Hazy AIPA

I still remember the first time I had a hazy IPA. Brewing at a lager-focused brewery at that time, we prided ourselves on not brewing an IPA at all, an act of hubris to be sure though it did help tell a story. We focused on traditional beers, priding ourselves on technique. For those who haven’t brewed beer before, just know that we brewers worked real damn hard to get beers clear, especially our IPAs where you’re desperately trying to get all that hop matter to drop out.

IPAs were already something of a weird subset, our cousins who just got hoppier and hoppier by the day. And now you’re telling me we’re not even gonna let the dang things drop clear?

I was open to trying it, though perhaps snidely was excited at the idea of not liking it. At proving us “purists” right.

And then I had one.

And it was absolutely delicious.

Hazy IPAs took over, and we came to learn that making good ones was also incredibly difficult. That keeping a stable haze wasn’t easy. That there was just as much science that went into making a perfect hazy as there was making a perfect lager.

The purists said it’d be the end of craft beer.

They said the same thing about Kettle Sours.

And about Milkshake IPAs.

And then the darndest thing happened, lagers actually became in vogue again.

It’s an eerie time to make the comparison I’m about to make given that the craft beer industry is struggling mightily. But as a former craft brewer turned software engineer and now CTO, my mind can’t help but see similarities between the launch of hazy IPAs and the explosion of AI ever since its big bang moment back in 2023.

I was weirdly at similar points in my career for both. At this juncture where I was transitioning from the beginning of my career in those fields into this point of having more of a voice, taking on some more leadership, and forming strong opinions on the right and wrong ways to do things. I had thoughts on the perfect dosing amount of biofine to help a beer drop clear. On the exact temperature to ferment a Hefeweizen at to get more banana phenols than clove.

I was finally getting confident in my coding skills and now here’s this thing threatening the entire paradigm. I was desperately trying to learn more, diving headfirst, worshipping at the altar of Uncle Bob and others, and now everything is changing?

With AI it wasn’t quite the same lightbulb moment for me as that hazy IPA was. Using ChatGPT to help troubleshoot was helpful, but it wasn’t what using Codex or Claude Code is today. It was never lost on me, though, that then, just like now, it was the most primitive it would ever be.

Now it’s everywhere, much like hazy IPAs, with people who never coded before enjoying it for the first time and many hardened engineers begging the world to see it for the existential threat it is. A lot of what’s put out there is bad, much like many of our first attempts at hazy IPAs, broken apps people are shipping prematurely. Horrible CI/CD practices making any future fixes on an app seemingly impossible. Tools claiming to be AI that are the same products companies always offered just rebranded.

But a lot of it is beyond good. It took me a while to begin trusting Claude Code, my current tool of choice, but the reality is I do trust it now. I’ve set up structures around it. I’ve become religious about plan mode and reading all output. I have ample documentation for both me and my team to follow and our coding tools are forced to reference it too. But, in the same breath, it’s hard for me to imagine a world without it now.

That scares me far more than it comforts me, but it’s true. The things I’ve had it help me build are tools I am genuinely proud of, and they were done way faster than I could’ve possibly done on my own. In a role where I was the only developer until very recently, the ability to have essentially a fleet of junior developers on call who I could deploy at a moment’s notice was invaluable (certainly worth far more than my subscription cost) and it’s only getting better. Even the “junior developer” framing feels disingenuous these days. The good ones aren’t junior anymore. They’re whatever you are, but faster, and with a deeper bench.

But it’s not magic. The speed with which it outputs code means it is uniquely talented at exposing all your faults, personal or systematic. Bad assumptions surface in minutes instead of weeks. Sloppy architecture compounds in real time. The brewery comparison holds here too: the breweries who made great hazy IPAs were the ones who were also good at making incredible lagers. They had their systems and processes dialed in, which allowed them to take full advantage of this new exciting thing.

AI is no different in my mind.

If you have rigid structures in place. If you run tests. If you have a deployment pipeline. If you have strong version control. If you have style guidelines that everyone is on the same page with. If you have strong oversight.

If you have all of those things, then you can make incredible beer, and you can make incredible things utilizing AI.

If you don’t, you still may stumble onto an incredible product, but you also may make a batch you have to dump entirely. You may have an infected batch you don’t catch until it’s too late. Your ability to be consistent and safe evaporates.

The existential threat of it is much harder to quantify, and much more far-reaching, than beer. It’s changing everything, and it’ll continue to. What that means on the other side of this is impossible to know, and I am in no way trying to handwave that away.

What I am saying is that I was an idiot when I handwaved away hazy IPAs or Kettle Sours or any of the other beer trends that hit while I was still in the industry. That I was short-sighted when I thought it’d be a long time before OpenAI’s products would be anything more than high-end bug fixers.

The other side of this very well may be horrible. But it might also be really damn good. And the people best positioned to find out are the ones who already know how to make a clean lager.