Once you have a chat feature, you become a chatbot. Once you’re a chatbot, you’re a wrapper in competition with every frontier model developer, permanently at risk of having your lunch eaten by the next zealous engineer, drinking from a pool of infinite compute.

Nicolas Bustamante’s (@nicbstme) recent article on software moats in the age of AI is the latest to do the rounds of the AI discourse hypercycle. I thought the article was valuable, but there’s one argument in particular that I think deserves closer scrutiny.

Specifically, that natural language interfaces are prima facie superior to the application-specific interfaces and workflows that characterized vertical software up until The Current Moment™

In short: I believe the chat interface is appealing, and for good reasons (among them: its tight relationship with the underlying technology, and the notion of text as the universal human interface). BUT it suffers from a number of drawbacks that prevent it from being all-consuming in the way @nicbstme supposes.

Subscribe now

Why the chat interface?

1. Chat is general-purpose.

Chatbots can take nearly the entire corpus of human language as an input, and are incredibly flexible as a result.

This is basically the argument of Text is the universal interface, paraphrased by @tszzl in the below quote.

“The language model works with text. The language model remains the best interface I’ve ever used. It’s user-friendly, composable, and available everywhere. It’s easy to automate and easy to extend.”

2. Chat is a lindy interface.

Conversation itself is, in some sense, the oldest technology. Chat-based interfaces are a particular instantiation of this technology, used widely across text, instant messaging, search engines, social media comment sections, and more.

It’s thus an incredibly lindy form factor, which ~100% of users will be familiar with, and that offers close to zero learning curve.

3. Chat is a natural medium.

Current frontier models are predominantly large language models, trained on large amounts of language. Natural language is where they excel.

These are all totally legitimate motivations for adopting a chat-based interface. For reasons I describe below, they also drag firms in a direction that is likely worse for their product, and is almost certainly worse for their defensibility. Let me explain why.

Against the chatbot

I see four main limitations to chat as the universal interface: typing, conversation, chat as all-or-nothing, and chat as a pure substitute for labor.

1. Typing is a bottleneck.

Typing is simply too slow. Average type speed is around 40-50 words per minute. A good typist might hit 70-80. By contrast, the average spoken word speed is closer to 150. This is still considerably slower than our ability to process speech, at around 400, and pales in comparison to estimates of the speed of thought, at over 4,000 words per minute.

Speed can be improved with speech-to-text features (ChatGPT’s voice mode, or apps like Wispr Flow) but these are still limited as:

1. It’s usually inconvenient to speak out loud in quiet places, or be heard in noisy ones.

2. At its fastest, speech is still sluggish relative to thought.

So the problem remains: any form factor that requires the user to type out a fully formed thought before acting simply cannot compete with one that allows the user to just do the thing.

2. Conversation is linear, knowledge work is networked.

A conversation with a chatbot is always a scrollable stack of prompt-response pairs, each barely visually distinguishable from the next, proceeding until the model loses coherence.

For even mildly complex topics, this is a painful format. Real knowledge work is messy: ideas branch this way and that, looping back on themselves, in ways that a two-way conversation simply can’t represent in an accessible way.

Conversation is inadequate, even relative to other fairly rudimentary means of representing information, like an Obsidian-style graph.

Popular products try to address this with features (search, targeted replies and branched chats being popular ones), but these too become unwieldy quickly, and fail to manage the complexity of the work we regularly engage in.

3. Chat is an all-or-nothing medium.

If you give the user any ability to interact using natural language, you have two options:

1. Permit any topic of their choosing (see: a chatbot) OR

2. Place heavy restrictions on the acceptable inputs and outputs, feeding the user rejections if they ever stray off topic (“Sorry, I’m unable to answer that…”).

We’ve seen how awkward the former can be, even for more general-purpose applications, and the latter can only ever create a frustrating product.

I think of this quote from Will Ahmed, about his company Whoop and their fitness wearable:

“Once you have a screen you tell the time, and once you tell the time you’re a watch. I didn’t want to compete with watches.”

Once you have a chat feature, you become a chatbot. Once you’re a chatbot, you’re just a wrapper in competition with every frontier model developer, permanently at risk of having your lunch eaten by the next zealous engineer, drinking from a pool of infinite compute.

4. Chatbots are strict substitutes for humans.

It would be unfair to say that LLMs do not increase the output or efficiency of their users, but chatbots are strict substitutes for human labor along one dimension in particular. That is, with respect to time.

At any one point in time, either the LLM is working OR the human is typing, but each must wait until the other is finished to continue.

This creates a kind of ping pong pattern of work, where 100% of the labor is performed by EITHER the human OR the model, and the two never achieve the true complementarity of a carpenter with his hammer.

Maybe in the endgame this won’t matter. In the not too distant future, the entire economy may comprise swarms of AI agents, acting and interacting at the speed of inference, in their own unnatural language. And maybe the gains from advanced AI models are so great that they simply dwarf the inefficiencies of the medium (I think this is reasonably likely).

But the issue remains that chatbots will have you flipping back and forth between windows, or staring at a model-in-thought, while you wait to issue your next command.

Subscribe now

A great product feels like auto-complete

In his paper Agency plus automation, Jeffrey Heer uses spell-check and auto-complete as products that are illustrative of a set of design principles and patterns. Principles that, if followed, can square the circle of a natural language-forward technology with a world that demands more sophisticated product design.

“First, they provide significant value, promoting efficiency, correctness, and consideration of alternate possibilities that a user might not have otherwise considered.

Second, they augment, but do not replace, user interaction: These aids blend into the interactive experience in a nondisruptive manner and can be directly invoked or dismissed.

Third, these augmentations require neither perfect accuracy nor exhaustive modeling of the user’s task to be useful (e.g., consider the subtask of “spelling” vs. the more general goal of “writing”).

Fourth, through interaction, both people and machines can incrementally learn and adapt (e.g., autocomplete may alert a search-engine user to relevant related topics, while other user queries may train the system for future recommendations).”

Thinking through these in turn, the performance of the current crop of LLMs is middling to straightforwardly poor.

1. They can certainly provide significant value and promote efficiency. Instances of ‘incorrectness’ (see: hallucination) have mostly been resolved.

2. As shown above, chatbots do replace user interaction. This is their dominant mode. They are pure substitutes with respect to time.

3. Most current LLM products do require (or at least benefit from) “exhaustive modeling of the user’s task.’ That is, these products work best when the user describes the task and desired output in excruciating detail. Some of this can be scaffolded away (e.g. with skills), and will likely diminish as the underlying models improve. For the time being, it remains a hurdle, one that reduces their value relative to 1. Humans 2. Better designed products.

4. Memory, skills, and other features allow the tools to learn and adapt over time, but this learning still feels limited in some fundamental sense, and most improvements in the product come overwhelmingly from the models themselves.

There is a vision for an alternative product, one that addresses the weaknesses of the chatbot: the image model.

Take for example, Photoshop’s Generative Fill tool. With Generative Fill, the user can simply click and drag to expand a photo in a particular direction, and Photoshop will fill in the space with newly-generated images.

Not happy with the result? Select the area of the image you’re not happy with, generate variations, and choose the one you like best!

This is but one (relatively limited) use case. Nonetheless, it shows a different way of thinking about generative AI tools. A different design pattern that does not require exhaustive modeling of the task (the user doesn’t need to type out “extend X element of the image in Y direction”, they just do it).

A form factor that augments, but does not replace human interaction.