October 14, 2024
Moats in GenAI Startups
First published in Oct 2024 on LI
7 months ago, when our tech and VC brothers and sisters in SV overnight became AGI bros and sis, I wrote quietly here that LLMs do not have Intelligence as they cannot Reason and Plan- two key tenets of Intelligence.
As usual, unpopular things get discarded by crowd. Our Sequoia crowd in Sonya Huang et al, even ended up writing thought articles on AI’s next act- Reasoning. Credit where credit is due however, her writing is articulate and well composed, underlying premise I naturally disagree.
7 months later, Apple drops the research on how LLMs don’t reason including new ones in o1, which are even influenced by minor irrelevant perturbation in prompt.
Took 7 months for a company to drop this for ppl to maybe start getting their act correct, but it was obvious back then when I wrote about it.
So, what’s the actual deal here?
As I wrote, in that post, LLMs are very cool things, but you need to understand where they are going to be hot in reality- UX/UI.
The real mass value opportunity in LLMs isn’t in the Infrastructure or Orchestration layer.
Zuck is one of the smartest businessman of our era. He will outsmart everyone trying to play in Infra layer. With LLAMA Stack, he will even try to create a dent in orchestration layer.
Your effort is not to compete in that with him.
The mass value of GenAI is in Application Layer.
There is a plethora of existing business and consumer problems that can really be disrupted how things get done using LLMs if you refrain yourself from falling for the hype on Reasoning and focus on UX/UI layer unlike what the current call is in Silicon Valley.
However, the real moat in application layer is not the myriads of Agentic workflow you can create. Any Tom, Dick and Harry can do that if they have spent some time in figuring out MoA/Multi Agent Systems. There is no moat there. If you have written code in that area off late, you will know there is ZERO moat in your Agentic architecture. Sooner or later, someone tinkerer will have it for themselves.
The true moat of GenAI application is in using Agents in solving a problem where you are able to add in the entire network. Your value is not in solving a specific user problem using Agents. Your value will be bringing the entire network on your product away from everything else out there.
Now I will be the first one to say openly that breaking the network is one of the most challenging task in life. Network don’t easily break if others in network don’t move.
However, there is one way you can break a network and have them move. If you provide enough value to majority of the network participants that their incentive no longer resides in staying and keep forming the current network.
If the value you are providing is markedly different than what the existing system and it’s network offers.
That’s when the Network breaks. Those ideas are gold!
