[simonw]

I think "They wanted an engineer to build a chatbox that called ChatGPT with company documents as prompt context" fits the term "AI Engineer", personally - see https://www.latent.space/p/ai-engineer which uses it for "applied" AI.

[gavinray]

You don't call someone who integrates the Twilio API a "Twilio Engineer", or Mailchimp a "Mailchimp engineer"

Integrating third-party libraries to build an application is a significant chunk of the work in any SaaS product and the expectation is you can read the vendor docs and figure it out

[simonw]

I think the difference here is that it's possible to know bowering there is to know about the Twilio API. Read the docs, build a few things and you can consider yourself to have mastered that entirely.

Nobody on earth can tell you that they've "mastered" the art of building software on top of LLMs.

They're weird. They don't behave like other APIs. They're non-deterministic and unpredictable and not even the people who created them fully understand what they can and cannot do.

(For one thing, if someone claims to have mastered LLMs ask them how they would 100% protect against prompt injection attacks...)

[seattle_spring]

Why would a self-described "AI Engineer" be any more capable of building that sort of functionality over any other backend engineer, especially one who is familiar with agent-assisted development?

[simonw]

Because building on top of LLMs is really tricky. You need to understand things like writing evals, configuring agentic loops, creating and iterating on system prompts, designing tools that work well with LLMs.

It's a speciality, just like being a payments engineer who integrates with systems like Stripe is a speciality.

Being familiar with agent-assisted development helps a little bit because at least you understand prompts, but there's a whole lot more to building software on top of LLMs than that.

Any engineer can get familiar with these things of course, just like any engineer can figure out what it takes to work on payment systems.

[gavinray]

  > It's a speciality, just like being a payments engineer who integrates with systems like Stripe is a speciality.

At $PREV_JOB, we had physical Point-of-Sales systems as well as a mobile app, and provided multi-merchant marketplace functionality with things like disbursement reports and support for multiple bank accounts for vendors.

I had to migrate all of this from Braintree to Stripe. It probably encompasses the most complex payment system I've worked on in my career.

But that's not a job title, it's just part of "make the app work"

[simonw]

At my $PREV_JOB we would have called you a payments engineer for that.

I don't think AI Engineer is an exclusive job title. If anything, coding agents are pushing us all to become generalists much more so than before.