Are you using AI?

Ooch. I can feel your annoyance/pain. Not many get to do what they love in a corporate workplace, and now AI has come along and disrupted matters. Reading through this post I can now appreciate the significant disruptive nature of AI in business due to commercial pressures. It seems a common trait now where code ownership and creative joy of coding is lost, and replaced with the trudge of just “managing” these tools do a job. Getting things out the door seems more important than getting it right these days. Interesting that you've at least seen some of AI’s strengths in pattern-finding and exploration (that's me too). Let's hope we figure out how to use AI for our own benefit and that AI is not used purely to enhance greed. If it facilitates the parts of the work we actually love then great, if it's there just for churn, not so great.

I wanted a new kind of like for your post. I don't like what you have described at all, but very much agree with you.

We seem to have gone in a circle  - I was brought upon the notion that you can't test quality into a design (software or hardware) but via the route of  test driven development and AI we seem to being pushed into a regime where the vast team of AI monkeys bash out rubbish code and squeeze it through a sieve of tests. While we stand by and hope it'll work.

Not for me - I like to know what every line in my code does - even if it takes me longer to write. (But I'm not sure that it does - because the AI costs so much in chips and energy etc that once you factor that in - is it really any more productive ?)

MK

I've been using AI for a few month now, mainly tinkering with lots of caution, carefully reviewing what commands it runs, what files it touches to be sure it does what I tell it, and nothing more. I've been in "programming" for more than 30 years and I have quite an opinion on what is acceptable and what is not, hence the careful approach.

But it appears I'm quite an exception here as we've had a two day workshop at work to help us reach the 20x speed gains purported by many advocates in the field. The consultant doing all the talking was adamant that my approach is no more: "craftsmanship is dead" was his motto, parading amongst our desks.

Apparently, the "human in the loop" is just there to act a a shepherd to a flock of agents, that are writing both the code and the tests that verify it directly from the ticketing system or the written specifications. And the warden is just there to proofread everything at the end, getting the hang of the code in minutes, just like we've been doing code review for human created content.

But with the amount of code generated it's very strenuous to achieve this, there is a very heavy cognitive load. And for me, it's worse, I love writing the code, trying various options, so that I am really confident the code I wrote is the best possible approach, after having justified rejecting the others. Doing code review has always been a pain point, even if it has its undeniable use, human interaction is quite fruitful.

Sadly, this AI usage means I'm left with doing only the thing I do not like: managing "other minds" and "reviewing" their work, when they have the attention span and propensity to lie of bored teenagers: if you give strict rules to match your desire for clean code, it either loops indefinitely or straight up ignores them!

It's not all dark in there, though, because the LLMs are quite formidable at finding correlations (not causation) in lots of data, finding patterns that I would have had a hard time identifying. So for initial exploration of an issue, providing suggestions, it's quite great.

What's bothering me most, is that it's all built on a giant theft. I mean, I kinda of regret taking part in so many open source projects where the premise was to share the knowledge with fellow humans, which the license I released it under was quite clear, but now we must pay to get back the knowledge we gave for free over the years!
Or even worse, I gave for free the basis for destroying my own employment. For those versed in history, the revolt of Canuts in Lyon or Luddites in the UK should ring a bell, with, I fear, the same sad outcome.

And all this is even ignoring the economical and environmental sustainability part. We are collectively getting hooked up on a "drug" that is (almost) free today as a "tryout product", but when the dealers will start to recoup their costs, I'm worried many won't be able to afford it anymore with unforeseen circumstances.

In the end, I'm starting to seriously look at what I like doing which would be valuable to others to pay me decent money for it, and that would be quite safe from this AI monster lurking in the shadows. Maybe I should ask a LLM, after all these are just gain function optimization platforms, so it should have ideas for me to solve this equation and escape its grab...

I am using AI for work (writing software). I've found quite a lot of variability in the models with Claude series working best for my platform C# .Net / React / K8s. Have seen some subtle bugs slip in and have had moments where I felt I was fighting rather than colaborating. Typically take an iterative approach rather than leaving it to its own devices. Some interesting discoveries/approaches have been seen so it has proven useful as a learning tool.

On the whole, software has been written and quality/productivity has not suffered.

in the past when working on a old domestic device if its audio there is usually someone that has a manual. 

if its industrial if the owner doesnt have a manual, then sit down and nut out how it works 

then check sensors or relays and if boards have capacitors that have dried up

then if any moving cables that could have broken.

After that make a flow chart of events  

AI can do lots of things but they don't shorten design time or improve designs - yet.

AI can select components, but it takes a lot of work to get it to select an optimal component and then how can it prove it did a good job without showing you hundreds of datasheets. Sure it can organize the data sheets, with enough coaching, but a parametric search engine is going to compete favorably with AI at this task.

AI can generate a netlist, but how can it prove it is a good design? Do you have to translate the net list to a schematic and PCB layout and build the card and test it to know if it was hallucinating? Verifying the design from a netlist introduces far more problems than it solves.

I haven't seen  anyone show that AI is good at designing circuits. Presumably it will happen at some point. 

lol sensored the title of a story about a big white whale.

Aw man, theres so many things I would love to see AI trained on. 
Id like to say, hey AI, read my PLC program, and tell me where I have 36 contiguous registers free. or, optimize this ladder logic. or find and clean up unused registers. or read this engineering manual, protocol guide, or spec sheet, and return functions, values, whatever data, that pertains to the specific model hardware we use. Getting AI to read tech documents is nearly impossible to begin with, every OEM has sort of random formatting. A common theme is that every model in a product line is in one document. so then youre reading that document looking for functions pertaining to your specific hardware, and you have to carefully parse which functions track to which units, which makes designs and implementations hard. 

My main goal is process automation to maximize production, though for the last 9 months Ive been on sensors and logging of every type of thing.. 
while systems to do certain processes exist in the generic sense, even if IEC61131-3 exists, it doesnt mean you can utilize it in a practical way. PLCs have been out for decades, and many, new or old run the same code that ran 40 years ago. But their niche, regarded as such, theyve existed for so long, yet the documentation is persistently awful. PLC design patterns. Theyre not PCs, even if they operate similar to PCs. I have never been able to get good design tips out of AI for PLC systems. 

so AI wont / cant read the published docs, and to be fair, many of the published docs are garbage. 
theres many possible solutions, and opportunites for AI to succeed, but its all dependent on specific OEMs.

Theres other things that annoy me about AI, on the brink of being useful, but not there yet. 

Take Andriod studio for example. Andriod Studio now has integrated gemini AI. But the integrated AI doesnt know anything about andriod studio. its a generic code bot. Theres no features to it. The thing could do all sort of amazing operations, but its all generic. Hey AI, teach me how to use Android studio... then it could highlight menus, or help navigate the complex code it generates. but its not really aware of the platform it is in, which is infuriating, because both are Google products. 

I was using confluence today, and its the same sort of thing. I wrote an article on how to do a process. (Ive wrote many) did I write it in my personal articles, or the company articles? I'm not sure. Ask ROVO. Rovo tells me how to search. user = currentUser() ok thats cool, but it doesn't search keywords, or ignore case, or basic text search, so it has AI, but basic functionality is missing. I wrote about a bug in Rpiboot, I know the keywords, but can't recall the exact title, or project that the paragraph was published in.  We have this AI with access to all our company documents, and it cant do a proper search. 

I feel like its that way with many AI implementations, there's the low hanging fruit, that makes it interesting, but then where it could be legitimately useful, those features are not implemented, or may even be missing, because everyone assumes the AI implementation can do anything. 

it goes on and on like that. Part of it is AI, but part of it is the system integrator swapping out basic functionality for AI. For AI its context, but for a human user, its a half baked product. 

Theres an example I like:
you ask an AI did Captain Ahab die in Moby ***?
The current NLP paradigm is something like read the entire book into memory and parse it. 
which is horrendously inefficient. The answer, is on page 1. 
but thats not a smart systems approach. What could be done, is the book moby *** is a Canonical Novel. Books of this genre follow the pattern of having a beginning, middle and end, going from character intro, plot development, Climax, conclusion. 
being a lazy kid in high school, who doesn't want to read the whole book, but knowing the type of book, we could assert, that if moby *** is 200 pages, that the climatic section of that book would be around pages 180, and read that. but this is an intelligence assertion against a known structure.  That's what I dont see in AI. Maybe its there. I, the user of AI can not impart any of my knowledge onto the AI. Instead I have to repeatedly craft linguistically detailed instructions to get results. which from an automation perspective is pretty insane. 

You can of course see similar attempts at max lazy being deployed in AI. its just not smart about it. you upload a 2000 page pdf, and say hey parse this. AI will read the first few pages and try to answer your questions. Being sentient, you notice this. so you break a 2000 page document into 100 distinct PDFs. suddenly the AI can return better results. But again, its all things like tricks to leverage the system to compensate for a general lack of intelligent design. if the AI could find chapter 37, and read the 3 pages there, and not be confused or miss lead by sparse parsing.  then maybe it could be more useful.

Its good some time , and some time it  can mislead  , so better to to start with minimalist code and use ai to expand it according to need

Nice, thanks for the screenshot/explanation, I was wondering what the VSCode implementation would look like. 

I just learned GitHub Workflows works nicely when Gemini contributes to that too; I know Jan Cumps uses the Workflows feature, but I've only used it for one project, since it takes quite a while to create a decent workflow file. With Gemini, I just repeatedly pasted any errors into the text window, and it kept correcting things one by one.

For anyone not familiar with GitHub workflows, it relies on a specific file that instructs GitHub how to (say) test the code, or to build it, and it can even package up or zip up the final binaries and place them in the Releases page so that users don't need to build the code themselves if they don't wish to.

I'm now at up to 80% free Gemini quota usage, so realistically the (presumably daily) free quota is sufficient for perhaps half a day of work when working slowly, or a few hours of work when working intensively.