Raising model children

I recently saw a user send a message in the Focaldata platform, a qualitative research AI agent, and it read the following:

No I want you to use the words EXACTLY as I've written them, DO NOT change what it says on the discussion guide.

I could feel her frustration quite viscerally. I've sent many a message to an AI tool along these lines.

Back on December 20th 2024, when I was using Cursor in my early days and 'vibe coding' as a term hadn't been coined yet, I would be up until 3AM asking the AI to stop rewriting my middleware code, after multiple requests.

At the time, if anyone had been watching me, they'd see the frustration of requesting something so logical not playing out as it should.

But digital systems are complex. AI-powered digital systems are even more complex. But that complexity doesn't exist in the way software complexity used to exist before. It's not a language to learn necessarily, it's like a child to raise.

The Deep Mind team felt this whilst they were training the very early AI systems to 'walk' in a digital playground.

Now you have members of Cursors team telling you which model is best for what use case, like which player should play which position in basketball.

I want a tall & lanky Tim Duncan type of model in the post to just pop in the shots - this is consistency.

I want a Steph Curry model in the corner dropping ungodly dimes - this is creativity.

I want a Lebron model that is versatile in most situations - this is realiability.

I'm not making predictions here, this is the current state of these models. There is not a single one that does everything you want, well.

And maybe that's fine. We used to go from website to website, searching, reading, analysing.

Now we're going from model to model.

We will have raise multiple model children.

There's these nuances that you see when you grow a little human being, and you begin to capture those in your mind subtly, knowing when to push vs. let go vs. give up vs. go to another model.

You can't feed them too much (context), you can't expect them to always tell you the truth (hallucinations) and you can't delegate completely to them (irresponsibility).

Our models today aren't adults, and I wonder if they'll ever become adults.

If they do, maybe then we'll get AGI, but if they don't we'll just get another piece of technology that improves the world.

I've never raised a child personally, but I've grown things from 0. I've seen nature grow things from 0.

A little at a time, 1% every day, that's the key and that's the vibe.

Too much too soon and crank a bone's broken, your trust in the system deteriorates and god-forbid you become 'Anti-AI'.

Treating these things like children helps me reframe my expectations a bit, so that I don't expect Carlos Alcaraz 2022 when I'm really getting Alcaraz at age 2. Many more reps to go.

The funny thing is that we're seeing all these children being born at the same time.

Just last week I used Figma's 'Prompt to design' AI and it is the best AI design execution I've seen.

It's fairly fast on existing designs, and is excellent at generating UI from scratch.

What used to take me 30 mins to come up with a few options as a designer is now taking me a 30 second prompt and a 30 second generation. That's a 30x boost.

If anyone gives you a 30x boost on what you used to do, you're taking all the rough edges that it has around it.

The thing with design is that in the majority of cases, we're not reinventing the wheel. So it's just about finding the right wheel that already exists elsewhere and applying it to your situation.

As I improved my craft in design, I would spend more and more time researching for things. The time I spent researching depended on how much I cared and how much was on the line.

A revamp for a new client? Max research.

A simple request from an engineer? Take the first example you don't hate.

But now, I just get AI to come up with an implementation, to start with. Again, I'm not a purist that thinks you should be doing this with your own brain.

For most of us, most of the time, the job is the job. There isn't much joy in it, there isn't much of a big deal in it, it's just execution of a task that's been requested. Fairly transactional, fairly straightforward.

Being able to lean on AI for that is quite insane. So now the work for everyone switches from delivering work, to raising the new child on the block in the best way. Working with that child's capabilities and limitations, and sometimes waiting for the child to level up mentally and have their age and hippocampus increase.

The following are my favourite use and don't use cases:

| Works well ✅ |

Generating designs from scratch with minimal additional context Create me a simple login screen with our brand colours

Creating relevant copy for the context, instead of Lorem Ipsum Give me this but for X demographic

Giving a single task to output repeatedly Claude project to draft LinkedIn posts

Generating things where the finer details don't matter yet Creating a first draft of a design Creating a prototype for user testing

When creativity is the most valuable thing Creating a custom image of characters in a new setting

Single long-horizon goal where you don't care how it reaches it Deep research report on how new flight routes get determined

| Doesn't work well ❌ | Generating complex endpoints, without directly accessible documentation from scratch Create me a cron job to run every day and pick an article from Arxiv that's relevant to my job, and then send me a PDF with Resend

Giving an entire artboard to read and transform Translate this long landing page into a security page with the following content

Switching models for the same implementation Letting Grok code start structuring the codebase, then changing to Claude

I've enjoyed raising our model children in the last year. There are so many things that exist today that did not exist when I started doing this in September 2024:

• Claude code - prototype generation will no longer be the same
• Figma AI - creating design screens will no longer be the same
• GPT 4o image gen - Image generation will no longer be the same, and Gemini took it to another level
• NotebookLM - Audio learning will no longer be the same
• Wispr Flow - Writing things down will no longer be the same

I look forward to seeing our model children's personalities develop into bratty teenagers.

To the future generation of models