Dynamic Progress Bars, UnoReverseGPT & the new-age UX of LLMs

Dive into the first version of ‘The Sketch’ on Substack, exploring how fixed navigation opens up use cases, using strategies inspired by card games with ChatGPT and the new UX of LLMs.

🚀 Experience of the Week — Dynamic Progress Bars

New experiences push the boundaries of what we can allow our users to experience. Whether it be small touches to add delight, or larger value-adding features to improve user goals, there are always ways to enhance what we create.

This week, we’re talking about progress across the interwebs.

Who doesn’t love to see some progress? We all want to see some movement towards our goals, even in the micro and digital sense.

Enter, the Dynamic Island.

While Dynamic Island is akin to the Touch Bar, in that it was highly touted at the keynote but hardly used in reality, it does give a new fixed navigation point for experiences to leverage, which can be 22% faster to use 🤯.

This experience came via a tweet from Raul Dronca, where he shares an example of Dynamic Island being used to show reading progress for a page. This actual experience is not technically feasible, due to the sensors on the phone, so Raul came up with some alternatives.

We’ve seen different sites do some form of this, and blog posts often come with some showcasing of progress. TikTok, Medium and Zapier (amongst many other blog sections of sites) do some form of progress:

It’s a powerful feature that can incentivise users to complete the task at hand.

Think to yourself, where would my users want to know how far along they’ve come? Progress bars aren’t just reserved for onboarding journies, we can also use them in more minor applications.

✍🏼 AI prompt of the week — Negative market research

We’re very used to asking AI for recommendations in a positive way, and that helps us drive what we should do for a feature/idea that we have.

But equally, you can reverse that logic.

Next time you’re working on something, ask what is done poorly on the topic.

I did this recently while researching what Diversity & Inclusion pages on marketing websites are like:

Think about how you could UnoReverse ChatGPT the next time you’re doing research.

Ask things like:

“What are the weaknesses of this [content piece]”
”What do [industry competitors] do badly in user acquisition?”
”What [part of experience] is particularly frustrating or negatively viewed by users?”

Hit Open Ai with that +4.

🧠 Thought piece — Probabilistic UX of LLMs

LLMs (Large language models which are the foundation that AI tools like ChatGPT are built off) are leading to new products every week. It’s almost becoming normal how crazy fast new products are coming to our screens.

But that’s not the only thing it’s leading to.

In past applications and software, there was always a limit to the experiences a user could have. It’s what Computer Scientists & nerds all over the world call ‘deterministic’.

Deterministic = finite set of outputs.

When you hit a CTA, you’re going to get a limited number of outputs. Take a quote checker for example, the product team that would build the systems to generate a quote for you would only have engineered a limited set of scenarios. These would be something like:

1. Approved
2. Denied
3. Needs further documentation (or equivalent)

That’s all well and good for our usual deterministic experiences.

LLMs however, are not deterministic. They’re quite literally the opposite.

Probabilistic.

This means there aren’t a finite number of results that an input & subsequent CTA could produce.

It’s why when you give ChatGPT the same prompt twice, it can potentially come up with two different answers. It’s how DALL-E 2 comes up with multiple variations for the same text description.

And this raises an important question about experiences.

When we’ve all been used to deterministic, closed experiences, how do navigate users through ones that are the complete opposite?

If hitting the CTA doesn’t lead to exactly what you wanted anymore, how do we tailor the user experience for that experience?

Here are a few concepts I’ve come up with that could spark some ideas and help generate solutions

The image above visualizes a few different ways in which we could adapt to probabilistic experiences. These could come pre, before the user prompts, or post, after getting the result. Let’s break them down:

1. Pre-output prediction — the ability to give a confidence level similar to statistics, that would signal the chance they are going to get the response they’re looking for. This could be expanded further to give tips on how to increase confidence level, e.g. provide more context.
2. Pre-input direction — inspired by the Mirror #1 Mirror #2 days of pirating movies, this experience could give users the chance to point the LLMs in a certain direction, to give an answer in a specific style, e.g. harsh, casual or blunt. The possibilities are endless and adaptable.
3. Post-output feedback scale — let’s say you keep the same initial experience, but you create an iterate post-input experience, this would take more of the form of nurturing the LLMs, similar to how you nurtured your Sims when you were in school.

These are all of course very conceptual, but I hope it raises some new experiences to cater to our new world.

Thanks for reading, this was the first article of “The Sketch” newsletter, that you can subscribe to below 👇🏼

The Sketch | Sakky Baral | Substack