DALL-E/Every illustration.

Why Building in AI Is Nothing Like Making Conventional Software

Introducing Source Code, your backstage pass to Every Studio

68 6

Comments

You need to login before you can comment.
Don't have an account? Sign up!
Jennie Pakula 9 months ago

Really great post. Interesting to consider with AI for legal services - part of the feasibility risk is ethical, and it's likely that current regulatory frameworks around unqualified practice are not going to be able to properly capture the real risk profile here.

Edmar Ferreira 9 months ago

@JPak this makes a lot of sense and it will become more relevant as the model capability increases.

Oshyan Greene 9 months ago

Very good and timely (for me) outline of key considerations for builders in this new environment! I think for many people the pivot to "deep AI" will be infeasible but that doesn't necessarily mean the project is did. One of the exciting things about the AI revolution-in-progress is how quickly the baseline models improve *without* your input. So one reasonable option, given the pace of foundational model improvement, is simply to set the project aside for even a mere 3-6 months. There is a very real chance it will be viable when you pick it back up again, which is just insane and to my knowledge hasn't really been the case for cutting-edge tech pretty much ever (in terms of how long you'd have to wait for someone else to figure out the core problems you're facing).

There is also the key factor of foundational models being so highly accessible so immediately, which again has largely not been the case with other tech TMK, whether through genuine withholding (not explaining how your tech works, much less open sourcing it), or de facto inaccessibility (the new tech is hard to understand or make use of due to its complexity). AI has the incredible qualities of moving very quickly *and* being something that almost anyone can immediately make use of the moment it is updated. A new foundational model comes out and it's almost immediately exposed via the same or a very similar API (and chat interface) as you've already been using. New features like custom training, Embeddings, etc. take a little time to figure out to be sure, but are still incredibly accessible compared to many "open" but technically challenging platforms and systems of the past (try writing your own IMAP client for example).

But yes, the broader points of the article remain strong. This is just more color and nuance to the playing field.

Edmar Ferreira 9 months ago

@Oshyan "So one reasonable option, given the pace of foundational model improvement, is simply to set the project aside for even a mere 3-6 months. There is a very real chance it will be viable when you pick it back up again, which is just insane and to my knowledge hasn't really been the case for cutting-edge tech pretty much ever"

This is an excellent point! I have some projects on my backlog that I revisit from time to time to see if the models have become good enough.

Manoel Lemos 9 months ago

Thanks for writing that, Ed! It's a valuable piece of content and a helpful framework for everyone out there building AI-based products. Keep rocking, writing, and publishing! ;-)

@myfamilylovespaper 9 months ago

In a few months, I am going to embark on a project to build a knowledge bot in my company to assist the Product Managers. This frame is going to help to approach the project. I would have started with wireframes, like you said. Thanks for writing this.