Technology

Interesting what he wrote about LLMs' inability to "zoom out" and see the whole picture. I use Gemini and ChatGPT sometimes to help debug admin / DevOps problems. It's a great help for extra input, a bit like rubberducking on steroids.

Examples how it went:

Problem: Apache-cluster and connected KeyCloak-Cluster, odd problems with loginflow. Reducing KeyCloak to 1 node solves it, so it says that we need to debug node communication and how to set the debug log settings. A lot of analysis together. But after a while, it's pretty obvious that the Apache-cluster doesn't use the sticky session correctly and forwards requests to the wrong KeyCloak node in the middle of the login flow. LLM does not see that, wanted to continue to dig deeper and deeper into supposedly "odd" details of the communication between KeyCloak nodes, althought the combined logs of all nodes show that the error was in load balancing.

Problem: Apache from a different cluster often returns 413 (payload too large). Indeed it happens with pretty large requests, the limit where it happens is a big over 8kB without the body. But the incoming request is valid. So I ask both Gemini and ChatGPT for a complete list of things that cause Apache to do that. It does a decent job at that. And one of it is close: It says to check for mod_proxy_ajp use, since that observed limit could be caused by trying to make an AJP package to communicate with backchannel servers. It was not the cause; the actual mod was mod_jk, which also uses AJP. It helped me focus on watching out for anything using AJP when reviewing the whole config manually, so I found it, and the "rubberducking" helped indirectly. But the LLM said we must forget about AJP and focus on other possible causes - a dead end. When I told it the solution, it was like: Of course mod_jk. (413 sounds like the request TO the apache is wrong, but actually, it tries internally to create an invalid AJP package over 8kB, and when it fails blames the incoming request.)

LLMs are useful to provide generic examples of how a function works. This is something that would previously take an hour of searching the docs and online forums, but the LLM can do for very quickly, and I appreciate. But I have a library I want to use that was just updated with entirely new syntax. The LLMs are pretty much useless for it. Back to the docs I go! Maybe my terrible code will help to train the model. And in my field (marine biogeochemistry), the LLM generally cannot understand the nuances of what I'm trying to do. Vibe coding is impossible. And I doubt the training set will ever be large or relevant enough for the vibe coding to be feasible.

Note: this comes from someone that makes a (very good) ide which they only monetize with an AI subscription so it's interesting to see their take

(They use Claude opus like all the others so the results are similar)

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 done a test of 8 LLMs, on coding. It was using the J language, asking all of them to generate a chess "mate in x solver"

Even the bad models were good at organizing code, and had some understanding of chess, were good at understanding the ideas in their prompts. The bad models were bad mostly on logic. Not understanding indexing/amend on a table, not understanding proper function calling, or proper decomposition of arguments in J. Bad models included copilot and openAI's 120g open source model. kimi k2 was ok. Sonet 4 the best. I've mostly used Qwen 3 245 for better free accessibility than Sonet 4, and the fact that it has a giant context that makes it think harder (slower) and better the more its used on a problem. Qwen 3 did a good job in writing a fairly lengthy chess position scoring function, and then separating it into 2 quick and medium function, incorporating self written library code, and recommending enhancements.

There is a lot to get used to in working with LLMs, but the right ones, can generally help with code writting process. ie. there exists some code outputs which even when wrong, provide a faster path to objectives than if that code output did not exist. No matter how bad the code outputs, you are almost never dumber for having received it, unless perhaps you don't understand the language well enough to know its bad.

The LLM worship has to stop.

It's like saying a hammer can build a house. No, it can't.

It's useful to pound in nails and automate a lot of repetitive and boring tasks but it's not going to build the house for you - architect it, plan it, validate it.

It's similar to the whole 3D printing hype. You can 3D print a house! No you can't.

You can 3D print a wall, maybe a window.

Then have a skilled Craftsman put it all together for you, ensure fit and finish and essentially build the final product.

I think it's going to require a change in how models are built and optimized. Software engineering requires models that can do more than just generate code.

You mean to tell me that language models aren't intelligent? But that would mean all these people cramming LLMs in places where intelligence is needed are wasting their time?? Who knew?

Me.

Well, they will simply fire many and leave the required number of workers to work with AI. This is exactly what they will want to do at any convenient opportunity. But those who remain will still have to check everything carefully in case the AI made a mistake somewhere.

To those who have played around with LLM code generation more than me, how are they at debugging?

I'm thinking of Kernighan's Law: "Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it." If vibe coding reduces the complexity of writing code by 10x, but debugging remains just as difficult as before, then Kernighan's Law needs to be updated to say debugging is 20x as hard as vibe coding. Vibe coders have no hope of bridging that gap.

Clearly LLMs are useful to software engineers.

Citation needed. I don’t use one. If my coworkers do, they’re very quiet about it. More than half the posts I see promoting them, even as “just a tool,” are from people with obvious conflicts of interest. What’s “clear” to me is that the Overton window has been dragged kicking and screaming to the extreme end of the scale by five years of constant press releases masquerading as news and billions of dollars of market speculation.

I’m not going to delegate the easiest part of my job to something that’s undeniably worse at it. I’m not going to pass up opportunities to understand a system better in hopes of getting 30-minute tasks done in 10. And I’m definitely not going to pay for the privilege.