The Leading Edge #16

Welcome to edition sweet sixteen of your weekly dive into the tumultuous waters of tech! ???

This week we have one single topic - but a pretty long one at that. How does AI startup survive the onslaught of incumbent tech behemoths? We’ll explore the nuances of competing in a market dominated by players like OpenAI, Google, Microsoft and Meta - who not only set the rules but have the resources to change the game at will.

Answers to Perplexing Questions

Five months back Perplexity.ai announced a $73.6 Mn fundraise, at a valuation of $520 Mn. That’s an impressive number for a startup that’s just two years old.

But the CEO Aravind Srinivas still has a problem. ?

Because Perplexity is a thin wrapper around popular LLMs.?

They scour their trusted web indices, feed the results into AI models like GPT-4 and Claude 3, and voilà – formatted answers at your service. All for the modest sum of $20 per month (or $400 annually) per user.

This does not mean they provide no value. Far from it.

I am a frequent user of perplexity. And it’s not just me, there’s 9999999 other people using their services (Aravind mentions he gets 10M queries per day ??).

But then, the incumbents also realized that Search Engines are a thing of the past - Including the proverbial 900-pound Gorilla of search business - Google (They serve 8.5 Billion queries a day).?

In the Google IO 2024 event, google put a lot of emphasis on “Search in the Gemini Era”.

OpenAI is not sitting idle either. The world is anticipating them launching a search engine soon.?And they seem to be well on their way to announce it. Financial Times recently entered into an agreement with OpenAI where their content will be used for user queries. OpenAI has similar agreements with many more such companies - including Axel Springer Germany, Le Monde France and Prisa Media Spain.

Aravind realizes this is coming - and he said as much this in his CNBC interview below:

(Zoom to 1:30 for the relevant content)

“Look, startups have to be aggressive in terms of competing against incumbents who already have billions of users (Google Search), and OpenAI has 100 million users. We don’t have that today.”

OpenAI just killed my Startup

A common complaint from “AI” startups is that everytime OpenAI rolls out a new update, a few hundred startups die.?

Think about the first set of "thin wrapper" startups that was just adding plugins to OpenAI - such as the ones which can read pdfs and answer queries, or let you analyze data in excel and create dashboards.

Here’s the problem :? If your idea predominantly relies on an API - the company that offers the API has the IP and the moat, and you have none. The parent company can easily absorb the features of your service if it wants to.

But it need not even be the API company.

Remember that the same API is available to everyone else. If you have no moat anyone can (and will, if you are successful) duplicate it. And Perplexity got enough alternatives too - for example Poe (built by the Quora gang).

But what has Sam got to say on this??

In essence, he advises assuming that OpenAI will continually enhance their API, adding features galore. Your product plan should account for this relentless march of progress, lest you find yourself steamrolled by the OpenAI juggernaut.

Model Madness

Ah, but what about building your own foundation models, you ask?

Surely, that would secure a solid intellectual property advantage and grant you a competitive edge, right?

Allow me to burst that bubble with a hearty pop.

Mistral is apparently planing to raise 600M at a $6Bn valuation . If they do, then they are tripling the valuation in just six months - they raised at a $2Bn valuation in Dec 2023.

Does this makes sense? Is there still money to be made in foundation models?

This is the latest LMSys leaderboard which lists all the top LLMs. Mistral comes at #19, and that's the commerical one (Mistral-Large).

Even if you remove the duplicates of GPT, Mistral does not hit top 10.

The leaderboard itself is very unstable. While OpenAI has still very much managed to stay on top sinc ebeginning (depsite the short-lasting win for Claude 3 Opus), they are no longer as untouchable as they were before.

Anthropic has raised $8Bn so far at a valuation of $18.4Bn .

And let's not forget OpenAI itself, now valued at a jaw-dropping $90Bn.

I will go back to my analogy of search engines. There are many wannabes - but there’s only one with absolute dominance. And that is (surprise!) Google.

There is a consolidation coming for foundational models. There’s potentially 4 to 5 that will survive ( I am thinking one for each major cloud providers, plus OpenAI) - others will realize it costs too much to do R&D on these models and no way to monetize them - and fold shops or get acquired.

The Incumbents are investing heavily in keeping their advantage. Microsoft and Google will keep investing tens of Billions in this year after year.

No start-up can really compete.

If you are a startup wanting to build a foundation model now, the investors would think of you as crazy. Or delusional. Or both.

Does it mean that Mistral at $6Bn post valuation is not worth investing? Feels too rich to me, and I don’t even know if they have any competitive advantage anymore. But I am sure somebody assumed OpenAI was too expensive at a $6Bn valuation, while others invested. Those who invested in that valuation now have a 15x return.?

The point is: commercial models are highly commoditized now. There’s no guarantee that today’s winners will continue to be winners of the future. In the future, you might just pick a foundational model from your top service provider (Azure, GCP, AWS) and may not even think of much else.

And then there’s Zuck.?

The Zuck Strikes Back

Back in 1995, Netscape Navigator was a thing - web was just being discovered and Netscape had captured the most significant share of the browser market. Microsoft came up with a technically inferior Internet Explorer to compete - but it was free, and it was bundled with all windows computers.

Eventually Netscape just faded away. A good thing that the investors go their money though as AOL picked it up with 1998 for $4.2Bn.

Zuck is giving away Llama3 for free.

And unlike Internet Explorer, it's a technically competent foundaitonal model (it's up there in the Leaderboard).

He’s investing at a scale no startups can match to build AI capabilities. He ordered 150,000 H100s last year. He estimates that by end of 2024 Meta’s owning 600,000 GPUs. Nearly 350,000 of them are H100s. That's supposed to be a cool 14% of the H100s available in the entire freaking world.

Now, being the modest guy that he is, he only states:?

“We have built up the capacity to do this at a scale that may be larger than any other individual company”.

But that’s not just GPUs - it’s also the sheer amount of data they have. Facebook has more than 3 Billion monthly active users (MAUs) and guess where the data goes to train.?

And I repeat: he’s giving it away for free?(just out of the goodness of his heart - he has no business angle here ?? ).

So if some startups believes they can build something better than Llama3, they can hold Zuck’s beer.

The Thick Wrappers

Alright then: no Thin Wrappers, no Foundational Models.?

What else is left?

I would argue that the real value of AI comes from building applications. When your product solves a specific user problem end to end for a specific domain (i like how specific this gets ??).

And to do that, you need to really understand the industry, the pain points of the users, understand the user workflows, their tools and platforms, their language and jargons.

So, mostly a software solution that’s enabled by AI.

As per Sam Altman (i’m sumarizing this): “the long term differentiation will not be the model in itself. It will be the model most personalized to you and? best integrated to your life“.

In other words:

1. Find a way to integrate business data into the model
2. Find a way to implement a persistent memory into the model (I am not talking? of a lazy RAG implementation)
3. Give it agency ( at least a domain limited one)

In summary, as we navigate the evolving AI landscape, the focus should perhaps shift from chasing the capabilities of giants to carving out niches where technology meets genuine user needs in seamless, innovative ways.

Until next week...

We're still growing - this is the 16th edition! Your feedback stays as crucial as ever. Hit reply and let me know what you think! Want to see a specific topic covered next week? Don't be a stranger, share your ideas!

And of course, if you find this newsletter valuable, spread the knowledge! Share it with your network and help us grow. ????

See you next week!

#artificialintelligence? #generativeai #leadership #ai #productdevelopment #startups