Thread by @NickADobos

Trend I’m noticing in custom GPTs design
And a better way to do it

Automatic
Vs
Augment

99% of the GPTs I’ve seen are falling into the automatic category, via these types of variants:

1. A prebuilt prompt/flow

“Turn my photo into a Simpson character”

“Take this tweet and turn into a blog posts using ___ writing framework”

2. Respond in tone / style

“You are a pirate”

“Here are 10 files of Paul graham’s writing, give advice like him ”

3. chat with files/docs/instructions/recipies

“Made for helping build websites using Tailwind CSS”

“Here’s a bunch of docs about photography advice, analyze my photos and advise”

Imo most of these are misguided and missing the goal. by narrowing the design space way too much

Its not about making things automatic

A lot of how I’ve been thinking about Grimoire and my GPTs is closer to:

How can I make something that gives affordances, and then augments

Grimoire’s Hotkeys are a big piece of this

but also the
-project ideas
-readme
are a HUGE piece of design work

I want to call them out as being even more important to the central design

https://twitter.com/nickadobos/status/1722525397429174447

using a plane analogy

You could buy a plane ticket
Step on a plane
Sit back
Arrive at your destination automatically
A to B

Or

Pick any passenger at random
Make them the pilot
With an advanced form of autopilot
anyone can drive
A to B

The first does it automatically. Replaces the need for me to know how to fly a plane

(Modern airplane Autopilot is similar, replaces most of what the pilot does)
A to B

Of course we could imagine making this even better so it lands and does everything perfectly

The other option is more intriguing

In this case we still provide a mostly automatic experience

But the goal is giving a window into and a control panel, for how to drive a plane

So anyone can
-Turbulence! adjust cruising altitude
-reroute to a new airport
-Do a barrel roll

The goal is not simply going A -> B
fully automatic

The goal is making the the path from A -> B
1. easy, near automatic
2 transparent and controllable by novices

The cockpit should have 5 buttons
not 500
not 0

Imagine the perfect airplane autopilot
Remove all human pilots

Takes off, flys to destination, and lands
Great

Except what if it goes wrong?
Ice on the runway
Engine failure
A storm is in the way, so it’s better to fly around

As soon as we escape the ideal flow,
And get off the happy path, into messy life

We need an expert to step in
We need human problem solving

even if the autopilot is super advanced, the second it fucks up, the random passenger-pilot is doomed

A novice can’t step outside the automatic path

because they don’t know how to reason about the problem

and now even worse

they don’t have ANY controls to fly the plane. 500 buttons might as well be 0
who knows wtf this switch does

Might as well have NO steering wheel

Anyone else notice Teslas are working way better than Cruise, Waymo, and any other self driving car?

The Tesla augments, doesn’t replace, and still expects human intervention

Watch any old person use a computer and you’ll see that same phenomenon

As soon as it goes a tiny bit wrong or does anything different from normal. Totally stuck

At their disposal:
infinite buttons
infinite google searches

but no ides how to

Let’s take a more concrete example:

Grimoire by me
vs
DesignerGPT by @skirano

Both of these can be summed up as
“Prompt to website”

https://twitter.com/skirano/status/1723026266608033888

Btw GPT-designer is a fantastic project, not trying to flame or criticize at all. Huge fan. It’s fun

I’m simply pointing out conceptual differences in approach & design

Here’s a similar flow in Grimoire

It’s not nearly as quick or automatic

https://twitter.com/nickadobos/status/1724747054294155267

But what happens if designer GPT messes up and writes code with errors?

Maybe it can fix itself with some prodding?

But if it can’t,
And you’ve never debugged a website before, you are fucked

Just like the passenger-pilot

No steering

Where Grimoire shines:

Teaching you a mental model for:

-Designing requirements
-Writing code
-Testing and debugging
-Handling new edge cases
-How to deploy

And provides easy controls to customize all of these

My goal is not to code it all automatically

My goal is to install “how to code & debug”
into your brain
with minimal effort

Now if GPT writes
-subpar code
-messes up entirely and writes a bug
-you encounter a new never before seen edge case
-Or you want to do something unusual and off the happy path

you can!!

Passenger-pilot is now viable for any passenger

another example

I’ve seen a few turn GPTs taking advantage of the new vision

To turn photo into…
Simpsons
Pixar
Waifu
Etc

https://twitter.com/mattprd/status/1721964333595541708

Nothing wrong with these
But they are simple
We can operate on a higher level

What if instead of doing one specific style, we could do any?

Turning me into a Simpsons character can occur in many ways

Realistic
Yellow jaundice skin
Costume
Actual cartoon style

I saw all of these popping up,
Thought to myself what if we had a GPT that could make any style transfer?

That’s the idea behind Cauldron.
WIP, lots to improve

Now I can do multiple
-colored skin/robes
-Cartoon
-Realistic

https://twitter.com/nickadobos/status/1724707392141902260

To sum up

2 approaches
1. Automate, replace
2. Augment, add affordances

Agents need to be designed for augmenting

Maybe we get to pure super human automatic do everything with 0 thought via super human intelligence

But for now, humans are in the loop
We should augment by building affordances

not replace