The last 6 to 12 months of open models has pretty clearly shown you can substantially better results with the same model size or the same results with smaller model size. Eg Llama 3. 1 405B being basically equal to Llama 3.3 70B or R1-0528 being substantially better than R1. The little information available about GPT 5 suggests it uses mixture of experts and dynamic routing to different models, both of which can reduce computation cost dramatically. Additionally, simplifying the model catalogue from 9ish(?) to 3, when combined with their enormous traffic, will mean higher utilization of batch runs. Fuller batches run more efficiently on a per query basis.

Basically they can’t know for sure.

LLMs have made it really clear when previous concepts actually grouped things that were distinct. Not so long ago, Chess was thought to be uniquely human, until it wasn’t, and language was thought to imply intelligence behind it, until it wasn’t.

So let’s separate out some concerns and ask what exactly we mean by engineering. To me, engineering means solving a problem. For someone, for myself, for theory, whatever. Why do we want to solve the problem, what we want to do to solve the problem, and how we do that often blurred together. Now, AI can supply the how in abundance. Too much abundance, even. So humans should move up the stack, focus on what problem to solve and why we want to solve it. Then, go into detail to describe what that solution looks like. So for example, making a UI in Figma or writing a few sentences on how a user would actually do the thing. Then, hand that off to the AI once you think it’s sufficiently defined.

The author misses a step in the engineering loop that’s important though. Plans almost always involve hidden assumptions and undefined or underdefined behavior that implementation will uncover. Even more so with AI, you can’t just throw a plan and expect good results, the humans need to come back, figure out what was underdefined or not actually what they wanted, and update the plan. People can ‘imagine’ rotating an apple in their head, but most of them will fail utterly if asked to draw it; they’re holding the idea of rotating an apple, not actually rotating the apple, and application forces realization of the difference.

I’ve used it most extensively doing Ruby on Rails greenfield apps, and also some JS front ends, some Python mid sized apps, and some Rust and Nix utilities. You’re absolutely right about it struggling with code base scale, I had to rework the design process around this. Essentially, design documentation telling the story, workflow documentation describing in detail every possible functionality, and an iteration schedule. So the why, what, and how formalized and in detail, in that order. It can generate the bulk of those documents given high level explanations, but require humans to edit them before making them the ‘golden’ references. Test driven development is beyond critical, telling it everywhere to use it extensively with writing failing tests first seems to work best.

So to actually have it do a thing I load those documents into context, give it a set unit of work from the iteration schedule, and work on something else.

It does go down some seriously wrong paths sometimes, like writing hacky work arounds if it incorrectly diagnosing some obscure problem. I’ve had a few near misses where it tried to sneak in stuff that would bury future work in technical debt. Most problematic is it’s just subtle enough that a junior dev might miss it; they’d probably get sent down a rabbit hole with several layers of spaghetti obscuring the problem.

I’m a professional developer and have tested AI tools extensively over the last few years as they develop. The economic implications of the advancements made over the last few months are simply impossible to ignore. The tools aren’t perfect, and you certainly need to structure their use around their strengths and weaknesses, but assigned to the right tasks they can be 10% or less of the cost with better results. I’ve yet to have a project where I’ve used them and they didn’t need an experienced engineer to jump in and research an obscure or complex bug, have a dumb architectural choice rejected, or verify if stuff actually works (they like reporting success when they shouldn’t), but again the economics; the dev can be doing other stuff 90% of the time.

Don’t get me wrong, on the current trajectory this tech would probably lead to deeply terrible socioeconomic outcomes, probably techno neofeudalism, but for an individual developer putting food on the table I don’t see it as much of a choice. It’s like the industrial revolution again, but for cognitive work.

Most if not all leading models use synthetic data extensively to do exactly this. However, the synthetic data needs to be well defined and essentially programmed by the data scientists. If you don’t define the data very carefully, ideally math or programs you can verify as correct automatically, it’s worse than useless. The scope is usually very narrow, no hitchhikers guide to the galaxy rewrite.

But in any case he’s probably just parroting whatever his engineers pitched him to look smart and in charge.