That’s a compelling story. But that doesn’t make it accurate. We are early on and it’s hard to determine how large or small the ecosystem and value that will be created on top. Some went big, like personal computers, the internet, databases, Java, electricity, and so on. Others did not (yet), web3 (maybe), other databases, low code platform, website builders, workstations, many old personal computing systems, etc.

Yes, it’s easier for incumbents to capture value because they own customer relationships. However, if that was always the case open ai wouldn’t exist. So some do breakthrough.

Early on could people see the future and know what Ms dos, Oracle, or google would become? No because it’s new and it’s the future. It’s only obvious in hindsight. Yes, all of these ai companies on top of chatgpt could become nothing. On the flip side the complexity and depth of chat gpt and other models will continue to expand, creating more possibilities for potential value creation. Or the whole thing can turn into an overvalued version of Clippy. It’s hard to see how it will play out.

Good points here.

I certainly see the opportunity for start-ups to break through here. I just don't have much faith in ones where GPT-4 is their core tech.

On the other hand, companies going after the entire vertical where AI is a feature and not the product, or well-funded companies with innovative models that outperform OpenAI is totally a possibility.

I'm not convinced of the alternative either though. If you think of the AI as an intelligence akin to an employee.. at some level it becomes a commodity and then it becomes what is the AI doing.. what is it harnessed into?

> Early on could people see the future and know what Ms dos, Oracle, or google would become? No...

In the case of Oracle, yes, because Oracle was created directly to compete with DB2 (which was already a success) but as a product that could run on all the popular platforms of the day. So success of their early RDBMS-based-business was predictable, assuming they could execute.

I used to think I wouldn’t rationalize a losing position while I was invested, but after a hedge fund I was in divested from a large position, I was able to see the other way so clearly

This is quite funny. If it weren't for tax (and maybe liquidity in certain situations), this suggests a quarterly "sell everything" day might make sense. You'd sell everything on a friday perhaps, and come monday buy again after having the weekend to freshen your eyes.

alternatively you could take unbalanced bearish positions, like too many at the money put options, everyone can start thinking about the ways that becomes profitable by noticing the bear thesis and rooting for that for a little while.

then people could see which thesis is stronger and more probable, the bullish or the bearish ones.

Good points. Premature optimization and over engineering when one hasn’t been part of either not having any esoteric or unfair knowledge is like shooting in the dark.

YC’s premise to start small, serve the problems and solving them and they will light a path has merit to it especially since they took on a B2B focus.

Except PCs, Internet, etc were new things or in case of Java, improvements

These "AI Startups" were just 2 lines of code wrapping a call to OpenAI (or 10 lines if js with langchain - which already makes it easier for devs to use features like "string templates" or such /s)

> if that was always the case open ai wouldn’t exist

Because there were no incumbents offering a ChatGPT similar product

There is value created, but the argument was that it is the big ones with existing customer relationships that are going to benefit, not the 100th startup that offers an AI assistant to schedule appointments via voice commands. I think all of that second group of thin GPT wrappers will be eaten by the entities developing and running the foundational models like OpenAI.

In general, AI is data hungry and college kids in their dorm room will always be at a disadvantage in that regard compared to large companies (many of which locked down scraping on their properties, even if "their" content is mainly generated by users, i.e. non-employees).

> In general, AI is data hungry and college kids in their dorm room will always be at a disadvantage in that regard compared to large companies (many of which locked down scraping on their properties, even if "their" content is mainly generated by users, i.e. non-employees).

It's a two-way street too. The big players have more data to start with and they're already the "place" that you go to add more data.

Google's ML/AI stuff (the behind the scenes stuff, not the consumer facing chatbot type stuff) in gmail/calendar/etc is already useful because of their vertical integration. They "see" everything and can add new records in any part of their productivity suite, then make correlations / recommendations across the entirety of their offerings in a way that ChatGPT wrappers outside of their walled garden will never be able to replicate.

Now with OpenAI, MS has the same opportunity. Issues on github? Code? Emails? Teams conversations? Calendar appointments? Bing search history? It's all just part of the same bundle to them and it's all read/write.

Many mobile and web apps are database wrappers too.

Replicability isn’t the only part of value.

Solving a valuable problem is. What about the tools quietly moving mountains and not seeking the spotlight?

Most customers of software are not in the software development space.

I think that those tools not seeking the spotlight might also be at risk. I'm fairly sure some of these tools have a 'twin' that has found success by securing early adopters and has become the de-facto tool for 'X' job, largely due to marketing.

Very plausible.

Still it feels like there’s way more problems that software hasn’t quite solved yet that LLMs may only go near.

What risk? Most of those tools are created by inexperienced people that will only notice how bad it was in the future (and they will fail to achieve the same success with much better products later)

yes the HN bubble is real. A no-coder developing a successful GPT wrapper won't care about moat

There are still many use-cases where further control offered by open-source models (i.e. fully supporting all constrained text generation techniques) is superior. Stable Diffusion XL fine-tunes with controlnet, regional prompting, and the further extension ecosystem is still better than DALL-E-3 in the hands of skilled users.

But yes, the distance between the leading foundation models and the leading open source foundation models for LLMs seems to be getting farther, not closer.

This is a prediction that there will be zero surprises. An obvious prediction to make, but probably wrong.

> I see the profits in the current AI hype cycle going to the leading foundational model company (why pay for the 2nd best?)

Because it's a race to the bottom. Look at crowded spaces like TTS. Eleven Labs looks like hot shit, but in actuality they have over a dozen well-funded direct competitors. Including OpenAI.

There won't be a "Twilio of [X] Model Inference", nor will there be a "Best Foundational Model for [Y]". These are features, not moats.