Millennials and the Magic Wand
Sam Altman - the YC guy? Before there was OpenAI, there was Loopt; before there were LLMs, there was ML; and before there was chat, there was the magic wand.
2010s were a vibe, money was free and the jeans were skinny. The magic wand in a way accurately captured the mood of the decade. You’d digitize a niche, sprinkle a Bayes classifier, slap the wand label and raise a million dollar seed in the name of ML. Quick and lazy, and it worked until it didn’t.
Co-Star and the First Seamless “AI”
I am sure there were many, but the first place I saw an LLM being seamlessly integrated was in the co-star. Why am I, as a man in technology, using co-star? Go figure.
But that doesn’t change the fact that the app never bragged about AI, leaned into it’s mystic brand and quietly charged $3/question. Or better said, the feature never overshadowed the product.
ChatGPT and the Water Rising
ChatGPT standardized chat. I remember the 6 months between GPT3.5 and chatGPT. The possibilities were endless. But once users tasted the ability to prompt at will, it was all anyone was talking about.
And chat is a very good universal interface. Anyone can use it. Every app now needs a fully functional chat interface, which can practically replace the UI.
With MCPs, the architecture is also painfully simple and horizontally scalable. You spin up an LLM as the central controller and expose APIs through MCPs. Of course it’s more nuanced, and there is a whole intuition on when to use Agents vs Workflows, but that’s for a different post.
But it’s Not Enough
Everywhere chat shines as an interface is precisely where it falls short:
- Humans Don’t Think Linearly
- Your Power Users Deserve to Be Treated Better
The Linear Chain of Thought And How to Fix It
Humans don’t think in a neat line. In a real conversation we branch, backtrack, and pick up dropped threads. A single chat forces users to juggle context or drown in it. Memory helps, but it’s a band‑aid.
At Kognitos, I am experimenting with nested chats and pop up dashboards. Example: an analytics thread fetches data; a “Create graph” button spawns a mini‑chat dedicated to the chart. Users can tweak the graph there, then hop back to the main thread. Thought becomes a map, not a rail.
The Power‑User Problem And How to Fix It
Chat’s uniformity is also its ceiling. Noobs move as fast as experts. But experts crave speed. Muscle memory trumps natural language for repeat actions. Quick, invisible motions are how we level up. Great UI encodes those shortcuts and makes the experience delightful (think iPhone’s haptic timer).
TL;DR - Invest in HCI/pay that designer.
In Closing
The space is evolving and the models keep getting better, but I am almost certain that Specialized User Interfaces are not going away. Art might become more interactive. All these blog posts are really motivated by my experiments with how art and apps might evolve and intersect. More on that later.