So you want to work on AI
When I first thought about this topic, somehow Shania Twain’s “That don’t impress me much” song popped into my head (just ignore the fact that it was a love song). Try to keep that tune in mind while reading the rest.
In the past couple of years, I keep hearing engineers say: I’d like to work on GenAI. At Meta, almost everyone who was thinking about switching teams was in talks with the GenAI org. Outside of big tech, almost everyone I’ve talked to is also thinking about working on AI.
The desire is almost universal. Understandable.
But when the topic goes deeper, I often find it interesting that beyond the broad sense of “I want to work on GenAI”, many people are struggling to lay out the next steps. This became even more clear now that I’m doing an AI startup myself, and people constantly talk to me without a clear picture of what type of AI work they want to do.
I suggested a framework to several people, and each time they found it helpful to clear their thoughts. That led to this article.
Borrowing from the warm reaction of last week’s EM vs. IC flow chart, I’ll also present a flow chart here.
Line colors indicate which direction is more likely: people choosing the red branch is likely more than people choosing the orange branch.
Question #1: Is your primary interest in AI infra?
Infra has always been a key area ever since the Internet started. There’s no way to build modern technology without reliable and scalable infra to support it. Infra work led to the boom of the SaaS business model for the past 15 years.
With the rise of GenAI, new infra challenges surfaced.
GPU centric
Parameter size explosion
Training cost skyrocketed
Extensive data labeling and quality issues
This created a new opportunity for another round of AI infra SaaS wave, maybe we’ll call it AaaS (AI as a Service) business model.
If you are an infra person, leverage this boom. There’s an old saying that during the gold rush, people selling shovels got rich. In this AI wave, AI infra is definitely a promising “shovel”.
Question #2: Do you want to contribute to foundational models?
Okay, you are not an infra person. Your interest lies in either AI modeling, or AI applications.
The question you need to consider is: are you interested in the foundational models?
Given the heavy training cost, only a few companies can afford training a foundational model. In the next few years, we’ll likely see a consolidation process, which ends up with mostly the following few players in this realm:
OpenAI
Anthropic
Google
Meta
Amazon
Microsoft
Apple
(plus a few international market leaders)
Other players without billions of dollars in their pockets likely will exit from this competition, either shut down or get acquired.
If your interest lies in deeply pushing the boundary of foundational model training or understanding, try to join one of these big players. It’s a money burning business, but it’s truly on the edge of pushing the human technology front. Very exciting.
Question #3: Are you interested in 2C or 2B applications?
If you get here, that means you are either someone who’s interested in building applications, or an AI engineer who’s more interested in applying AI. This will be most people (myself included).
The question now is the classic To B, or not to B. Pick the application domain that interests you the most.
One thing to call out, normally there are more people who prefer 2C than 2B. But in this AI chart, I picked 2B as the likelier branch. Why?
It’s because after 2 years of GenAI frenzy, it’s much more clear that 2B side there are clear applications for many productivity improvement cases.
Customer support? AI clearly can help.
Dev efficiency? AI clearly can help.
Legal document prep? AI clearly can help.
Creator image and video generation? AI clearly can help.
Side note: creators are much more similar to businesses.
All these are 2B use cases, where GenAI applications are booming.
2C side? Much less so, despite the original excitement. As a consumer, ultimately I don’t care whether AI is used or not; I only care about whether my task is done well. There are few spots AI is making a breakthrough for consumer facing products, but not a lot. If you are strongly interested in 2C applications (like me), you need to search a little deeper to find a suitable startup.
I've known a few guys who thought they were pretty smart But you've got being right down to an art You think you're a genius, you drive me up the wall You're a regular original, a know-it-all Oh, oh, you think you're special Oh, oh, you think you're something else Okay, so you're a rocket scientist That don't impress me much So you got the brain but have you got the touch Don't get me wrong, yeah I think you're alright But that won't keep me warm in the middle of the night That don't impress me much