How to Use AI for Better API Creation

Last week I had the privilege of presenting to the 2023 CarMax/Edmunds internal developer conference. Like many people, AI has been at the top of my mind. The result of that concern and this speaking opportunity resulted in me delivering Separating AI Fact from Fiction for Accelerated API Development. Like nearly all of my talks, you can find the slides and script posted for free.

The presentation features a high-level overview of what these tools are, lists various ways in which they can be helpful throughout the software development lifecycle, and points out areas where users need to continue to be wary. It is anything but slight at more than 6,200 words chock-full of examples. There's real utility there, but user beware. In lieu of a full-fledged article in the newsletter, I'd encourage interested API and AI folks to check it out on my website.

'Where's the Prompts?'

A frequent question for this presentation (and any other discussion of generative AI) is "What was the prompt?" Sites like LearnPrompting.org promise "a completely free and open source" way of learning to use ChatGPT to "accomplish your goals".

I understand the desire - if we only knew the exact set of words, we'd be able to get the exact outcomes we're looking for. Our programming education creates a deterministic expectation among our toolsets. We expect our machines to only do what we tell them to, and - if we're to recreate the outcomes that we see in a presentation or quickstart, we need to follow the exact recipe or set of dance steps.

However, as I stressed several times in the presentation, generative large language models, like ChatGPT, are non-deterministic. This means that even if I use the same prompt as someone else, there is no guarantee that I will get the same answer. Furthermore, these systems are constantly being tweaked behind the scenes. Just because a prompt worked yesterday is no guarantee I'll see similar results today. For those accustomed to the vast majority of software systems till now, that ambiguity can be scary.

Successfully using a tool like ChatGPT is less about memorizing a set of commands. Rather, it is more about being comfortable iterating through the lossy medium of language. I attempt to show a process for breaking down, identifying, and reassembling a concept that ultimately gets me closer to that end state. That, thus far, seems far more durable than any set of 'magic incantations' that will only become obsolete in the future.

OK. Enough teasing. Separating AI Fact from Fiction is here.

Milestones

• Twitter has launched a new 'Pro' API plan for 'startups'. However, given how badly the 3rd party-ecosystem was treated, it remains to be seen if anyone volunteers to play Charlie Brown to Twitter's football-yoinking Lucy. Oh, and the plan is still $5,000 per month. <sad-trombone sound>
• For years, Darksky.net was my preferred weather source. That ended, however, on March 31st of this year when Apple ended both Dark Sky's delightfully elegant website and API. Into that void have emerged two wonderful homespun efforts: MerrySky.net and the Pirate Weather API. Both projects take weather data provided by government agencies and turn it into easily comprehensible, beautiful, and useful utilities. Pirate Weather replicates much of the Dark Sky API design, making it a possible drop-in replacement for the numerous API quickstarts and tutorials that relied on that API. If you are in a position, please consider supporting Guillaume Carbonneau (MerrySky) and Alexander Rey (Pirate Weather). And thanks to Dan Ciruli for bringing these great projects to my attention.

Wrapping Up

Thank you my paid subscribers. Because of their support, this newsletter remains free from ads, data collecting, or paywalls. These fine few mean the rest have freely available access to my various flights of fancy. They're awesome. Thank you.

That's all for now. Till next time,

Matthew (@matthew and matthewreinbold.com)