I started out very sceptical. When Claude Code landed, I got completely seduced — borderline addicted, slot machine-style — by what initially felt like a superpower. Then I actually read the code. It was shockingly bad. I swung back hard to my earlier scepticism, probably even more entrenched than before.
Then something shifted. I started experimenting. I stopped giving it orders and began using it more like a virtual rubber duck. That made a huge difference.
It’s still absolute rubbish if you just let it run wild, which is why I think “vibe coding” is basically just “vibe debt” — because it just doesn’t do what most (possibly uninformed) people think it does.
But if you treat it as a collaborator — more like an idiot savant with a massive brain but no instinct or nous — or better yet, as a mech suit [0] that needs firm control — then something interesting happens.
I’m now at a point where working with Claude Code is not just productive, it actually produces pretty good code, with the right guidance. I’ve got tests, lots of them. I’ve also developed a way of getting Claude to document intent as we go, which helps me, any future human reader, and, crucially, the model itself when revisiting old code.
What fascinates me is how negative these comments are — how many people seem closed off to the possibility that this could be a net positive for software engineers rather than some kind of doomsday.
Did Photoshop kill graphic artists? Did film kill theatre? Not really. Things changed, sure. Was it “better”? There’s no counterfactual, so who knows? But change was inevitable.
What’s clear is this tech is here now, and complaining about it feels a bit like mourning the loss of punch cards when terminals showed up.
[0]: https://matthewsinclair.com/blog/0178-why-llm-powered-progra...
Core to Ptacek's point is that everything has changed in the last 6 months. As you and I presume he agree, the use of off-the-shelf LLMs in code was kinda garbage. And I expect the skepticism he's knocking here ("stochastic parrots") was in fact accurate then.
But it did get a lot of people (and money) to rush in and start trying to make something useful. Like the stone soup story, a lot of other technology has been added to the pot, and now we're moving in the direction of something solid, a proper meal. But given the excitement and investment, it'll be at least a few years before things stabilize. Only at that point can we be sure about how much the stone really added to the soup.
Another counterfactual that we'll never know is what kinds of tooling we would have gotten if people had dumped a few billion dollars into code tool improvement without LLMs, but with, say, a lot of more conventional ML tooling. Would the tools we get be much better? Much worse? About the same but different in strengths and weaknesses? Impossible to say.
So I'm still skeptical of the hype. After all, the hype is basically the same as 6 months ago, even though now the boosters can admit the products of 6 months ago sucked. But I can believe we're in the middle of a revolution of developer tooling. Even so, I'm content to wait. We don't know the long term effects on a code base. We don't know what these tools will look like in 6 months. I'm happy to check in again then, where I fully expect to be again told: "If you were trying and failing to use an LLM for code 6 months ago †, you’re not doing what most serious LLM-assisted coders are doing." At least until then, I'm renewing my membership in the Boring Technology Club: https://boringtechnology.club/
This was actually the only point in the essay with which I disagree, and it weakens the overall argument. Even 2 years ago, before agents or reasoning models, these LLMs were extremely powerful. The catch was, you needed to figure out what worked for you.
I wrote this comment elsewhere: >>44164846 -- Upshot: It took me months to figure out what worked for me, but AI enabled me to produce innovative (probably cutting edge) work in domains I had little prior background in. Yes, the hype should trigger your suspicions, but if respectable people with no stake in selling AI like @tptacek or @kentonv in the other AI thread are saying similar things, you should probably take a closer look.
Sure, but I would argue that the UX is the product, and that has radically improved in the past 6-12 months.
Yes, you could have produced similar results before, manually prompting the model each time, copy and pasting code, re-prompting the model as needed. I would strenuously argue that the structuring and automation of these tasks is what has made these models broadly usable and powerful.
In the same way that Apple didn't event mobile phones nor touchscreens nor OSes, but the specific combination of these things resulted in a product that was different in kind than what came before, and took over the world.
Likewise, the "putting the LLM into a structured box of validation and automated re-prompting" is huge! It changed the product radically, even if its constituent pieces existed already.
[edit] More generally I would argue that 95% of the useful applications of LLMs aren't about advancing the SOTA model capabilities and more about what kind of structured interaction environment we shove them into.
But I think my other point still stands: people will need to figure out for themselves how to fully exploit this technology. What worked for me, for instance, was structuring my code to be essentially functional in nature. This allows for tightly focused contexts which drastically reduces error rates. This is probably orthogonal to the better UX of current AI tooling. Unfortunately, the vast majority of existing code is not functional, and people will have to figure out how to make AI work with that.
A lot of that likely plays into your point about the work required to make useful LLM-based applications. To expand a bit more:
* AI is technology that behaves like people. This makes it confusing to reason about and work with. Products will need to solve for this cognitive dissonance to be successful, which will entail a combination of UX and guardrails.
* Context still seems to be king. My (possibly outdated) experience has been the "right" context trumps larger context windows. With code, for instance, this probably entails standard techniques like static analysis to find relevant bits of code, which some tools have been attempting. For data, this might require eliminating overfetching.
* Data engineering will be critical. Not only does it need to be very clean for good results, giving models unfettered access to the data needs the right access controls which, despite regulations like GDPR, are largely non-existent.
* Security in general will need to be upleveled everywhere. Not only can models be tricked, they can trick you into getting compromised, and so there need to even more guardrails.
A lot of these are regular engineering work that is being done even today. Only it often isn't prioritized because there are always higher priorities... like increasing shareholder value ;-) But if folks want to leverage the capabilities of AI in their businesses, they'll have to solve all these problems for themselves. This is a ton of work. Good thing we have AI to help out!