what the actual reason for layoffs is, it is economically and politically expedient for companies to claim it is because of AI. It is very bad to admit that you overhired (which a lot of companies did during the COVID era). It is much better to pretend (or to even convince yourself) that you are firing people because you are forward-looking and adopting AI or whatever.

There are some people (even in quant firms or AI labs) who are terrified that AI will take all of their jobs. I view this to be fairly unlikely. Even if AI is at a state where there is a 90% chance that you can replace your rank-and-file with AI, I doubt senior leadership would ever replace them instead of empowering/augmenting them with this AI. 1. In the 10% where AI does not materialize, you have just tanked your entire company. 2. It will not be the case that one day, people wake up and realize “AI can replace everyone.” I think it is more likely that people have some probability in their heads that “right now, AI can effectively replace everyone at my company.” For most people that probability is around 0 right now.

Ergo, the first company to replace everyone with AI will be taking a great risk that I think very few major companies will want to take. _Especially_ the AI firms/quant shops who have spent so much time and effort acquiring talent; the left tail of "the AI sucks _and_ my people are permanently gone" is completely unacceptable.

Furthermore, if you think that some startup or smaller company can pave the way for a bigger company to do the same, I suspect big companies are just fundamentally different. So even if you know small companies can get away with it, it does not entirely eliminate the risk for bigger companies.

Outcomes are more top distributed

This is mostly for CS/math graduates. I don’t know what the world looks like for people in very different fields, and nor do I have many predictions on what the future will look like for them.

Maybe 5 years ago, the vast majority of CMU CS graduates who went into industry were going into companies like Google or Amazon. They were respectable, but almost nobody was getting the insane outcomes you see today. Now, a much smaller portion of people are getting jobs that pay much higher. And outcomes are becoming more stratified as well; either you are graduating and making 150k-225k, 500k-650k, or 1m; basically nobody is making 300k out of college. More importantly, the 500k-1m range is capturing an increasing portion of the outcomes. It is no longer enough to merely be good at what you do; if you are a college student now, I think the most important thing you can do is target excellence.

Furthermore, I think a lot of people are capable of being excellent more often than they realize. If there are 100 people with the potential to do something amazing, in practice I think 90 of them get scared and 7 or 8 of the remaining 10 succeed. So your odds are really good! In fact, it may be correct to condition on success: set an ambitious goal in the medium term, condition on yourself achieving it, and do the things that best explain it happening post-facto.

With this stratification of outcomes, you cannot settle for mediocrity. You have to be ambitious, both in terms of building your own skills and in terms of what you aim to accomplish. The AI can already replace mediocre people, and as AI gets better, this will become increasingly true.³ So you must become more and more ambitious, especially if you are young.

College

We are at an inflection point where it will soon make sense to start dropping out of college/stop going to college entirely. For what it is worth, I don’t think that point is now, but maybe in a few years. With tuitions rising and MOOC-like classes (i.e. any class with more than 30 people) getting increasingly bad because of AI, and a ton of people using AI for pretty much every assignment, colleges may stop conferring the status and legitimacy that they used to. At the very least, the cost-benefit analysis is really bad for most people, and there are genuinely only a handful of companies which make going to university a justifiable decision. I don’t know what the other paths will look like, but I am optimistic we will see more alternative routes besides a 4-year degree.

I have thought for a while that most undergraduate classes are sinfully inefficient for ambitious college students, and you should keep in mind that I go to a university known for having more rigorous CS courses than most. This means the problem is probably even worse than I am making it sound. The problem is essentially one of trust. In a MOOC-like, 1. you cannot trust your students to actually know many things, because the distribution of ability and preparedness has high variance, 2. you cannot trust your students to approach your course in good faith, especially because they all have to be there, 3. and because everyone gets the same treatment (e.g. grading scheme, lectures, HWs, tests), you cannot do anything about these differences.

From an academic perspective, you really are getting very little out of college if you just take these MOOC-likes. The specialized grad classes and research are really where most of the value lie. At the same time, it is probably a good idea to clear out your required classes fast (as fast as possible), because having optionality in the future is so valuable. In particular, this means you should aggressively skip any classes you have the chance to skip, because if you believe the intro MOOC-likes are the best use of your time, you are incredibly bearish on your abilities. And the entire theme of this section is that you shouldn’t be!

It is not about intelligence

Intelligence is not actually the separator of outcomes here; it is courage and (as previously mentioned) taste. I think “I am not intelligent enough” is almost always just a massive copout, and functionally, I think intelligence is trimodal/separated into three strata:

1. Cannot string together basic logical arguments. (Insert your average bigot here who thinks X people are bad because Y book told them so.)
2. Can reasonably understand logical arguments and learn new things.
3. Can self-recurse: is capable of reflecting on and improving their own thinking.

If you are reading this, you are almost certainly in the third strata. You may superficially see that someone is smarter than you. Maybe they even are younger than you! But you are capable of improving your own mental processes and thus can reach the same level with hard work and time. If you squint your eyes and look at the big picture a decade or two from now, you will see that high school olympiad scores, GPA, or your first internship when you were 18 do not really matter.

You are capable of so much more than you think you are, and you just have to optimize towards reaching it. This is the advice I would give to any college freshman or sophomore today.