Hacker News Reader

Comment by A_D_E_P_T | Hacker News Reader

Do "AI Startups" even make sense?

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.

View on HN

I sit on an AI evaluation committee for a huge law firm (It's just a regular old consulting gig) - we get so much inbound from (mostly kids) folks trying to build wrappers for some aspect of legal workflow, but behind the scenes thomson reuters is slowly adding everything they're going to need to software they have been using for 10 years now.

Can confirm. One of my best friends is a senior engineer at thomson reuters and they are focused on that.

Have they published anything yet? Would love to read

In many fields there is no moat. It’s an execution battle and it comes down to question: can the startup innovate faster and get to the customers or can the incumbent defend its existing distribution well enough.

Microsoft owns GitHub and VSCode yet cursor was able to out execute them. Legora is moving very quickly in the legal space. Not clear yet who will win.

A business is nothing more than defined, established, repeatable systems and processes. That is the difficulty with tooling businesses.

> Microsoft owns GitHub and VSCode yet cursor was able to out execute them

Really? My startup is under 30 people. We develop in the open (source available) and are extremely willing to try new process or tooling if it'll gain us an edge -- but we're also subject to SOC2.

Our own evaluation was Cursor et all isn't worth the headache of the compliance paperwork. Copilot + VSCode is playing rapid catch-up and is a far easier "yes".

How large is the intersection of companies who a) believe Cursor has a substantive edge in capability, and b) have willingness to send Cursor their code (and go through the headaches of various vendor reviews and declarations)?

Windsurf was acquired for $3B by OAI and it's clearly the worse of the two. Cursor is trying to raise at a $10B valuation and has $300MM in ARR in less than two years.

So in short, yes, companies do appear to be showing some willingness to send Cursor their code, even with all the headache associated with getting a new vendor.

Could you elaborate on the 'headaches of various vendor reviews and declarations'? (I thought paperwork etc was table stakes for a SOC 2 certificate)

They do exist, but not in the simplistic "AI but for X" you described. The article even explicitly says there were zero of those companies in that round.

You can see "AI startups" examples there, A company that manages your business outbound communication with lots of AI features. AI powered code generation for business operations, Accounting services with lots of AI features, business finance software with lots of AI features

This is the way. Use predictable logic and the control of automation with AI as augmentation. So not AI on the control loop, but AI as a helper. Who manages to find solutions that will find the sweet spot of this setup, will be the next success.

Yes, they do, because ultimately software UI is what gets regular people to use things.

To a technical user there may be little/no difference to you between prompt engineering into a chat box vs. clicking a button with premade text slots.

But to the average non-programmer, a chat app like ChatGPT is somewhat pigeonholed into the chat format, and so use cases that don’t lend themselves to this interface will be outcompeted by specific apps that do.

We should never forget the infamous 2007 Dropbox comment! https://news.ycombinator.com/item?id=9224

The exchange was quite nice and thoughtful compared to the usual AI conversations we have on this site.

  - This is the best AI tool ever!

  - Does it fix the reliability issue?

  - Not now, but wait 6 months, because that’s when better models will be coming out

And so since ChatGPT came out.

"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."

Has the download now been deleted

Will it be shared with others

500GB/1TB of storage is not expensive

Were any permissions obtained prior to download

I think it absolutely makes sense. ChatGPT's strength is how generalised it is as a tool, but openAI will never able to adapt the platform to every single use case. You can absolutely use it to learn a language for instance, but a great AI language learning platform needs a better tailored UI, it needs all kinds of non-AI functionality around it like idk a spaced repetition system, it might need to integrate into other platforms, and good prompting to be effective. AI isn't the product itself, but a component to try to solve a problem. And honestly I wish more startups focused less on simply "AI" and more on the problems it should solve.

If for nothing else openAI won't be able to market itself for every single use case, and so long as people aren't using chatGPT for some use case (even if it could perform the task) there's still an opening.

But any use case that gets large enough and makes money will be absorbed by openai direct, based on the market developed by the startup. OpenAI is using the Amazon model. Let someone else spend the money figuring out which market segments are profitable, then steal them with their inherently better access to the platform.

There are use cases that don’t align with OpenAI/Anthropic’s business model, which is to always use more tokens to get better. As the models improve they become way more expensive, with order of magnitude increases in price. I think there is a lane for detecting what can and should be deterministic after doing it a few times, and making it concrete so you don’t just burn tokens every time.

> going so far as to download the fulltext of every legal ruling ever made in the US -- something like 400GB

Where can I find this?

Maybe he refers to something like https://law.justia.com/

PACER exists at the federal level. Otherwise you have to piece it together from each jurisdiction yourself, defeating any anti-scraping measures in the process. Unless someone happens to have made such a dataset available via torrent at some point?

Start here: https://com-courtlistener-storage.s3-us-west-2.amazonaws.com...

The "opinions" are what you want.

These are huge files heavily compressed, so they're quite difficult to handle.

Why are they huge? Is it just PDF overhead? The opinions themselves should just be some finite number of pages of text no?

Nah, they're .csv files, not even PDFs. It's just that it's a lot of text. (The valuations of the LLM giants don't seem too crazy when you realize just how much of the US economy is dedicated to creating and shuffling text.)

There's so much text in each of those monstrous .csv files that you can learn quite a lot if you run a statistical analysis on just one of them.

> Do "AI Startups" even make sense?

I'm doubtful. Remember when Google said their strategy was AI First? Baidu too? I'm old enough to remember that the criticism then was along the line "AI is technology. What problems do you want to solve?". The line of thinking seems still relevant to me today.

Its a good point and it comes back to why Google can't take up projects that other startups are working on.

Google has all the technical infrastructure, talent and everything to make something like AirBnB, Docusign and hell even intellij. Why not?

Thing is, OpenAI/Anthropic/etc. are demonstrably taking up those ideas. There actually were (are?) AI PDF reader startups and AI research assistant startups. (And AI coding startups, AI video startups, image analysis startups, etc.)

"AI startups," if they make sense, seem to have a very short shelf-life. They're either overtaken by the continuing improvement in LLM context windows, or, if there's a real and general unmet need for what they offer, the giants will tend to integrate it.

Google actually did make their own Docusign, it's called eSignature [1] and it was built into Google Workplace

[1] https://workspace.google.com/resources/esignature/

And in true google fashion, it only works with Google accounts; if you send a signature request to a non-google account, it says it's sent but does not work...

There’s a specific reason why google doesnt promote a Docusign like product even when they have superior technical abilities.

It probably comes down to the fact that code is not that crucial but all the other non technical aspects like distribution, supplier relations and marketing that makes a product.

Maybe LLM wrappers turn out to be that way. The model may not matter but the distribution and customer relation etc would matter more.

Because of a mix of comparative advantage and opportunity cost. Google as an entity absolutely dwarfs those other companies, and competes at that scale. Airbnb’s annual revenues are lower than Googles annual r&d spend. Google’s “wins” need to move the needle on a $2tn valuation, and an Airbnb size win doesn’t do that.

This is also the case in the past with companies like Xerox and why didn't they do X with Y that came out of PARC.

There's only so much product bandwidth a company can take on that makes sense, look at the graveyard of Google products.

Because everything they attempt is inevitably compared to Google's ad business - which makes everything else look like a starving beggar.

Reply on news.ycombinator.com