AI, as amazing as it seems, is still a buggy experience at many real world tasks.

Your choice is whether to consider the bug a limitation or make it a feature.

Choosing the latter requires you to leave the comfort of determinism for the creative joy and frustration of predictions.

And to redesign your workflows and mental models

for a very different type of compute.

My first AI experiences flip-flopped between enlightening analysis and sycophantic lies.

Tinkering with it, sometimes provided you a week of work from minutes of investment.

The closest to the feeling of using a credit card instead of saving up.

And as with credit cards, there's never a free lunch.

Initial AI-generated 'answers'

sometimes required endless hours of verifying and adapting to my actual needs.

A lot has happened since.

The AI product experience has gradually improved.

And users have become smarter about how to leverage it.

The first 'modern' phase of AI was all about scale.

Broad models fed with 'everything' from the internet, had an answers for any question you could think of,

but weren't always right

(as with the internet, by the way).

More data and more parameters gradually got them better, and AI got its normy a-ha moment with GPT-3.

The second phase went from scaled guessing to structured reasoning.

The products started to divide work into steps,

verifying themselves, and showing their work.

The O1 model crossed the line from guessing to reasoning.

The third phase is where we are today.

Models and applications melt together in verticalized tools, made for specific tasks and situations.

With a large share of the work happening in the harness rather than blindly depending on the model.

And, users has increasingly embraced the bug,

Integrating applications into their workflows, to provide their new tools with relevant context, and making sure to test and verify their work.

For many types of work, the new tools represent a truly exciting promise.

But it runs straight into a wall