Ask HN: The Problem with "AI Startups"?

I think people creating "AI Startups" are looking at things backward. There is this new technology and they are trying to look for ways to use the tech and call it a company. They should be looking for a problem people have that they can solve, if AI makes that solution better, great. Customers aren't buying AI, they buy good solutions to their problems. This is where we're seeing a lot of bad reviews on some of these AI Startups... they are making these AI products that do stuff in areas where adequate or better solutions already exist, so customers don't care.

Don't be so blinded by the tech that you forget what the point of a company is.

I don’t think most of the startups are trying to create anything resembling AGI.

However, on that subject I would still throw out a counter argument. AGI might not be a matter of “throwing infinite money at the problem”. It might be that the building blocks already exist but need to be arranged in the right combination under the right circumstances to create autonomous agents.

I think it's fair to implement complex preprocessing and external integrations (such as memory). But you can actually prototype llms at a tiny scale! A cluster of 3090's (SLI) or 4090's with some kind of splicing (4-8 input(s) are spread across 4 gpus and then merged together). The cost to entry is roughly around $15k.

While you won't be making anything even close to llama1 it is possible to develop technologies that can be used to train llama4.

I've personally been using gh copilot and it's an amazing tool for the 'braindead' parts of my workflow. I usually let it generate a rough template and then fill in the parts by entirely replacing them, but it at times manages to do exactly what I wanted when working in larger files which already have much of the functionality defined and it's rather trivial to figure out the rest.

Copilot does what it says, it's a copilot. It feels like I am iterating over a problem with a 2nd developer suggesting ideas as I code.

Having co-founder DNA in the market is easiest, and second best is being open minded (ex: young and naive, a PhD, ...) and immerse yourself in the customer space. We did the latter, and it took us ~4 years to match our inhouse deep tech with customer problems, and another few years for the market to catch up. And now genAI has put it into overdrive, and our Product #2 (louie.ai) is much easier bc it is now more like Scenario 1.

For related reasons, I'm bullish on growing a consulting arm of most b2b startups to help accelerate this process + figuring out profitable scalable revenue.

Domain expertise and narrow models can definitely be an advantage but you need to be somewhat successful before you can get sufficient data to train or refine a model with sufficient domain expertise to be a differentiator.

You definitely do not need to train your own base model to be successful, but if your entire pipeline is a system prompt or three in a small agent graph... You didn't build a product, you built a hobby tool over the weekend no matter how much UX/UI polish you put on it.

I do think this question, and many start-ups are thinking about how they want that sweet AI money and are starting their business from there rather than from a problem. If you see a problem that is labor intensive and could be done by mechanical turk... Well you're probably on to something and an AI language model can probably solve that problem.

The companies and AI products I think that are going to last either aren't starting with AI they're focusing on a problem and have reasoned their way to AI OR they're doing some deep stealth research into the problem over a long period of time to be able to develop domain-specific data, techniques, and plans so they can fine-tune their own model before ever showing it to a potential customer. You better be sure you have a good plan if you're going to try and be the latter.

Yes I think I agree. You're describing data startups more than "AI" startups (although of course to investors and prospects they are AI STARTUPS).

If you build a good system to collect hard-to-gather, rich proprietary data then improvements in AI will help you squeeze more and more insights out of it.

People think these tools are more capable than they are because marketers encourage that misrepresentation. This leads to a huge glut of scam companies that are nothing more than wrappers on OpenAI's GPT tech, or chaining together tools from cloud providers.

I had the chance to talk with one of the healthcare startups, they're trying to make a product that competes with Dragon DAX Copilot (which is incredible) by cobbling together some of the tools from Azure, Google Cloud Platform, and AWS. I didn't believe their pitches, and their prices were insane, because they got some seed funding and are building an MVP still, while also trying to get some revenue. They shared with me a document that was confidential because anyone with technical know how could see they had nothing unique, they were only shuttling data between a couple cloud providers to utilize different pre-built technologies. When confronted with this they got rather defensive and moderately offensive, and even tried to end-run around IT by going to providers and filling their heads with nonsense.

Right now AI can do 10% of what people promise it can, and it's so new audiences don't have the tools to know who is real and who is a fly by night operator. It's like the beginning of the app store, AI is full of incredibly low effort apps right now and most are garbage. People are rushing in like it's a gold rush, they want some easy money before bigger, more competent products come out.

I'd say this is the biggest issue, too many ethically challenged people are spinning up cheap minimum-viable-product apps and slick marketing materials to take some cash off the table before collapsing. The industry is full of snake oil.

I think this is changing. Previously you had to train something massive off a huge data set. But now it's moving towards having a pipeline of pre-trained models that are trained on massive data sets and then smaller models that you train in house to tweak results from that pipeline. Any start-up should be able to get its hands on enough data to train a LoRA for example. There are good enough open source components to build a moat out of a good pipeline with one or two components in the pipeline trained in-house and the rest pretrained.