Thoughts on slowing the fuck down

Now that the pop media is finally letting go a bit of the topic "AI is the new X!", I'm starting to notice a few more high quality posts seeping through. This is one of them.

I really want to read people's perspectives on LLM's, it was just impossible to find quality when everyone wanted to give their opinion. This is the worst on LinkedIn, where mentioning AI gives you free "brownie points" (I have yet to figure out what Managers gained from this). I don't care what you use it for, unless you have a new perspective I can ponder over.

Regardless, nothing is black and white, and most things are a shade of grey. LLM's have been more positive leaning, making the CTA for working on something a lot simpler. Although, I end up refactoring my day away (which I am fine with, I quite enjoy putting the dots on the i's).

LLM systems at their core being probabilistic text generators makes them easily produce massive works at scale.

In software engineering our job is to build reliable systems that scale to meet the needs of our customers.

With the advent of LLMs for generating software, we're simply ignoring many existing tenets of software engineering by assuming greater and greater risk for the hope of some reward of "moving faster" without setting up the proper guard rails we've always had. If a human sends me a PR that has many changes scattered across several concerns, that's an instant rejection to close that PR and tell them to separate those into multiple PRs so it doesn't burn us out reviewing something beyond human comprehension limits. We should be rejecting these risky changes out of hand, with the possible exception when "starting from scratch", but even then I'd suggest a disciplined approach with multiple validation steps and phases.

The hype is snake oil: saying we can and should one-shot everything into existence without human validation, is pure fantasy. This careless use of GenAI is simply a recipe for disasters at scales we've not seen before.

At which point you've gained very little efficiency in most large organizations given that by the time you're actually doing development work at the ticket level 90% of the project timeline (identifying issues, prioritizing, creating requirements, architecture, ticket breakdowns, coordination, etc) has already passed.

If AI can enable engineers to move through the organization more effectively, say by allowing them to work through the service mesh as a whole, that could reduce time. But in order to evaluate code contributions to any space well, as far as I can tell, you still have to put in leg work even if you are an experienced engineer and write some features which exposes you to the libraries, quirks, logging/monitoring, language, etc that make up that specific codebase. (And also to build trust with the people who own that codebase and will be gatekeeping your changes, unless you prefer the Amazon method of having junior engineers YOLO changes onto production codebases without review apparently... holy moly, how did they get to that point in the first place...)

So the gains seem marginal at best in large organizations. I've seen some small organizations move quicker with it, they have less overhead, less complexity, and smaller tasks. Although I've yet to see much besides very small projects/POCs/MVPs from anyone non-technical.

Maybe it'll get to the point where it can handle more complexity, I kind of think we're leveling off on this particular phase of AI, and some headlines seem to confirm that...

- MS starting to make CoPilot a bit less prominent in its products and marketing - Sora shutting down - Lots of murky, weird, circular deals to fund a money pit with no profits - Observations at work

It's really kind of crazy how much our entire society can be hijacked by these hype machines. My company did slow roll AI deployment a bit, but it very much feels like the Wild West, and the amount of money spent! I'm sure it's astronomical. Pretty sure we could have hired contractors to create the Chrome plugin and Kafka topic dashboard we've deployed for far cheaper

I've been working on some parts of this problem, specifically capturing and retaining other semantically useful layers of the systems we build as we build and maintain them.

By introducing progressive semantically enriching layers (starting with prose, reasoning and terminology and going all the way into specifying interaction surfaces), we can reduce the dark matter between spec and code, make code more disposable – if your semantics live in the spec layer rather than the implementation, you can throw away and regenerate the implementation without losing understanding – and, critically, give LLMs a way to navigate a graph of knowledge instead of gobbling up walls of text.

https://clayers.com -- https://github.com/CognitiveLayers/clayers

The productivity gains are somewhat real in a sense, but are not really about "moving faster", as the hype would have us believe. GenAI agentic systems instead boost individual developer "efficiency" by allowing a single, reasonably qualified developer, to approximate an entire software team. As those developers, however, we're still required to manage the workload of those teams and ourselves to ensure quality output, just as ever before.

The problem is that it's VERY easy to overload oneself with the output of these new tools. Human comprehension is the bottleneck, as much as it always has been. Anyone that tells you otherwise is shilling for these companies.