What won't change with AGI?
Code will be the first domain of LLM’s to be ‘solved’ at a truly AGI-like level. This is true because of technical reasons (easy to run code and learn from the output) but also because of motivated reasoning (solving code means you can run code to solve math and other harder problems).
Bezos has a great quote:
“I very frequently get the question: 'What's going to change in the next 10 years?' And that is a very interesting question; it's a very common one. I almost never get the question: 'What's not going to change in the next 10 years?' And I submit to you that that second question is actually the more important of the two -- because you can build a business strategy around the things that are stable in time. ... [I]n our retail business, we know that customers want low prices, and I know that's going to be true 10 years from now. They want fast delivery; they want vast selection.”
So forget what’s next, let’s think about what stays the same. First, though, a note on the next few months.
You’re going to see a LOT of apps offering a way better no-code approach towards building small, purpose-built apps. I think this is pretty cool! Claude Artifacts is probably the most mainstream instantiation, and Websim and val.town are two of the more well-known startups, but there’s so many others. The general thesis is basically a better version of any of the no-code app builders, which makes sense, and will be lots of fun to see what people do.
Which begets a hard problem for startup builders: if it’s easy for somebody to build the code for a competitor to you, what ends up being durable? Two big factors: taste and connections.
Taste is taking an opinionated way to solve a problem. You need to develop your own approach and have others recognize it. Sometimes this requires really understanding a very particular niche, and solving that better than a general solution. Maybe it’s just a prettier and faster version of something that already exists.
The thing about taste is that it requires not being static. The outwards manifestations of taste often change but those in the know can recognize mimicries. To take fashion as an example: we went from baggy clothes in the 90s to extremely slim fit in the 2000’s and 2010’s and post-Covid are back to flowy styles. Every trend is a reaction to the previous one, and understanding each trend requires a certain level of dedication to understanding every little cultural signifier.
Connections matter because even if there’s a million copies of the same thing, you can win a deal by just being in the room. Most people just want to solve the problem and move on, so you need to reach them where they are: be the first Google result, run a billboard outside their office, or meet them at a dinner party.
For most people, connections are difficult to bootstrap. This is probably the most valuable part of doing an accelerator like YC, some level of social proof but more importantly meeting some folks who have social currency today, and your batchmates, who may grow up to have it at some point. You can say the same things about going to a good college or MBA program or working at a hot startup early on.
To some extent, these two forces counteract. Taste often manifests outside of the mainstream, and traditional connections demand some level of conformity to existing norms. Being able to pull off both is very rare.
These two elements, taste and connections, also define most of human society. Why do we dress a particular way? Why do we broadcast the music and movies we like? These are social signals of who we are, what our ideals and shared beliefs are. Society is big and there are so many people to connect with (and or sell B2B SaaS to), very few markets are winner-take-all.
Investment banking and VC/PE offer examples for what businesses might look like in the future. IB services are essentially undifferentiated between each bank, the advice and tooling you’ll get from Goldman is the same as at Morgan Stanley, etc. But maybe the Goldman MD goes to the same country club as you or MS gave your failson kid an internship. If two YC startups build basically the same product, how do you choose between them? Maybe you share a run club with one, or you like the color scheme on the landing page of the other.
These firms represent the closest thing we have to apprenticeships. The senior bankers are there because of their ability to win business, through some combination of taste and connections. Junior bankers do spreadsheets and slides, but the good ones are there to learn some semblance of how the senior bankers operate and start to borrow some of their connections. The connections could be lateral, eg if you meet similar aged folks in client companies or other banks you share a deal with, or upwards, eg going to the MD’s Hamptons house.
Within the core IBD departments, desks are usually actually quite small. Big banks have lots of people, of course, but if your business is basically selling vibes to your clients, you don’t actually need tons of folks. Similarly if software offerings become commoditized and differentiate on higher leverage points, companies don’t need to be that big.
In companies that rely heavily on LLM building, most of the junior work will look closer to apprenticeships and learning taste and connections. Certainly you will still need to make your decks and make sure the prompted code runs, but we’ll focus more explicitly on how it integrates and solves a user problem. I would expect the partners to then focus on selling their taste and networking to better sell their products too.
The natural question, of course, is what happens to all the programmer jobs. One framing is to look at a career ladder from Google. In general, more senior engineers are expected to take on more leadership and system design responsibilities, along with reasoning and planning to break tasks down.
I made L7 at a BigCo with about 5 years of experience, depending on how you count, so it’s a little amusing to read the typical expectations. At that point I had already mostly stopped coding - the time turned into leadership planning, cross-team alignment, and mentorship.
The last few years have seen a big push back versus these kinds of skills - lots of middle managers fired, tech CEOs ranting about how you go faster with fewer, etc. It’s weird to believe simultaneously that AI will write all your code and also that you only need leetcode grinders.
So maybe that’s the change to really understand. What happens if you have a glut of engineers with obsoleted shallow skills but not much more? Thought leaders have opined about the impact of AI on customer service, truck drivers, and more, but not that many on software engineers. ‘Learn to code!’ - well, thanks for creating training data, I guess. (bad GitHub code actually messed up the training of most sub GPT4 models on beginner problems).
Very curious what other people think the second order effects will be!
> It’s weird to believe simultaneously that AI will write all your code and also that you only need leetcode grinders.
I think this is extremely true — amusing to watch Silicon Valley shit-talk PM pool girls while they transform into PMs themselves telling AIs what to do/write/build. I've felt a flip in the last 3-5 years where I now feel far more secure in my "fake email job" (all soft skills, intuition, trust) than I would if I spent undergrad becoming a mediocre software engineer 🤷🏻♀️