Only Three Kinds of AI Products Actually Work

In general, it feels like the early days of the internet. LLMs have so much potential, but we’re still mostly building copies of the same thing. There have to be some really simple product ideas that we’ll look back on and think “that’s so obvious, I wonder why they didn’t do it immediately”.

Almost all AI products are just chatbots (e.g. AI-powered customer support). These suffer from having to compete with ChatGPT, which is a superior general product, and not being able to use powerful tools, because users will be able to easily jailbreak the model.

But scrolling feeds has become the primary way users interact with technology in general, so the potential here is massive. It does not seem unlikely to me at all that in five years time most internet users will spend a big part of their day scrolling an AI-generated feed.

Agentic products are new, and have been wildly successful for coding. It remains to be seen what they’ll look like in other domains, but we’ll almost certainly see domain-specific research agents in fields like law. Research agents in coding have seen some success as well (e.g. code review or automated security scanning products).

The technology behind current human-generated infinite feeds is already a mature application of state of the art machine learning. When you interact with Twitter or LinkedIn, you’re interacting with a model, except instead of generating text it’s generating lists of other people’s posts. In other words, feeds already maintain a sophisticated embedding of your personal likes and dislikes. The step from “use that embedding to surface relevant content” to “use that embedding to generate relevant content” might be very short indeed.

The reason agents work and chatbots-with-tools don’t is the difference between asking an LLM to hit a single button for you and asking the LLM to hit a hundred buttons in a specific order. Even though each individual action would be easier for a human to perform, agentic LLMs are now smart enough to take over the entire process.

This kind of product is usually called an “AI assistant”.

This doesn’t work well because savvy users can manipulate the chatbot into calling tools. So you can never give a support chatbot real support powers like “refund this customer”, because the moment you do, thousands of people will immediately find the right way to jailbreak your chatbot into giving them money. You can only give your chatbots tools that the user could do themselves - in which case, your chatbot is competing with the usability of your actual product, and will likely lose.

chat is not a good user interface. Users simply do not want to type out “hey, can you increase the font size for me” when they could simply hit “ctrl-plus” or click a single button

For the first couple of years of the AI boom, all LLM products were chatbots. They were branded in a lot of different ways - maybe the LLM knew about your emails, or a company’s helpdesk articles - but the fundamental product was just the ability to talk in natural language to an LLM.