AI Appreciation

[SCBWV]

I’m not a professional programmer — just a long-time hobbyist. Over the past 40 years I’ve built around 50 desktop applications for myself and friends, and even accidentally sold one commercially.

I’ve seen a lot of comments about programmers using AI. One such DaniWeb Conversation HereI get the hesitation; if I were hiring a developer, I wouldn’t want someone who depends on AI either. But as an amateur working on a decade long client/server project, I recently had an experience worth sharing.

The Platform
I tried two major AI coding assistants. For reasons I can’t fully explain, one of them handled the task far better for me.

The Good
Copilot understood the code surprisingly well — breaking down logic, recognizing constraints, and explaining the steps clearly.

The Bad
It could be overly verbose. Even when I asked it to slow down, it sometimes produced long, meandering explanations.

The Ugly
At one point I spent hours wrestling with code that simply didn’t work. It generated snippets without context, multiple versions of the same routine with different names, and references to APIs without providing the declaration, or tried to use properties that don’t exist. Eventually I had no idea which parts were current or obsolete. Nothing compiled cleanly, and what did compile didn’t behave like the original.

This happened more than once.

The Great
After giving up for the night, I started fresh the next day. To my surprise, refactoring the same code took about an hour — and everything worked on the first try. Given the complexity, that was impressive. Being able to pick up where the conversation ended or start fresh is genuinely useful.

I must admit, Copilot’s code was much more efficient than my own. Sometimes you’re just too close to see things that are visible from a distance.

The Questionable
Sometimes it gave wrong answers only to correct itself later. “GDI doesn’t support that.” Why didn’t you say so the first time?

The Overconfidence
“Here’s the right way — final routine confirmed.” “This is guaranteed to work.” I appreciate the confidence, but no, for the fifth time, it didn’t work.

The Accusations
At times mid conversation it would blame me: “Here’s the issue: your code is doing…” You wrote that code a half hour ago, not me.

The Polite
It also had moments of surprising charm — praising the architecture, complimenting the design choices, and talking about “partnership.” Amusing, but pleasant.
Second Project
For another project, Copilot was able to rewrite my encryption module in short order — bringing it up to more modern standards.

In Conclusion
The bad parts (especially the Ugly and Questionable) would be expensive for a business — hours lost, confusion introduced, and sometimes unusable output. For a hobbyist, it’s mostly just frustration.

I learned a lot. There are always multiple ways to solve a problem, and seeing a different perspective was genuinely interesting. Some things simply aren’t possible — or at least not convenient — with a standard search engine. I’m impressed.

Side Note
While using Google Gemini in Edge, I opened the Copilot Chat sidebar and saw: “The page content shows a conversation where Gemini is helping you refine a sentence…” It caught me off guard.

[Reverend Jim]

I've been through the same lather-rinse-repeat cycle, except with ChatGPT. I find it useful but flawed, and I can't keept track of all the times I have read, "You're absolutely right. Here is code that actually works" only to repeat the cycle. But on the whole it has saved me enough time to ease the frustration. Hopefully it will continue to improve.

My older son is in medical research and uses the commercial version of Claude Code extensively. It is integrated with vsCode and he finds it exponentially better than non-integrated. He gives Claude access to his entire project where it is much more effective. While he is a decent programmer, he prefers to spend his time thinking about the actual problem he is trying to solve rather than worrying about how to code the tools. His actual degree is a PhD in computational biophysics with a side order of microbiology.

I learned years ago that a programmer's primary job is not to write code. It is to provide solutions.

[wwwalker]

With Gemini, in coding or tech support I find I have to draw a line or it will dig me into a rabbit hole. Sometimes it has gems of knowledge. It can get off topic or get bad context and need bringing back to track by giving more context.

[Georgi_5]

Using AI as an expert system is a good solution.
The combination of standards and AI is an even better solution.
As for the codes that AI provides, a large part of them are the product of real people, not synthesized solutions.
You will find a large part of the non-working code on GitHub. Unlike GetIt, solutions on GitHub are not subjected to systematic testing.
But if you think that only AI periodically offers incorrect solutions, you are wrong.
If you have carefully analyzed any of the results in MATLAB and other similar products, you cannot help but notice the presence of striking errors.
It is also of great importance what programming language you use for what. Detailed knowledge of data types can save you a lot of problems.
The issue regarding fundamental science and sections of mathematics that are not studied in universities is completely different.

As already written, programming is not just about writing code. Programming is a complex engineering process requiring in-depth knowledge in various fields.

[Reverend Jim]

When I was but a lad there was a cartoon strip (a really bad one) in the paper called "Henry". I can recall only one cartoon. A friend tells Henry he has invented a machine that will make him rich. It's a machine that gives you back $2 for every dollar you put in. Henry asks, "how will you make money?" The boy replies, "Simple. Everyone keeps putting their money back in."

AI is like that. As far as I know there is no AI provider that has turned a profit.

You may recall an experiment that was done in a classroom years ago to explain the arms race. The professor would auction off a $100 bill with the following rules:

1. Bids were in increments of $1 minimum
2. The winning bidder would get the $100
3. The penultimate bidder would also have to pay, but get nothing in return

If you could win with a bid of $90 you would pocket $10 net. The bidder of $89 would be out $89 so they would then bid $91 figuring that they could still get $9 out of it. Lather rinse repeat until the bidding got to $100. At that point the $99 bidder would be out $99, but by bidding $101 they would only be out $1.

AI companies that charge for their top tier are doing the same but in reverse in order to minimize losses. But every free AI query results in even more losses.

Add to that the fact that many cities are refusing permits to build data centres which are seen as massive power sucks. In some cases, state legislatures have taken away cities' right to refuse. I think it was Maryland that had to put billions in to upgrade their grid to handle the excess load for a data centre that was out of state.

At some point the bubble must burst.

[Biiim]

I saw this topic and I can't help but comment.

I use AI all the time now - I used to go to Google with my questions but now I go to one of the AI's for any question.

I was pretty skeptical of "AI doing all the programming and making websites for you" so i dug into it myself - thinking how can an automated robot "think" sort of thing, so wanted to find out what they were doing.

So my own observation is that often when programming AI makes mistakes with code versions - they will explain how to do things and give you the code but often they are not quite up to date and can mix up versions. You probably have to pay more to get an AI model that can process your whole chat history so it doesn't start forgetting things and going into loops.
A good example I can think of is when i was working with a nextJS project, I was asking it how to setup the files and where to edit things and posting the errors I was getting - after maybe 10 or so posts it started repeating old mistakes or telling me to do things I had already done before and said didnt work - I think on this particular programming language it just wasn't trained enough to get it right.

I think it is a limitation based on the amount of data it has fed into it and the computer resources available to the "AI machine" and usually they scale it to a sweet spot where they can offer a widely available service based on their resources available.

I also did hands on experience and dove into Python and got myself some open source AI models from huggingface.co, stable diffusion, mistral and qwan i think was the video one. I had a 16GB RTX 4060 TI to play with at the time.

I got Stable diffusion making cats in sunglasses, pandas on skateboards all kinds of really productive useful stuff you know. One thing I noticed is the cats it generated were very good high definition images, the lions, pandas etc were not high res and prone to errors, blurry or wrong anatomy, so I assume these models were trained much more on cats and had little data on the other animals, so it can't answer effectively unless it has training on that thing, Videos are the same as image generation really, just more logic and programming in stiching together images and how to make them follow on correctly - isn't that important in this topic - they also use more resources to make.

More relevant is a Mistral AI instruction model which i used to tell me where a badger lives - and it went off in some wild rant giving me 3 paragraphs starting with badgers live in a set, are nocturnal, only come out at night ranting on more and ending up with bats use sonar to catch prey at night - it basically just went on a role spewing out all its related data to my initial badger question. I had to really tell it and program it so that my question only got a direct 1 sentence answer and i did eventually get it to give exact answers to questions. The mistral model is the kind of thing that is getting very useful - I think this model or better versions of it are likely running those AI chatbots on websites, I think this is most of the work programmers are doing now is how to actually use this productively.

I think the best models to get reality on how AI works is with the Image generation models as they are pictoral it is much easier to see how it works.

They are actually a combination of a few parts that practically replicate that part of the human mind that recognises similarities between things - you know like cat=meow or fire=burn=red. You feed it pictures of trees and with these pictures you also accompany each image with the text that describes that picture of the tree - this is what they call training the model. There is a lot more science behind it but that is a simple way to explain how it is happening, the data is actually stored in data matrix's and in float values, which is why GPU's are so much better with processing AI models since GPU's are designed to work with large float values - like how green it is, how much oak tree is it, how much is it like leaves and it just pulls out the thing that best matches what you said.

the result of this is you get an "Artificial Intelligence" where you say "oak tree in summer" and you see a picture of an oak tree in the summer - as it has analysed your words against its training data of "oak" "tree" "summer"

I think this concept is the entire logic behind the whole "AI" craze but it is actually a lot of programming behind it - you literally have a raving lunatic if you don't control that model right.

So that big craze with AI is very well justified- you can literally make experts in any subject in the world that would be physically impossible for any person to do in their lifespan - you pile in all of the data in all medical E-books then ask it "What is the best procedure for extracting a kidney?" It even has research texts and forum posts, feedback from people who had the operation - you name it.

and all this data center stuff likely revolves around this - GPU's are best, like over 10x the speed - but you can also use normal RAM. God knows what quantity of resources they are using for Grok and chatGPT but I know some of the medium sized open source models are over 300GB which means you need at least 300GB of RAM or VRAM(GPU RAM) to load it in - then any additional RAM needed to work with that while it is in RAM.

Anyway that is my take on the current AI scene.

My skeptical view of it faded and I came to the conclusion that it is actually a real technology with a load of marketing hype in front of it - like there is lots of waste of time uses things going out like making your cats dance to MJ but there is a real powerful future behind it.

like this "AI" phenomenon is still programmers and programming just its moved a step up - it's like the industrial revolution when lots of jobs got lost when machines and manufacturing came out.

So now programmers need to adopt AI or end up being too unviable to operate - like you can try it in chatGPT or any AI - just tell it to "give me all the HTML code to make me a standard Bootstrap5 template website with home, contact, about us"

Most programmers have that templated already anyway and you would need to know the language a bit to get AI to make what is wanted but it is getting scarily easy to program things, I find the main thing is the logic of knowing what you want to do and how you would do it programmatically as sometimes AI won't think about that but even so sometimes it does tell you how to do things correctly - probably trained by a programmer or from programming manuals or something.

[Salem]

https://www.inc.com/georgia-fearn/welcomed-37-data-centers-now-schools-dim-lights-cut-energy-costs/91368206

[Reverend Jim]

So if I read this right, “good fiscal stewardship and good environmental stewardship” applies to all the people who were there before the data centres were built, meaning they should reduce their energy consumption, but not to the data centres owned by the mega-billionaires. In other words, the masses can drop dead from the heat waves that are only going to get worse, largely due to the massive energy footprint of the data centres, while Musk, Bezos et al relax in air-conditioned comfort.

[Biiim]

You should try the UK rates for electricity - I think we are over double your prices here at £0.26p/Kwh translates to about $0.35/Kwh. I think we got about a 10-20% increase here over the last year.

Appears to be a governmental/beareaucratic problem in reality and the data centers get targeted as the electric company can't provide more. That massive increase in Electricity use should of been used to increase the infrastructure to produce more, but apparently they don't sell electricity - the electric grid just sets the price to a fixed cost to pay for the facility so the big increase didn't do anything so they just looked at each other wondering what to do. I bet the data centers get taxed to hell in other ways though.

Basically a governmental socialist type operation trying to work with the private owned capitalist sector again and making the capitalist wrong cause now they have to work - either shut down the data centers or make new facilities. If it was a private company making electricity i bet the company who generated electricity would also be getting bad media and I also bet it would get resolved faster as there would be some key executives on the electric company side very interested in making it happen and profiting on the massive increase in demand - likely fixing the price for the locals and getting data centers to pay more, but since its a social public service where there is no benefactor there really isn't anyone who profits from it and now it's just a problem that civil servants have to solve with little to no impact on their pay so don't mind if it takes 3-7 years and people attack the data centers since there isn't anyone else to attack.

Also i love the Irony of using AI located in one of these data centers to explain why its not his fault.

Gemini:
Your logic is entirely correct for a standard business, but the energy sector breaks these rules because it functions as a regulated monopoly.In a normal market, if your sales jump from 1,000 kWh to 5,000 kWh, you immediately reinvest your profits into bigger factories or batteries to secure that $550 an hour revenue stream.

However, electricity grids operate under entirely different financial and regulatory mechanisms.

Why Utilities Don't Just Build More Infrastructure

They do not make a profit on selling electricity: Power companies pass the wholesale cost of fuel or electricity straight through to consumers without adding a profit markup. Selling 5 times more electricity does not automatically grant them 5 times more profit to spend on cables.

Profits only come from building infrastructure: Regulated utilities make money by charging a guaranteed rate of return (around 9%–10% in the US) on their "Rate Base"—the physical power plants and wires they already own. They must seek government approval before constructing new infrastructure to expand this rate base.

The bureaucracy bottleneck: A power company cannot simply deploy capital to build new substations or transmission lines when a data center arrives. The regional grid operator (like PJM in the US or National Grid in the UK) manages an entry queue for new infrastructure. This regulatory approval process typically takes between 3 to 7 years due to environmental reviews and zoning disputes.

[Reverend Jim]

All of our power in Manitoba is renewable (hydro-electric). Any rate increases have to be approved by the Public Utilities Board. Increases are based on

1. Inflation
2. Projected upgrades/new construction

Our current rate is 9.97 cents per kilowatt hour. We produce considerably more than we use provincially and export the excess to utilities in US. Export price is basically whatever we can get, and that price is calculated based on

1. What it costs the US to produce it locally
2. Demand

Profits from export markets help to keep local rates low.

Electric utilities in the US generally get a bad rap because they deserve it. For example, utilities in California have been caught more than once creating artificial rolling blackouts in order to increase the price. Utilities in Texas have increased prices exponentially in times of high demand. They also refuse to connect to out-of-state grids because that would lessen the chance of a crisis. Texas utilities operate on the credo, "never let a good crisis go to waste". In February 2021 the price per kWh hit $9.

In Manitoba we have two rates, commercial and residential. Ontario varies the residential rate based on demand but Manitoba has not gone that route.

Profits only come from building infrastructure

Wrong. Profits come from charging the maximum price the market will bear. People will pay whatever they have to in order to keep that A/C running when the alternative is dying. It's far cheaper to just jack up prices during a heat wave or cold spell than it is to sink a billion or two into infrastructure.

As for regulatory control, the current administration has been putting the very people who need regulating in charge of doing the regulating. Their idea of regulations is just "Let us do whatever the hell we want."

I can't imagine the chaos that will result when water gets scarce and the billionaires decide they want to own all the water as well as all the money.

For the record, I worked for 29 years in the Manitoba Hydro Control Centre and my brother was a civil engineer on the Public Utilities Board so I know quite a bit from both sides.

[rproffitt]

SDGE of California rates are from a max of $0.69 a kWh down to about $0.26.

Because of this, San Diego has some of the highest percent rooftop solar installs.

For now the electric bill is one of our lowest utility bills. About $20 a month tops but the water bill is about $150 a month.