Over 40% of the latest YC batch are AI/ML startups
ignorance.ai
ignorance.ai
"We are not a black box that sends everything to someone else's giant API" can be a selling point as businesses become more aware of their cloud dependencies.
Perhaps a better question is what new niches now open up, on top of the AI solutions.
Hugging Face is a nice example - it's mostly a python library, a github clone, and a discussion forum strapped together. "Old" tech used to serve a new and growing niche.
Great Library -> Large Community -> Large Coverage of Models & Datasets -> Larger Community -> Larger Coverage of Models & Datasets -> Revenue -> More Engineers -> Greater Library ...
Not all the steps (->) here are trivial: Contrast HF to Explosion (great folks behind spacy).
Can you elaborate more on this? about contrasting HF to Explosion?
This never transformed into a model hub. Despite a lot of people using SpaCy and probably building custom models.
Again in contrast, Explosion's other revenue stream (prodigy) is not a SaaS as well. Its a great software, and I presume it brings in a steady income. But in 2023, I would imagine that HF's LLM hosting, cloud training environment, brings in more money than Explosion's data annotation software.
There's no point on creating your own GPT before validating your idea and getting some traction.
Amazon didn't build their logistics and freight division first to then start selling books.
Once your product or service is off the ground and more resources are available then you can start moving away from OpenAI.
This could be much simpler that you think, particularly if you designed your systems with this in mind and considering future costs of training LLMs will be a fraction of what they are today.
Don't know how I feel about that, it quacks like a bubble but there is massive value that can be delivered.
How do you figure that? Given the speed that the gpt4 API returns data I feel like it's easily going to cost more than pretty pricy hardware.
I don't know for sure, but I've been in software for decades and you don't just get 10 million people to pay for something in a few months. Conversion rates are always lower than people expect.
For something like ChatGPT, the math is going to be something like this:
- something like 150M unique users have tried it[1], which is probably < 50M people (I myself would look like at least 5 users)
- most will never try it again
- of the ones who try it again, a very tiny minority will be interested enough to pay for it because it's just a novelty to the vast majority of people (the only people I know using it for work are researchers and content marketers)
- of the ones interested enough to pay for it, only a fraction will
So I'd wager they're in the territory of 20k-500k users, not 10M.
> Given the worldwide buzz and a small sample around me
People who are educated, live in the West, and pay attention to news always dramatically overestimate the number of people with a similar media diet. ChatGPT is discussed in your circles far more than elsewhere. There are literally billions of people who haven't heard of it or don't care.
Don't let your anecdata fool you. Your sample is not representative of humans at all. 60% of the world lives in Asia, after all.
1. https://www.reuters.com/technology/chatgpt-sets-record-faste...
Here is the math I do, starting from the addressable market.
There are 60M professional workers in the US [1].
The number is probably similar in EU and APAC.
This yields ~200M professional workers in the world.
If 5% of them have a paid subscription, that's exactly how you get 10M paid subscribers.
Of course, the debate is around the 5%. The rate in the pure tech industry is much higher (10-20% around me), and it is probably much lower in other areas (e.g. architecture). But 5% as an average is not unrealistic.
[1] https://www.dpeaflcio.org/factsheets/the-professional-and-te...
That's what concerns most about this, is how somehow it looks like they've optimized for a bunch of "follower" founders.
Isn't that the textbook definition of a startup?
For example, Tesla vs Nikola. One set of founders had a powerful vision and technical ability to see it through (yes, even Elon, say what you want about him but he has deep technical insights, moreso for his cofounders). The other set had a good funding catchphrase "Electric cars, but with trucks!" and tried to fake their way forward using copious amounts of VC money. Many of these companies sound like "AI, but with ... !"
How do you explain that the vast majority of startups fail, then?
IMO the best a VC can do is check that the startup they invest in is not an obvious scam. But then it's impossible to predict which technologies will work and which will not. So they just diversify, counting on the fact that a few of those startups will bring significant money at some point.
Sure, they like to share their "vision" and "predictions", just like traders probably like to share their opinion on the stocks they buy. But at the end of the day, they don't know, they just diversify their investment.
Right, how many products are just shims over the CPU instruction set.
so no, i don't think that the idea is what counts.
It doesn't need to be sustainable nor particularly innovative.
Also, YC seed isn't gonna get you very far when compute alone for training an LLM is in the millions currently.
I'm not saying it's going to happen, and I hope it doesn't, but when I did a little analysis for an idea I have it was absolutely the biggest threat I could see. OpenAI has the ultimate vendor lock-in - there literally isn't an alternative right now.
(This is a giant opportunity if you're a billionaire investor. Providing an AI backend is probably the next cloud infrastructure. OpenAI is going to be the AWS of that industry, but there's room for an Azure or a GCP too.)
Yes, keep it in the back of your mind and design your systems in such way that when the time comes replacing the GPT doesn't require rewriting your entire codebase.
Your immediate #1 priority in such competitive market should be to execute your idea and get paying customers, then you deal with breaking off OpenAI.
You can train GPT2 in a couple weeks on a dozen GPUs. Sure, it’s a lot worse than GPT4, but not insanely behind and the community is fast.
Same for RLHF. It is for sure fine engineering, but not unreachable for a motivated and competent bunch.
If you want >90% of projects you invest in, you need to invest in conservative fields, and grab 30-70% very early on, because your returns will be way lower.
As for not bringing anything to society - you mean to say that tech field and our startups didn't bring worth to society? Seriously?
Wanna talk about surveillance capitalism? Ultra consumerism? The biggest startups in the world have created a world where the goal is to win people's attention to sell ads. The goal of startups today is not to make good and useful products, but to sell stuff. Usually it means that time to market is more important than quality. Remember fitbit? "Doesn't matter if the product is crap, as long as it is cheap".
Not everything is bad, but I think we can definitely wonder if the world wouldn't have been better off without those startups.
Maybe it's a good way for YC/HN to find out which things don't work, in a short span of time.
Is what it is
My first startup was literally just some Javascript and a Postgres database. It didn't even call any external services, let alone one that's truly on the bleeding edge of computer science. I still thought it was a viable business (wrongly, it turned out, but still... I had a dream!).
"Some JavaScript and a DB" is a web service. It does stuff. Sure if it's small or a basic crud then the moat is small
But if I can replace the value proposition of a startup with "open chatgpt and type a short prompt" then there's not much value in it right. Maybe if there's value in the surrounding services.
This is how many successful businesses operate in the "atoms" world.
You take a product or service that's already popular and has a solid channel for distribution, you transform it or add additional value to it and sell it for a profit.
Think about a $8 precut, packaged pineapple at the grocery store.
Once you get your business going and get additional funding or generate cash flow, you can invest in getting rid of some of the middlemen so that you can take their profits too (start growing your own pineapples).
Sounds pretty straightforward, but the selling part is the difficult one.
Especially when these products can be built by a school student in an hour using ChatGPT?
I think the barrier of entry to atoms world companies are quite higher than that.
Generally yes, software is usually easier, although today you can start a new brand of vodka or cosmetics from your computer.
Customers acquisition is equally difficult for both though, so sales and marketing is where the secret sauce and differentiation is now, not so much in the codebase, I think.
The value proposition to the sliced pineapple is the time and skill to do it. Sure, it's "cheap" to do it at home (if your time is free and you have a good knife - I pay way less than $8 though ;) )
While I do believe there are startups with good value added, I think a lot of people are just throwing a simple call with very little thought added. And while the service is sellable, your moat is how easy people can duplicate it.
Is your service a product or a feature? If it's a feature OpenAI can just add it to their systems in a day
They sure 'really' are AI companies and totally not just wrappers around OpenAI or ChatGPT's API and giving Microsoft and Google more ideas to absorb for free with another AI bubble waiting to pop as soon as OpenAI increases prices. /s
Then we will see how unprofitable these so-called AI hype 'startups' are.
or rich get richer, poor get poorer and will not have access to AI tools, society will collapse, more jobs will be low paying, most on demand skills would be everything done by physical activity
The great promise of robotics and AI is the possibility of decoupling societal stability from the vagaries of the poor.
By having armies of semi-intelligent armed drones, that will kill the poor if they decide to riot?
We already have many things that are given away for free, whereas a 100 years ago they were considered luxury (books, pens, papers, access to knowledge in general, and in Europe - public transport and healthcare, plus to some extent food, heat and electricity).
If you’re a part of a startup ecosystem, you’ve seen how it works first hand - wealthy and successful sharing knowledge with the others.
In our current compounding-capital model, what you get is a depreciation of all human capital to zero, therefore everybody who doesn't already own capital is fucked for eternity because they have nothing to compound and get the ball rolling. Zero social mobility, absolute polarization between an AI-owning class and an own-nothing class for as long as revolution is prevented.
Computers, the web, mobile drastically increased the productivity of office workers, but somehow they don't work 15 hour work weeks.
For me, personally - before the startup ecosystem happened in my country, it was extremely difficult to break through. If you worked for a large entity - good for you. If you wanted to do something on your own - you weren't treated seriously, and it was very difficult to get through.
Now, people from anywhere, can visit meetups and hackathons to get up to speed, and gain connections. And accelerators like YC provide resources and mentoring to the ones that are the best.
Also, notice that YC has an open application, and they don't look at your social status, your wealth, nor your connections, when judging your application.
Of course it's not perfect - being from US, or the bay area, helps. Knowing people who got through YC and can help you prepare your application also helps.
But I haven't heard of anyone with a better idea for solving this.
If you look at the past HN comments there were users who predicted the raise of OpenAI, the war in Ukraine and so on.
But this is not at all the same kind of "predicting future events" as accurately saying where e.g. the CryptoCurrency or VR industry will be in a decade or so.
So even if they make money AND hire 'data scientists', there is a higher chance that they will run out of money quicker since they are competing against Google, Microsoft and OpenAI which they all can afford to run everything at a loss for free and raise their prices which eats these startup's revenue.
Either way, they are locked into these APIs and they aren't focused on profitability to even begin building their own models at this stage. If they are not profitable by the time the price increases come in, they are going to be struggling for survival.
The former are going to mostly be successful, particularly in the place of companies that don't adapt, but the latter are mostly going to be obsolete by the time they'll be several years into it.
AI as a product is a TERRIBLE play. But AI as part of a solution is going to be necessary for nearly every business vertical.
Similarly, lots of companies spent thousands of dollars on a website or a mobile app that no one looks at because they're a small town landscaping business that really doesn't need those things.
The same goes for AI. There are definitely use cases where it will improve how we do things. For some businesses that will make them more profitable and more competitive. A lot of businesses won't adopt it for decades. Some will adopt it when it brings them no real benefits.
AI is a tool. People will often use it poorly, just like every other tool.
I think language models will become a commodity, like a spell checker is today. They might open up the market for translations into more exotic languages (but I don't know how the quality is for such languages and by definition, the market ain't big) and I think every author will use one as part of their workflow (think of technical writers using a language model for the textual parts). Computer games might profit massively in terms of immersion.
No, I don't think that new applications will become that important with language models. I think it's the other way around. There are so many applications that we simply don't need any longer. Language models and image recognition models together will take over the user interface industry. Why bother with web design and user experience if you can just throw the chatbot at your client? That's going to become the real value here.
If I was a junior web developer, I would begin to seek out new programming languages and stacks.
What about all those technologies that just died? Companies that did not embrace Minitel, blockchain or the metaverse seem just fine right now.