Better AI chat interfaces (1)

Lately, I’ve been exploring different ways of interacting with LLMs, as the current interfaces don’t fit my workflow. There are many use cases: mine involves reading papers, books, tweets, then working with LLMs for brainstorming ideas and generating code. This process is not linear, I often need to go back and forth.

As I delve deeper into the chat session, I suddenly find myself lost in the enormous amount of information that I and the LLM generate. And as the content piles up, the experience starts to turn sour.

As computing becomes increasingly commonplace, the challenge lies in how easily we can work with AI to achieve what we want. Human-AI interfaces are a topic that has been left behind compared to significant strides in model training efficiency/performance—algorithms, GPUs/TPUs, etc., and for a good reason. But I also believe we need to start exploring better ways to collaborate with AI systems: to accelerate the acquisition of intelligence, the launch of new products, and the discovery of new ideas.

The main issue, which underlies all the features I'll highlight shortly, is the raw vertical layout of these UIs. It's the industry standard, but it's clear that beyond quick queries, they have a wide range of applications.

These are some features I’d love them to have (some of them were actually added!):

1. Diving deeper into a specific prompt

Branch out a particular item and keep exploring in parallel. This solves two problems: first, it avoids having many chat instances, and second, it keeps the main chat clean. My workflow often leads me to end up with very long chat sessions.

2. Collapsing and removing prompts

I'm always dealing with the scroll hustle. It's super frustrating to scroll up and down trying to find something I remembered. If our chat sessions get long, it'd be great to have the option to collapse prompts and remove any unnecessary ones. The goal is to keep our sessions tidy and easy to navigate.

3. AI explaining

I'm using fewer browsers these days. Honestly, if I had my way, I'd hardly ever leave an LLM interface. One step closer to making that happen is the ability to quickly look things up. If you come across words or concepts you don't understand, instead of writing another prompt just for a quick check, let the AI explain them to you. Can you imagine how cluttered the main conversation would get if we added all those extra prompts?

4. Getting the answer right away

Having the feeling of a real-time human-like conversation is one of the use cases for LLMs. But sometimes, I don't care about that and just want the output right away. I’d be cool if we could interrupt the token streaming. Currently, the downside is that we'd have to wait a bit longer since the model needs to generate the full answer. But over time, we can expect response times to drop.

5. Jumping between prompts

If our chat sessions get lengthy, having a map of the prompts can be super useful. It lets us jump between them easily and skip the hassle of scrolling through everything.

Another idea I've been toying with lately is to have an interface where the LLM gets a sense of what you're working on—whether it's coding, studying, summarizing, generating reports, having a casual chat with an AI companion, or just looking for a quick answer to something specific—and adjusts on the fly to suit your use case.

I’ll continue researching new/better ways of interacting with AI. I find it really exciting to explore this territory nowadays, as it’s a relatively uncharted one.

I'm curious to hear about your experiences and struggles with them.