The Front End Theory
I know I have been talking a lot about AI lately, but I have something brewing in the space, so expect to hear a lot more. I have a couple things brewing, actually. But I can’t talk about it yet. Hopefully next week.
If you’ve been reading along, you know my thoughts on how much of the economy is padded with various bureaucracies, and that technology both enables said bureaucracy by raising the ceiling of what is perceived as possible (letting middlemen in to “explain” it to the bottom ranks), as well as disassembling it through democratization of knowledge. Nothing is binary, everything is multimodal. For every action there is an equal and opposite reaction. Jobs get automated away in one sector and opportunities pop up in another. However, in our modern socio-economic hivemind, things are binary, AI is either threatening, or revolutionary.
Right now there is a common theme running through speech, crypto, business, education, and it has extended to AI. The theme is that less is somehow more. Speech is able to be restricted, but without explicitly knowing why, or by whom. Transactional data is not wanted on chain, because anonymity is illegal. Businesses use adjusted-EBITDA to inflate enterprise value to shareholders. Schools want to get rid of AP programs to homogenize learning. AI is up against a cohort of people who want to decelerate progress to protect us from the potential dangers of a runaway AGI.
These are large societal issues that I am not going to tackle on a Saturday morning in Los Angeles. I feel like that’s more of a Sunday morning, NY and SF kind of thing.
Specifically with AI, the decelerationists have a few different attack vectors, which are conveniently the same vectors that founders and VCs are trying to weight out in terms of value capture. Last week I posited the bottom-up pyramid of Compute, Model, and Apps, stating that much of the value capture would be lower on the stack, but that Apps would be the sticky spear to drive into the hearts of consumers (in a cute Valentine-y kinda way, not a violent way, you weirdo).
TBD how that all plays out. The further down the stack you go, the more centralized it becomes, but also more commoditized and moated. Compute is tied inextricably with energy and the silicon needed to manufacture semiconductors. Attacking this vector makes no sense barring regulation on who gets to consume said value down the chain. It’s also incredibly hard to make headway into this space as an entrepreneur. High barrier to entry, high barrier to regulate.
The LLMs (Large Language Models) are set to consume an enormous amount of this energy and silicon via compute, but so is everybody else. LLMs also have a high barrier of entry as the amount of money you have to spend on engineering and compute to build your own model from scratch is non-trivial. Forget about the fact that you would also be years behind, and each passing day these things are increasing exponentially in value. The value of LLMs is obvious to anyone who has used ChatGPT. It just works. So why is it all open source? Why has OpenAI given their API away for essentially free (the latest API now offers GPT3.5 turbo for the cost of about 350,000 words for $1)? Because they know that this is rapidly going to become a commodity. They know that the LLM industry will become balkanized, Bezos’s Law will come into play, and margin will get competed away. Even when Facebook tried to release their proprietary LLM, LLaMA, it was promptly leaked online. As I have written about before, the open-source ethos of many of these developers is inherently anti-centralization. They do not believe that AI should be closed-source, or proprietary. This makes it difficult for both the LLM creator to generate revenue from its end user, as well as regulators to step in and say what can or can’t be imported or exported from a model. All of that is not to say that LLMs can’t self-regulate, as OpenAI has done with their Trust and Safety protocols. Clearly this is done as a precaution to avoid the target on their back, as right now they are monopolizing the market. High barrier to entry, low barrier to regulate.
You would think that regulating LLMs would actually be extremely difficult from a technical perspective, but the complexity actually makes it easier to regulate. It doesn’t matter what’s going on inside the black box, just make sure that what gets spit out is kosher.
Now when we get to Apps, there is absolute chaos. Firstly, the barrier to entry to create an app is extremely low. The API is near free, there are github repositories for the initiated, and Google Sheets extensions for the un. Secondly, there are already a tremendous amount of non-AI apps already in existence, just waiting to be “enhanced”. The legacy companies here have a tremendous advantage because they already have a captured audience, recurring revenue, and provide a service. Adding an AI function to your product suddenly makes you an AI company! This will proliferate faster than you think. I am a big Notion user, and they have already added it seamlessly into their (already seamless) interface.
“Press ‘space’ for AI” is easily the best AI product integration I have seen thus far.
Seamless user experience is a must, but it has to have functionality, and above all else it has to work. The Notion AI integration is great, and it works, but it actually doesn’t do much more than ChatGPT would do if you copy/pasted your Notion page into a thread and gave it a prompt. It’s cool, fun, interesting, useful, and bespoke, but it’s not a game changer. Prepare for a lot of that.
Since everyone’s going to be tapping into the same OpenAI API in the near term (Facebook’s LLaMa turned out to be vastly inferior to GPT3.5), the front end of the application becomes of massive import. It’s not what LLM consumers are using, but how they are accessing it. If another LLM comes along that’s better, developers can point their front end in that direction. Low barrier to entry, high barrier to regulate.
I’ve discussed in previous articles that it would not be the AI itself that would be the tipping point of monumental value unlock, but the bespoke data embedded in that AI. Clearly, the way to attract users to your front end is going to be the unique experience that you are able to provide them. Once they are there, the product must be easy to use, solve an immediate problem, be sticky, and maybe even a bit social.
Bespoke embeddings are going to be the cutting edge of AI user adoption. ChatGPT saves each of your individual chat threads as unique “chats” that the AI keeps in its memory. Notion saves the entirety of each page. This makes each chat thread or Notion page its own unique, private LLM that no one else in the world can see or interact with.
Thus, the maximum value that a user can derive from an LLM is capped only by the creativity of the user itself.
I’ll leave you with that.