Do AI startups build their own AI or something else?
I see a lot of companies with fresh faced 25 year olds that are running all kinds of AI stuff. Everything from NPT, Vision, organization tools.
What I don't get is how they're doing this with such small teams. How are these young companies building out AI services?
For the most part, they’ve just put a thin wrapper on ChatGTP or some other big AI API.
This -- I think this is probably starting to die out already, but there was a massive wave of startups whose "product" basically amounts to a prompt template that they fill in with your input and pass along to ChatGPT. Any early stage startup getting plugged on HN or whatever, you can assume there's at least a 60% chance it's pure smoke and mirrors, and that percentage only climbs the more buzzwords it includes.
Once I started to work with an “AI” startup. After a short time, I learned that it was but “AI” just a bunch of if/then statements in Python.
Times have changed, now it's downloading an existing open source model and selling it back to people
Tbf, the term AI is fairly broad. Very few of these startups specify their use of machine learning
There are lots of things you can do with freely available tools if you have decent training data and the know-how to tailor the application to your use case.
I'd bet 95% are wildly overpromising to grab investment money because it's the hot thing right now though.
There are lots of powerful tools, such as pytorch, which take a lot of the complexity out of creating neural networks. A lot of people are studying this stuff in college now, so tons of young graduates are coming out of school each year with the ability to create these networks. There is also lots of open source code out there where someone has done all the hard work, and you can just train new models on top of it. Not everything needs to be ChatGPT, a lot of kinds of models can be trained relatively quickly and with little resources.
I took a Stanford AI course for free a couple years ago, while sitting at home in Boulder.
That such a thing is even possible is incredible.
There are good opensource models one can use and, if needed, fine-tune. Not every task needs chatGPT power. One can also build applications using openai models (directly by openai or by ms azure).
Google has AI models as a service as well.
It is important to understand that an application on top of openai model is not nessesarly a just chatgpt wrapper.
Yes