Paul Graham

"You've made it abundantly clear that you care more about work than about me, or this family." — a friend's 6 yo daughter as she left on a work trip

Amazon banning agents is the first opportunity I've seen since Amazon was founded for a startup to create an Amazon competitor. People will want agents to buy stuff for them. It will be one of the main use cases. And they won't want to use some Amazon-supplied agent to do it.

"Your restrictions on usage are my opportunity."

Any business that bans agents represents an opportunity for a startup to create a competitor. If they need to ban agents, it must mean people want to use them, or there would be no point. And that in turn means there's demand for a competitor that allows them.

"I feel kind of gross. I ate the same amount as you." — Jessica after going to Gott's

When you talk to successful founders, they always wish they'd hired fewer top managers from outside. I don't think I've ever met a founder who wished he or she had promoted fewer people from within.

The problem is not just that these hired managers are from outside. They're also a different kind of person. As one founder put it, "their loyalty is to their career, not to the company."

Talking to Amjad yesterday, I realized Replit is more important than I'd previously thought. It's the way AI pervades organizations. Organizations can become somewhat aligned with AI by using AI-powered apps, but the deepest change takes the form of code.

@wbhub I take this back. Although the 27-460 is a great movement, the 12-600 seems to have been very good too.

The IWC cal 853, one of the great automatic movements of the golden age.

This is the 853 inside this watch:

It's great to have competitors who are motivated by money. If they're any good, they eventually get it, and then they stop working. And if they aren't good, they aren't good.

Halloween 2014: A very serious 2 yo wants the full bat experience.

One very important thing that even the smartest people seem to have to learn from experience is taste in problems. There are juicy problems that tend to lead to interesting discoveries, and nasty problems that tend not to.

If I had to say in one word what distinguishes the juicy ones, the word might be unity. Nasty problems tend to be collections of random, often extrinsic constraints. Juicy problems have one main thing you're trying to solve

One very important thing that even the smartest people seem to have to learn from experience is taste in problems. There are juicy problems that tend to lead to interesting discoveries, and nasty problem that tend not to.

"You said it would consume all my time and attention, but it really does consume all your time and attention." — friend who started a startup

"You said it would consume all my time and attention, but it really does consume all your time and attention." — friend who started a startup

"There's no such thing as a complete set of tools." — Trevor Blackwell

"You said it would consume all my time and attention, but it really does consume all your time an attention." — friend who started a startup

I've been talking to a YC founder who's 15 years in and still doing big new things. One of the biggest popular misconceptions about startups is that founders have a brilliant idea, then implement it. In fact more often than not the biggest idea comes later.

The mistaken belief that everything depends on the initial idea hoses you in multiple ways. It makes you give up if the initial idea doesn't work. But it also puts too much weight on it: if everything depends on the initial idea, it has to be huge.

A company name that appeals to corporate buyers may not appeal so much to their customers.

@ahistoryinart You can tell this was painted in a cold room.

Don't own the .com? The good news is, there's no such thing. "The .com" is just the only .com you've been able to think of so far.

Apparently some people found this hard to understand. It's intended for startup founders who call their company <name> but don't own <name>.com. I'm saying they should just think of another name instead of e.g. resorting to expedients like try<name>.ai.

In England, students choose just a few subjects (usually 3) to study in the last 2 years of high school, so gender differences in their interests become more visible than in the US.

A friend of mine acknowledges that something is meant to be funny by laughing with 5 has: hahahahaha. Then if he actually found it funny he shifts to a higher pitch on the 6th ha and emits several more verses of has.

I recently met someone who'd printed out all my essays. I'd never seen this before and I was surprised how big they were, like two encyclopedia volumes (for anyone who even knows what an encyclopedia volume is).

"Do not bet against the airplane. The skeptics who say AI can't conjecture, can't theorize, and can't explain are describing last spring's models. Within months, they will be describing nobody. The evidence is already on the table, and those with eyes can see it."

There were a lot of things that only worked because there's a limit to the rate at which humans can operate. We're about to find out what all of them are, as they break.

Average accuracy of vintage watches I own, by brand, in seconds per day, dial up. Zenith 1.9 Omega 2.7 Longines 2.8 Patek Philippe 3.7 Eterna 4.0 Audemars Piguet 5.2 IWC 6.1 Vacheron Constantin 8.9

The sample sizes are small, and to measure accuracy properly you should really test in 5 positions. But Zenith, Omega, and Longines were the three brands that dominated golden age observatory trials. If nothing else, this shows they still regulate well today.

I had ChatGPT rank watches I own by accuracy in sec/day, dial up. Zenith 1.9 Omega 2.7 Longines 2.8 Patek Philippe 3.7 Eterna 4.0 Audemars Piguet 5.2 IWC 6.1 Vacheron Constantin 8.9

The sample sizes are small, and to measure accuracy properly you should really test in 5 positions. But Zenith, Omega, and Longines were the three brands that dominated golden age observatory trials. If nothing else, this shows they still regulate well today.

This explains why till now I'd never explicitly considered the question of what was the first man-made thing to make it into space. We don't talk about that milestone.

AI startups grow really fast. That's why valuations have gotten so high. Not just because AI is the cool new thing, but because so many AI startups have genuinely amazing numbers.

The sentinel hypothesis proposes that teenagers staying up late has evolutionary advantages, because there's always someone awake to notice if your tribe of hunter-gatherers is attacked. You can probably guess who told me about this.

AIs can now ship containers.

Jessica bought me a fabulous watch from my favorite vintage watch shop (Somlo) and conspired with them to have it in the window when I walked by. Of course it was the first one I noticed. But instead of a description, the label said "Happy early birthday Paul." I still didn't get that it was for me, because one of the guys who works there is called Paul. I turned to Jessica and said that he must have a birthday coming up too, but why would they put his watch in the window? Then I noticed she was grinning like the Cheshire cat.

It can be surprisingly powerful not to stream or even record a talk. It might seem that if a talk is happening anyway, there's no cost to streaming it as well. But I think there often is some cost.

It's strange how much easier it gets not to make certain mistakes after making them just once. You'd think a mistake that was easy to avoid making twice wouldn't be too too difficult to avoid making even once, but that's not usually how it works.

The good news is, if some domain has a finite number of possible mistakes and you never make mistakes twice, you have a chance of approaching perfection eventually. As Bohr said, "An expert is a man who has made all the mistakes which can be made in a very narrow field."

The excellent Omega cal 266. You could make a reasonable case that the Omega 30T2 family were the quintessential golden age hand-wound movements. And because they're not rare they're comparatively cheap today. You can get good examples for under $1000.

Someone asked if it's important to live with your cofounder(s). It can help, especially early on. But it's not so critical that you should force yourselves to do it if you don't all enjoy it.

A surprising proportion of the most successful founding teams lived together at some point. But this isn't all proof that living together wins. Some of what it's showing is simply that the founders liked one another, and liking one another definitely wins.

It's surprisingly revealing to ask: if someone's name were a unit of something, what would it be a unit of?

We talked about this over breakfast. We decided Trevor's would be (possibly excessive) generality in an engineering solution, Jessica's would be either curiosity about someone's character or degree of incapacitation when laughing, and mine would be amount of change in an idea.

Sure, there are some kinds of thinking that don't require writing. People could think before writing was invented. But writing is an extraordinarily valuable tool in the hardest kinds of thinking.

I just did office hours with a startup that needed (among other things) a new name. I asked what adjective scientists use to describe the situation their technology is designed to induce. The .com was untaken. If only it was always like this.

Future generations won't believe we let this happen.

One of the most interesting things about startups is that there are exceptions to practically every rule. E.g. it's bad to be a solution in search of a problem. Except .5% of the time it's good. The ultimate test of an investor is to see these exceptions.

When I notice such edge cases, I don't just agree that it's ok for the founders to go against the conventional wisdom. I urge them to, because I know I need to compensate for all the pressure they're going to get in the opposite direction from their other investors.

There is always a .com you can get for $5k that's better the tryblurgh.‌ai you're currently using. The reason you don't see it is (a) lack of imagination and (b) because you're attached to Blurgh, which you shouldn't be, because it's not that great anyway.

One of the most interesting things about startups is that there are exceptions to practically every rule. E.g. it's bad to be a solution in search of a problem. Except .5% of the time it's ideal. The ultimate test of an investor is to see these exceptions.

There is always a .com you can get for $5k that's better the t.co/sAEGzoHNNO you're currently using. The reason you don't see it is (a) lack of imagination and (b) because you're attached to Blurgh, which you shouldn't be, because it's not that great anyway.

Jessica says all my old watches look the same to her. A preposterous idea. But I admit there is a slight resemblance between the two I just got at auction.

Don't start a startup expecting you'll go to college (or back to college) afterward. Those two paths are incompatible, and you're lying to yourself if you try to pretend otherwise.

Though there is of course one case where they're compatible: when the startup fails. Which means to the extent you want to go to college, you'll be implicitly hoping to fail. And startups are so hard that if you want to fail, you surely will.

Our default assumption should be that it will, or has, but we won't hear about it.

I'm reading Shakespeare's Professional Career by Peter Thomson. Like most really good books, it covers much more than its title implies. It's so packed with detail that it's half way to being a history of the Elizabethan period.

If you know anyone looking for a good and fairly AI-proof career, watchmaker (= watch repairer) seems a promising one. Every watchmaker seems to have more work than he can handle, mechanical watches are increasingly popular, and it will be a while before AI does this.

Using "think" to describe what an LLM does reminds me of the 16th century, when astronomers didn't really believe in the heliocentric model, but used it for calculations anyway because the math was simpler. We'll start by calling it thinking, then gradually acknowledge it is.

The AP doesn't understand that this ship sailed years ago. What's the use of a style guide that doesn't understand current educated usage? Usage is the whole point of a style guide.

I looked at my current ChatGPT thread, and just in the last 5 days I've used guess, see, think, infer, and miss to describe things it has said or that I want it to do.

This is a big enough problem that there's not just one startup idea here, but several. For example, there will probably be one variant meant to filter out agents, but another meant to welcome them in an orderly way. This also seems an idea you could build more on.

Paweł Huryn tested models' ability to find bugs planted in code. Cost increases exponentially with performance (note the log scale on the x axis). I had ChatGPT superimpose a trend line. Bargain = above the line.

For some reason not all the points are labelled, but here's the key.

I happen to have read a lot of rock stars' autobiographies (Jessica buys them, and then I pick them up and read them), and the one consistent theme is the drugs. Back in the day, music and drugs were inseparable to a degree people would find hard to grasp today.

A CS undergrad asked me where he could have most effect in the AI age. I said probably at either extreme: either close to the technology, actually making LLMs, or close to the customer, using AI to give them exactly what they want. Or maybe both if you can stretch that far.

Prewar movements were often designed to fit in the rectangular cases that were in fashion then. The war made round watches fashionable (because all military watches were round). So inside 1940s watches you sometimes find this:

An alabaster bowl. When would you guess this was made? About 5000 years ago. The future was even more unevenly distributed in the past.

In the last 2 years, web traffic to news sites has fallen by about 28%.

The Democrats lack all conviction, while the Republicans are full of passionate intensity.

The reason so many people look for an ulterior motive for the AI labs asking to be regulated is that they don't grasp that models could be dangerous. But if you try assuming models are getting dangerous, or at least unpredictable, everything falls into place.

If models are getting dangerous, you want to slow down. But you don't want to slow down unilaterally and be left behind. So you ask the government to make your competitors slow down too. And you can't do that with a straight face without asking to be regulated yourself.

Jessica and I have been watching Professor T, and it's really good. If, like me, you wish Doyle had written more Sherlock Holmes stories, this may be as close as you're going to get.

Something 14 yo told me: You have to overact when doing Shakespeare, because the language is now so hard for audiences to understand.

Remember how eagerly the press swallowed stories about how Twitter was going to crash after Elon bought it? How could they have believed that someone who could run rocket and car companies couldn't run a forum?

It's isn't just to people with ADHD that this happens. Some tasks inherently take a certain amount of time. An appointment that breaks the afternoon into two blocks of time could leave you with no block long enough to complete a task.

In the 1950s, an aluminum company commissioned Vacheron & Constantin to make them some aluminum pocket watches. The case, dial, hands, numerals, and even the plates in the movement are aluminum. The resulting watch is 44 mm in diameter but weighs only 22 g.

Miriam Pawel says more than she realizes: "Students on University of California campuses may struggle with basic math, but they are graduating at record rates."

Notice anything interesting about this watch coming up for auction in a couple days?

I did office hours today with a founder who had so many different things to figure out that we walked for 5 miles. (These were mostly the good kind of problem. He's in the early stage, where problems ≈ deciding what to do in what order.)

In Texas they found 117 potential noncitizens out of 18 million voters. That's .00065%. And that's an an upper bound; these are merely "potential" noncitizens. This is a fake problem.

The problem is even smaller in Texas. 117 potential noncitizens out of 18 million voters = 00065%. And again, that's the upper bound.

Paul Buchheit meme-ranks with Jobs, Zuck, and Elon. (He's the one who came up with "Don't be evil.")

Although you pay for AI by the token, that's not the unit of inference, because you get more problem solving per token as models improve. Someone needs to define the unit, perhaps using a chain of increasingly hard problems, each pair of which can be solved by a single model.

A founder of a startup I funded had these brass plaques made to put up over his and his cofounder's desks. (He's Build Stuff, she's Talk to Users.)

One data point in technology making everyone richer: the cost of lighting has decreased by over 1000x.

Too much screen time may be bad for kids, but I can't wish scenes like this hadn't happened.

Vídeo

Eternas are some of the biggest bargains in vintage watches. This 1963 Eterna-Matic cost $509 at auction + $69 to be serviced = $578 total. It's still accurate to 7 sec/day (dial up).

Making Startups Powerful: paulgraham.com/powerful.html

Dynamic pricing means prices go down as well as up. So this is conclusive evidence that either (a) she's stupid, (b) she's dishonest, or (c) both. The strange thing is, I genuinely can't guess which it is. Could she lie this overtly?