Hacker News Reader

Comment by codingdave | Hacker News Reader

This is not the first story I've heard of a product that isn't quite right, but when deciding where to pivot, fell into AI as the answer and then failed. I've said it before, but it bears repeating - when we figure out what AI is good at and what it is not, we will start treating it as a tool instead of a goal, and our product ideation will stabilize.

View on HN

> when we figure out what AI is good at and what it is not

We know. It's good at making stuff up. It's bad at facts or for generating things that withstand scrutiny.

What we don’t understand yet is what that means. We have never had such a thing before and it’s not clear what the patterns are that will make the best use of its known attributes with the least exposure to downsides — whether those be related to the tools, the end users, or the business stakeholders, or any other parties. Volumes more nuance to understand, and we’re just getting started.

The product itself is fine. I was part of a startup that was happy to pay for said thing and got a lot of value from it (sdk from OpenAPI spec). So I don’t think product market fit was the problem here.

Are you saying AI can't rescue a bad product idea by automating things? Like more efficiently doing something nobody wants done isn't a business?

My opinion: In software, the best use of AI is to create a do-what-I-mean UI. That is so-called AI (more to the point, use of LLMs and vision, speech, and audio processing) can allow new modes of interface that help a user operate the computer and achieve something that was nearly impossible with the user at the raw controls.

This will not produce superior work to someone who knows what they're doing at the controls. A competent programmer doesn't need this. An excellent touch typist doesn't need this. But let's say someone who has a vague idea of what outcome they want, but an incomplete understanding of the mechanisms to get there, can get further with a do-what-I-mean UI.

Probably true with a more nuanced definition of "knowing what you're doing". I'm pretty good at writing code, Cursor is faster.

As long as the consequences of then producing something which was not-quite-what-they-mean but them being unable to recognize that are not too high.

Agreed. It moves the interaction to something like you'd have with contractors or service staff. Request, receive, review, accept or send for revision.

One of the keys there will be providing a way to nail down the parts that are correct, hold them constant somehow, and "in-paint" the bits that are wrong. For images, we have this figured out, but for larger works, we'll have to think through the equivalents.

Not OP, but this isn't what Konfig did. They didn't try and rescue a bad idea by automating things. They attempted an AI focussed pivot.

Using AI to automate things to rescue a failing product idea is kind of treating AI as a tool. But AI won't fix a bad product idea.

The problem is, people do want the product that the author was originally selling. A high quality SDK from OpenAPI spec is specifically a product that well funded YC startup I used to work for paid for. So I would contest the characterization that it was a bad idea.

The concept of an API should go away and we should just have AI's talk to each other in plain language.

Reply on news.ycombinator.com