Which kinds of GPT startups will thrive? | Hacker News Reader

Which kinds of GPT startups will thrive?

assistedeverything.substack.com

110 pointsgimili84 comments

Not many of them. Google or Microsoft can see what works for these startups and then put it in their own products. If you don't own the model, you don't have a safe business.

I think it's only a matter of time until open source models catch up.

Closed source models are also improving. Not clear to me why would open source models will catch up.

I mean if GPT-4 is good enough at its current level for people to start companies on it then they will be able to do the current level with open source models at some point.

right,

or giant corps will treat such startups as marketplace app creators for their LLM API and kinda "outsource" the innovation, while providing the platform and enjoying API income?

That’s assuming there is API income to be had. Assuming the “We Have No Moat” memo is accurate, the open source models may soon catch up to GPT4-level performance—and especially if inference compute requirements drop to consumer-level specs, there won’t be any money on the table.

Well there isn’t a sure way to guarantee they won’t tap into your business

Just look at Apple

They eventually will, and it is totally legitimate for them to do so.

In theory the same was true about the SaaS wave and the mobile wave.

In practice, many startups have built valuable businesses around things that Google could've done, but didn't do successfully.

The same will be true in the AI wave.

In theory, FAANG-co could do it. In practice, they can't.

The exceptions to this will be the things where the winner is determined by a dimension that Google will always win on. In practice, those things are rare.

To add,FANGA it does not go after $100 million ideas

And, FAANG doesn't go after ideas that look like $100 million ideas that end up becoming $100B ideas.

For example? Looking at killed by google I don’t see any obvious examples

Many of these https://en.wikipedia.org/wiki/List_of_unicorn_startup_compan... could've been in theory done (to the point that there was no room for a multi-$b startup) by Google (Canva, Instacart, Telegram, Miro, Airtable), but weren't.

Your contributions to this discussion have been very thoughtful and quite interesting to read. Upvotes stopped sufficing. Thanks for adding so much!

Thanks for mentioning, I appreciate it. What was most helpful or interesting?

This is a good list, but most of these Unicorns were product of zero interest rates. The valuations are down, we just don't know how much and some of the companies on the list are already bankrupt like Voyager. Read here [1] about the astonishing rise of the unicorns in the last 3 years.

Also a billion $ unicorn is not the same as billion $ revenue.

[1] https://wolfstreet.com/2023/04/26/what-are-we-going-to-do-wi...

You wouldn't see them there. The vast majority of $100-500m ideas explored by Google are never actually turned into Beta products. They get killed during prototype testing and early alpha, because teams can get that far without meaningful VP sponsorship. But to get the investment required to scale from what a team can do with a UXer, a couple of volunteer SWEs, and a part-time PM ... you need exec support. When I was there, back in 2015-2016 we proposed several product concepts to Prabhakar that we'd already run detailed TAM analysis on, created business cases for, and tested in local markets. All of them estimated at $100-300m ARR. All of them rejected for being too small.

The plus side of allowing experimentation like this at Google, though, is that the work doesn't just go poof even after a rejection. From that core team of 5, two of them are now successful PMs, with one of them having "re-pitched" our concept of appointment booking to the Geo team as a feature addition for Google My Business. It was adopted, as was he, and that's how & why you can book appointments within Maps' business panes now. Another of them took one of our ideas into A120 and ended up mildly pivoting on the advice of the Travel VP. After some additional work, the project was adopted by Travel and is now why you get rich tour creation and destination exploration features inside of travel.google.com. A third member of our team - the UXer - maintained our team site with all the high-res mocks, business cases and pitch decks, and another one of those ideas was ultimate adopted as part of the Workspace team's investment in creating the "Gmail Hub" features that are why you have an expandable right pane with lots of app integration.

The point is not that my team was a bunch of anomalous superstars (we were not - just high achieving normal googlers), but that these kinds of things happen constantly within Google. For every idea you see on a Killed By Google list, there are probably 100 things that were killed as product concepts before release but ended up baked into one of the "15 products with >500m DAUs" Sundar referenced at I/O a couple weeks ago.

> In theory, FAANG-co could do it. In practice, they can't.

It's usually more that FAANG doesn't believe in that product.

MS & Blackberry seemed pretty certain that keyboardless smartphones (the iPhone) weren't the future.

That turned out to be a really bad bet.

For Blackberry, it was way more than just the keyboard. And the whole MS-Nokia thing is a whole different story, because both screwed up really bad here, botching the potential shot at owning an Android / iOS alternative.

Or that the particular vertical just isn't big enough to warrant their focus. A $20m MRR is great for a small SaaS but isn't worth the time of any FAANG.

Google/MS make generic software for businesses and developers. They don't really get involved in specific verticals (aside from obvious ones like cloud, gaming, marketing). Successful start-ups will be the ones taking models from big tech cos and bringing them to the boring verticals those companies have no interest in.

Then one day Google tweaks their model in an unfortunate way and you fall all the way down that vertical.

Can you give an example?

A large majority of SaaS businesses can be done for free with an Excel spreadsheet, but many of them are still making millions a year.

Interesting. Such as?

YNAB, the budgeting app is a good example. Having it as an online app means you can use it from many computers or a phone, but at its core it's really just a spreadsheet.

The same can be said of any of their service offerings. You could build a business around something MS/Oracle/Google provides and they could one day decide to deprecate it.

Anything a SaaS platform provides can be replicated in-house. The same can't be said for a bleeding-edge 10 billion dimensional model.

Just build a product that is even slightly reliant upon customer service. They won't know how to compete.

This comment isn't getting the recognition it deserves.

The difference in cloud hosting and gpt api costs are enormous Though no?

Faang will tariff all these innovative co

If you or anyone reading this makes a GPT powered product that has enough usage and happy users that the main challenge is API costs, email me (email in profile) - I'd be interested in funding it.

As long as there is a product that users love, and a plausible path to a moat, VCs love businesses where the only bottleneck is hosting/API costs.

Has the fundraising environment shifted such that Google/Microsoft can more easily purchase their competitors? I'd assume raising funds has gotten harder, and that that change will make a selling more appealing.

I think there are a few opportunities for startups that want to leverage GPT technology:

1) Fine tuning base models with data that big tech doesn't have access to. E.g legal, medical, support data. Offering custom fine-tuned private hosted models for companies that can't leverage the base models APIs due to data privacy and lack of domain specific training.

2) Using GPT on the backend to do data transformation that the user doesn't interact with directly, e.g parsing logs, events, moderating content etc.

I don't think the opportunity lies in creating a thin wrapper to a custom prompt in a chat interface.

I would not be so sure about #1, OpenAI has been very cagey about what exactly the newer models are trained on.

I agree. People seem to largely be pretending it's not the case that GPT-4 significantly outperforms every other model right now.

I want to agree, but I think that’s what I thought at the dawn of Web 2.0 — no way slapping a basic CRUD UI on top of an open-source database will ever power anything but the most trivial use cases.

Turns out knowing Excel and programming and NOT knowing business and design (at that point in my career) might have cost me a couple bucks.

There's millions of small businesses that don't know what CRUD or open source means.

Hell just last year I helped a neighbor get a basic informational website up for a non profit he was part of. They tried it themselves but bungled it a bit.

They want something a bit nicer now and he said they're paying someone $16k to redo it.

A nonprofit that I was (voluntarily) running a mostly static wordpress website for paid a consultant thousands to rewrite the whole website because they wanted an entire conference registration and scheduling system. The new website is still in wordpress, they're paying hundreds per month in 'maintenance fees' but can't get the consultants to do anything for them, I don't have admin access, it doesn't have any new conference or scheduling functionality, and they don't understand why I don't want to help them any more.

If you or they have access to the database, just assign you/them an admin role and/or change the password and/or create a new user.

Consider any contracts currently in place before you do something like this, and figure out who is the name on the hosting, or whose server it is.

The nonprofit appears to consider it a success and is embarking on a new project to buy/build/whatever a CRM that will do the things they originally believed the website project would achieve. I can't even get them to use Mailchimp reliably, and they want to solve that with magic technology.

I'm surprised how long the "solution in search of a problem" trend has dominated tech product design, despite obvious and repeated failures of this approach to produce results.

AI/ML products fundamentally don't make sense compared to products that happen to use some AI/ML to aid in solving a problem.

It's sort of like loving to use redis (which I do) and thinking you want to found a company based on using redis in the product, or start a redis product team, dedicated to shipping products that use redis.

It's one thing if you want to host redis as your business, which is solving a problem involving redis, but if your aim is to use redis to solve a problem then you're going to be in trouble.

Imagine a PM on the "use redis" team rejecting a great idea for customers because it could be more efficiently solved using a traditional database, or forcing the use of redis when a cheaper, easier solution already works just as well if not better. This is actually the case on AI/ML teams.

GPT startups that will thrive are the ones that aren't GPT startups, but instead solving some other, real, problem that happens to only be solvable in a post-GPT word.

I'm not too surprised the trend persists, if I just follow the money. People like to bet on new(-ish) technologies and hypes, hoping to be the first in some brave new world - like with personal computers, the internet and mobile.

Didn't quite work for crypto or VR. I think that's partly because the economy changed: I don't remember such an insane amount of companies chase money so quickly with those earlier examples. For a long time there were lots of people just playing and tinkering with the new thing for fun, not immediately trying to get rich (even mobile, I fondly remember having PDAs long before there was an iPhone, tinkering with them and hanging out on XDA developers). If everyone is primarily trying to make money off something right from the get go instead of primarily trying to build something useful, I can see how adoption suffers. Bit of a catch 22.

What kind of deep learning startups survived or thrived? The underlying technology (of gpt et al) is racing towards being a commodity. There are very few "pure play" "AI" companies left from the deep learning wave, and what few there are pivoted to become SaaS with some ML features. Something similar will happen here. Companies that solve a business pain point, irrespective of whatever "AI" will get somewhere.

I wonder if the issues around owning of training data may allow larger, well funded companies to acquire ownership (for example, buying stack overflow content). This might allow larger players to starve smaller players.

If we're not already, I'd expect to see lots of previously public content get locked down so it can be monetized. It will be interesting to see what the long run equilibrium looks like, as in will large players play ball?

I'd guess that the future high value application will still have lots of fine tuning on more specialized datasets, which will continue to be a bottleneck and be considered valuable, while the large scale training data will be less important (with respect to transformer llms, who knows if there is some newer breakthrough). Same way that e.g. everybody uses CNNs pertained on imagenet, and fine tunes on application specific stuff, and there was not much of a commercial push (imo) to rush out and build a better general purpose large scale image set like imagenet.

It'll generally be the same as with other startups.

Using GPT of course isn't a moat in and of itself.

So, it'll be companies that can do the following:

1) Build a product that people love

2) Then, reach those people, eventually at scale

3) Then, monetize those users

4) Then, build some kind of moat to enable pricing power

For now, most AI startups are best focusing on #1. This is where most GPT powered tools fall short.

ChatGPT is popular because it was so good that people had to keep coming back and using more of it, and they had to tell other people about it. ChatGPT's utility was measured against a pre-ChatGPT world.

The bar is now higher, because that means whatever you're building has to meet the same 10x or 100x better than existing alternatives bar, except now your users live in a world where ChatGPT exists.

In short, the GPT startups will thrive are those that can build products that are 10x better than whatever users are doing now.

Practical takeaways:

Build deep relationships with your customers, understand their world, and have a high shipping cadence pushing out new versions of your product multiple times a week until you build something so good that they'd be really disappointed if they couldn't use it anymore.

The described step function changes are super apt and mostly overlooked by the pure hype crowd (although the step functions enabled by AI should really be what the main hype ought to be about!)

That said, once you get into the step function changes, the GPT-wrapper accusation might quickly become akin to a "AWS-wrapper" one, with traditional moats getting more important than AI-native ones.

We've had internet-enabled businesses without technical moats (but very real other moats, be it UX, social platform effects or a great b2b sales process) for the longest time, and might just see the same thing play out in AI native land

They need to find a niche that won't be threatening to any of the big honchos. Possibly marketing themselves as "lawyer AI", "MBA AI", "MD AI" etc. Ideally if there was some common opensource base like Linux distros but for GPT models, constantly getting updated and fine-tuned with anyone simply pulling them and using them when needed.

I can't imagine what kind of liability insurance would be needed for a GPT solution aimed at such heavily regulated fields as law, medicine or finance.

You can add all the caveats you want, but I suspect ml chat is stuck in the uncanny valley of not good enough to be trusted and not bad enough to be a toy.

Yeah I dunno about that one lol. Open AI backed solution for Law, already has deals with some of the biggest law firms in the world.

https://twitter.com/ai__pub/status/1644735555752853504

https://www.lawnext.com/2023/04/harvey-ai-raises-21m-in-a-se...

Microsoft plans for Medicine

https://arstechnica.com/information-technology/2023/04/gpt-4...

GPT-4 is more than good enough. People have no idea what's coming. There just needs some sort of supervision or oversight (and right now, it'll work with rather than completely replace practitioners).

Disgusting content marketing. The shtick works like this:

1. Think of a topic that is remotely related to the product you are trying to sell (GPT can even help with that).

2. Write a low-value (what new thing really did you learn from reading this?) article that looks like it's providing useful insight.

3. End with a shameless plug about a service you are trying to sell.

I thought this scam was debunked a few days ago right here on Hackernews? Why is it still getting upvoted?

I think AI startups will face headwinds because the established players have access to valuable model training data. For example, a large health insurance company can mine their claims data and create features and software offerings from this data, a startup will have to pay handsomely for valuable data.

I think the only possible moats for AI startups are in seedy or unethical areas like disregarding copyright or stuff related to sex (AI waifus). These areas are getting less attention and have no likelihood of getting zero'd by openai or google.

Not sure about disregarding copyright, like skimming around it?

Those that have a realistic acceptance of GPT's limitations.

And that's almost impossible at the moment, because there's so much progress. It's very hard to tell what the actual limitations are. But there will be some. GPT is not AGI.

Companies that think "just get 10 times as much training data, or a little tweak of the model, and those limitations will disappear!" won't last (though they may do really well at getting funding). But companies that think "the current limitations are permanent, so there's no point trying to get it to extend past that" will also not last (unless they adapt fast enough).

Since we don't currently know what the real limits are, it's very hard to give concrete advice here. Maybe "push hard against the limits, but don't bet the company on being able to overcome any particular limit".

Look at what thrives already. Fast food. Porn. Dominance fantasies. I think that covers it.

Until we can deliver cocaine over wifi.

Can we do that? Tell the GPT "get me high". Delivers some kind of genius AI contrived neuro-reactive audiovidio experience. No narrative. Just "chemicals".

THAT would be a good product.

We're experimenting in this space, and I think our approach is a nice balance. (I'm clearly biased).

OpenAI is giving us a chat-based interface to our core product, which we didn't have to build ourselves. Our platform was exclusively developer focussed (automated API integration), which has value on it's own, but was limited in reach to a technical audience.

By adding a chat interface, we get to make our tooling available to a whole new type of audience - non developers, who can "chat" with their API estate.

I personally like this balance - using AI to lower barriers and widen the reach, but the underlying offering has value standalone.

I thought this article could have been about 95% shorter (let's just say I saw high irony in the cartoon in the article about using AI to summarize needlessly wordy emails).

I think the only place "GPT startups" will really thrive long term is specific niche business areas where the big boys (Google/MSFT) are not likely to want to compete. For example, there was an HN post about a legal startup that used AI for various purposes. I could see that one building a sizable moat over time as "the go-to place for legal AI support" if their UI is good and very tailored to legal-specific workflows.

My primary point is that I think "generic" AI tool startups are likely to fail because the big boys will just build them into their products. E.g. a tool that just helps you write is going to have a hugely difficult time competing against the integrated functionality of Word or Google Docs (I'd be shuddering if I were Grammarly). Google and Microsoft, though, have largely stayed out of dedicated tools for highly specific verticals, and with all of the antitrust eyes on them I think they're likely to stay out of those spaces.

So you're saying that Google, Microsoft, etc., do have moat that smaller companies don't have. Otherwise, I see no reason why these companies couldn't continue to compete with large corporations. For example, if the makers of Papyrus Author successfully integrate AI into their writing application, they would continue to have the same competitive advantage over Microsoft Word they had before - unless Microsoft's AI is more powerful and much better.

I don't know, I consider it possible that at least in the beginning large corporations have moat, just wanted to point out that this is what people are wondering / don't really know at the moment.

I didn't really understand your point. But bringing up Papyrus Author is a great example I think of what I was saying. I'm not previously familiar with Papyrus Author, and I don't know how popular they are, but it's clear they are very much targeting creative writing authors specifically. Their hope is clearly that they can provide enough value to this group so that their price is worth it over a more generic word processor.

Any other examples? Law is a trillion dollar industry and the AI startup mentioned, Harvey is a backed by Open AI

Here are some other examples I can think of:

1. Architecture. I could easily see architecture-specific AI tools being incorporated into design apps.

2. Similarly, interior design tools.

3. AI tools for construction.

4. Anything in healthcare. Healthcare is so regulated and privacy is obviously paramount thus I'm sure there will be companies that spend a ton of time/money providing "HIPAA-compliant" diagnostic support tools.

We're taking a shot at hardware engineering: www.valispace.com/ai/

And I see similar things happening for major fields such as education, law, art, software development, management, etc. Here I looked into a few examples of what is happening in these fields already: https://assistedeverything.substack.com/p/the-age-of-assiste...

GPTs are distractions. Now's a great time to aggressively pursue something else.

Maybe a law firm that specializes in AI law. Didn't get a job because turned down by an AI, etc

It's plausible that a handful of these new startups can define themselves as the next generation software of their industry, but I'm not sure many of these startups will make out of the "Tug-of-War valley" as the article describes.

I'm amazed at the amount of seed deals being done around "X with AI" where X is an established area of software.

The bet is that a new startup will be able to deliver a better product than the incumbent players (often established companies with large adoption and distribution).

Of the many I've looked at, the hurdle the startups will have to clear seems to be massive compared to the incumbents being able to build these "AI powered" features.

Thousands will thrive. Startups have always been a Ponzi scheme.

The winners are the ones that grab funding now, achieve growth and sell quick to hand the bags to someone else.

Most AI startups for the next couple of years are just going to be scams that work of the back of OpenAI's API. My work place is currently under an onslaught of every person thinking there's a plethora of tools to do there job for them when it's just all OpenAI in the end. I believe AI is here to stay and will revolutionize the work place. I just feel way to many people are going about it the wrong way.

Uh you could make the same accusation about processors or cloud or any base technology. I think your view here is stunningly naive.

Fine tuning GPT4 produces real and tangible differences let alone building an interface for specific workflows.

have you found examples of new products with successful distribution using fine tuning

I mean, you aren't going to finding massively successful products in a matter of months -- people here are being very unrealistic. I'm just saying it works.

You cannot fine tune gpt-4 yet, unless there is some private access to do so I have not heard of.

https://platform.openai.com/docs/guides/fine-tuning/what-mod...

Make things more simpler, more convenient, and faster for the end user, even if it makes things a bit worse, stupider, and more annoying for everyone else.

Prompt assistance

Prompt/Result sharing

Easily fine tune your own dataset

Type barely anything and get something reasonably valuable

Click to build prompts without having to type

Automatic results based on other activity (location, browsing, messages)

Let's call the other guys "Gatekeepers" (Open.AI, Google with Bard, etc).

Option 1: none. Most of the value will be captured by Gatekeepers

Option 2: Gatekeepers will partially commoditize their service, and on top of them, several startups will thrive by creating something not easily replicable by Gatekeepers (via patents, via speed of execution, via viral growth, etc). Example: biggest GPT-powered media startup will compete with Netflix. Another: biggest GPT-powered e-learning startup will compete with higher ed - Stanford, MIT, etc.

Option 3: A single GPT-powered Coding startup will become > $100B. My bet is on Replit. (disclaimer: very early investor). When you hire a programmer, much like you pay for Jira, AWS and such, you will also pay for Replit. This partially overlaps with e-learning (see above).

Option 4: there's an even bigger revolution coming in AI, and it's not in the segment owned by LLMs. Or, it's a different interpretation of LLMs. Could it be... finally a real self-driving car?

Option 5, very unlikely: regulation will stifle competition and innovation, and most things will be killed by governments. Perhaps something smart can be said about US vs China. WWIII will be fought with virtual agents powered by GPT, over Twitter. Elon Musk will be kidnapped by GPT-6. /s

What else?

for option 3 wouldn't you need a substantial articulated value prop over co-pilot?

A start up in the: Agents, Agent frameworks, Tool Hubs: Huggingface, LangChain. Are probably the only "startups" that have a shot.

As for Data lakes I suppose that's true but alot of companies have their data in the cloud now. so I don't know how much data security/privacy/cost is a barrier for these gatekeeper companies.

It also all depends on how long AI is weak at using and generating a UI.

“ biggest GPT-powered e-learning startup will compete with higher ed - Stanford, MIT, etc.”

The higher up the content complexity pipeline you go, the more complex and correct the content will need to be. For this reason, the biggest e-learning AI product will disrupt the bottom of the market first (k-12) to replace legacy curriculum and assessment companies.

Could you expand on option 3?

The startup that is able to train new models from the ground up with new architectures of course. Nobody assumes that transformer-self attention is pinnacle of llm design we will not be able to surpass it, right?

ChatGPT's summary of the main points:

  | Point                          | Category     | Examples            |
  |--------------------------------|--------------|---------------------|
  | Real/Usable Objects            | Value Peak   | Software coded      |
  |                                |              | using GPT, hardware |
  |                                |              | products developed  |
  |                                |              | with GPT-enhanced   |
  |                                |              | requirements        |
  |--------------------------------|--------------|---------------------|
  | Experiences                    | Value Peak   | GPT-generated music |
  |                                |              | tailored to your    |
  |                                |              | taste, personalized |
  |                                |              | bedtime story, role-|
  |                                |              | playing game        |
  |--------------------------------|--------------|---------------------|
  | Solving Personalized Problems  | Value Peak   | Suggesting recipes  |
  |                                |              | based on your fridge|
  |                                |              | contents, creating  |
  |                                |              | personalized        |
  |                                |              | worksheets for      |
  |                                |              | students, etc.      |
  |--------------------------------|--------------|---------------------|
  | In-Context & Collaborative     | Basin of     | Platforms requiring |
  | Features                       | Success      | collaboration, like |
  |                                |              | Figma or Google Docs|
  |                                |              | enhanced with GPT   |
  |--------------------------------|--------------|---------------------|
  | Gated Knowledge/Data           | Basin of     | Company-specific    |
  |                                | Success      | data, domain-specific|
  |                                |              | data, complex-to-   |
  |                                |              | parse data          |
  |--------------------------------|--------------|---------------------|
  | Edge Computing / Offline Use   | Basin of     | Applications running|
  | Cases                          | Success      | locally for privacy |
  |                                |              | reasons or specific |
  |                                |              | offline use cases,  |
  |                                |              | like personal       |
  |                                |              | assistants          |

the sub-question in here of 'what is SEO for GPT' is interesting

how do you incept a model to tell someone to use your product

as a society that's been through ten years of extreme adtech, will we crack down on ad placement in LLMs?

ready for someone's dystopian arxiv post about how to inject RTB in the attention heads

I imagine gpt will use the same or similar heuristics as SEO to determine what to include I.e number of backlinks

To add to all the good ideas, there is also a "dark path" which no big company will go.

I heard an argument that many GPT startups (more specifically the products they make) could be the new consumer products. As in, they come up with something interesting that creates a buzz, makes some or a lot of money, then people move on to the next thing.

Reply on news.ycombinator.com