An AI Engineer Explains Why Most AI Startups Will Die

There’s many reasons to be pessimistic about AI startups (chiefly about just how incremental the technology is for companies with already existing user bases) but ‘most startups will fail’ is saying water is wet and this one reads like the sour grapes from someone who had most of their field take a sudden turn to the right superseding most of the taught state of the art with something new.

Really low quality article that skips the low hanging fruit on why it’s particularly tough on startups right now for generic ramblings about patterns from the past that do barely apply.

The big change we see in the field right now is that AI has become accessible to a much broader base of by trade software engineers and move from the realm of labs and science to applied.

That did not happen before at scale.

Until the technology plateaus and stabilizes, things will be high risk for companies without existing user bass. That’s the nature of start ups.

And also, while people are hating on the Thin Layer Over ChatGPT crowd - there’s nothing fundamentally wrong with that, you can make a quick buck and sell to a slower company and that’s just fine. Most of local retail these days is a thin layer over Aliexpress or Taobao, arbitraging on time to delivery. Nothing wrong with that once people get over the idea that somehow everything has to be groundbreaking.

There are no moats in this tech. The moat is in the data (look forward to exclusive licensing) and user base, which don’t favor start ups.

Is Langchain a 100M idea? I don’t know, not my place to judge but you know what - people paid more for obviously fake crypto bullshit, that’s between them and their maker, rich people games. I’m glad for the engineers to get paid and the open source community getting contributions, however they are paid for.

A lot of so-called "AI Startups" today amount to nothing more than a fancy front-end for OpenAI's LLM APIs (and similar setups). Those startups are going to die because their only differentiator is a prompt to an otherwise opaque 3rd-party API which they don't really control. Their secret sauce can be trivially replicated without much effort by competitors. Also - they don't really control the models themselves, so they can't do "proper" science when building the product - it becomes an entirely empirical process, subject to the full control of your competitors.

That was a really bad explanation of supervised, unsupervised and reinforced learning.

most startups die so not really a hot take..

I agree with this article under the premises of existing Generative AI, but it fails to consider a rapid advancement in reasoning and planning. If these capabilities are drastically enhanced and according to Elon Musk, AGI is achieved in ~ 2029, then its predictive abilities will be superior to humans. Frankly, I suspect while his conclusion is true, most companies will fail, in-fact the premises are the opposite case and the redundant task AI will be displaced by more general agents able to perform these functions with ease and much more.

TL/DR:

> So, if you are removing repetitive, predictable administrative tasks, that's a good use for generative AI. But if you are trying to create something that requires predicting something that's going to happen in the future? How does that technology work? That's the shiny new thing. And I would not work for, or invest in, that company today.