There appears to be a pattern. Unmet need is identified: "I want ChatGPT -- but able to read PDFs" or "I want ChatGPT -- but able to do research and produce lengthy reports." Startup gets funding for this and, if they're lucky, releases a rough beta that leans heavily on the OpenAI API. Two months later OpenAI launches a better, much more polished and seamless version, which is integrated into ChatGPT itself.

I had briefly considered forming a Legal AI startup (going so far as to download the fulltext of every legal ruling ever made in the US -- something like 400GB) but then o3/DeepResearch got so good that it became apparent that there'd be little point.

Steve Hsu claims to have solved hallucinations in a customer service context, which might be the only "startup-type" idea that has a head-start over the giants.

There is a clear business case and buying large trucks is already a capex play. Then slowly work your way through more complex logistic problems from there. But no! The idea to sell was clearly the general problem including small cars that drive children to school through a suburban ice storm with lots of cyclists. Because that's clearly where the money is?

It's the same with AI. The consumer case is clearly there, people are easily impressed by it, and it is a given that consumers would pay to use it in products such as Illustrator, Logic Pro, modelling software etc. Maybe yet another try in radiology image processing, the death trap of startups for many decades now, but where there is obvious potential. But no! We want to design general purpose software -- in general purpose high level languages intended for human consumption! -- not even generating IR directly or running the model itself interactively.

If the technology really was good enough to do this type of work, why not find a specialized area with a few players limited by capex? Perhaps design a new competitive CPU? That's something we already have both specifications and tests for, and should be something a computer could do better than human. If an LLM could do a decent job there, it would easily be a billion dollar business. But no, let's write Python code and web apps!

This makes sense. The entire engineering/tooling field is so gonna change. Picking a winner makes isn't really possible. Most people are just starting to solve real problems with it and starting to build patterns that are not complete nonesense. But it will still change a lot

> “AI for X” verticals are surprisingly narrow.

I think that makes sense too. Those were a significant part of the initial hype. A lot of people promising that they'll take a "generic" LLM (which you all have seen how already smart that is) but now train it specifically on parenting, or trivia, or your emails, or your help center. It's a service type that will continue to exist. Perhaps it needs to tailor to a specific enterprise scenarios to gain traction as a startup. Though the need for these companies to manage the privacy concerns of the customers with their ability to inspect and look at the data and clean it might not be fully solved yet.

> Reducto - Reducto is an AI-driven API that specializes in converting unstructured documents like PDFs and images into structured data.

This is an example of the type of companies where "extracting LLM relevant context from X" and are relevant for any company doing the "AI for X" schtick or enterprise doing AI development on their own. This company is specifically about PDF and images, but we probably gonna see others that are for videos, archives, isos, msoffice docs, and even the ultimate holy grail of "universal binary => very rich structured data" API.

> Developer Tools & Infrastructure

The picks in this category are the most perplexing to me.

This was trickier 18 months ago, but every major LLM provider has solid support for this now. You can just drop an API call to Google, OpenAI, etc. your existing pipeline. What am I missing? Maybe the selling point was batch, but all LLM providers have a batch product now too.

    classification_response = requests.post(
        "https://platform.reducto.ai/extract",
        json={
            "document_url": f"jobid://{job_id}",
            "schema": {
                "type": "object",
                "properties": {
                    "document_type": {"type": "string", "enum": ["W2", "Passport", "Other"]}
                },
                "required": ["document_type"],
            },
        },
        headers=headers,
    )

    a) what are the growth rate when measuring customers which are *not* YC companies?
    b) what is the churn rate for that same group.

It isn't a no-brainer. Many founders in the new era are weighing bootstrapping, seed-strapping, and VC money without a clear answer.

If you can see a path to growing MMs of revenue with little need for staff, you may just not go do a follow-on round.

I was pretty shocked that of 275 companies in the Winter 2023 batch, only 12 have received Series A deals. Granted, I know a huge part of that is that the VC environment has just collapsed due to the end of the ZIRP era, but those numbers at least sound pretty brutal to me.