Are you using AI?

Nice feedback shabaz I too mainly use Gemini for coding. It's what I've become used to, so I stick with it. However, I use the Gemini coding agent directly in VSCode. I prefer this to using the Gemini CLI because of the way Gemini presents the proposed changes inside the IDE. It is like a GitHub pull request. You can review each change section by section and accept or reject. You can even modify the proposed changes yourself, if you like.



Then on the left hand panel (not shown in the pic) is the explanation and rationale.

The only drawback I find with this approach is that the Gemini AI predictive code suggestions within the VSCode IDE is absolutely dreadful. Most of the time it hasn't a clue. So you have to reject a fair bit. But otherwise I find this a sound approach as you are the code reviewer. If it works for GitHub it works here too.

I agree with Shabaz in his closing remark. We'll soon be priced out. He explained it better than I did.

I do use AI for coding assistance, and have had to adjust the way I work slightly (but for the better I believe) to improve the results. For very tedious but obvious things, I'll take the easy way out and just request ChatGPT to do it all. That works for tools/utilities and for formatting data, creating arrays and so on, and does save a ton of time. However, sometimes I won't entirely be able to follow the resulting code, since I was not involved in writing it at all, I merely acted as tester and raised bugs/issues back to ChatGPT.

For more involved coding, where I do want to provide input and I want to be able to more closely follow what's going on, my habit is to only provide ChatGPT with the specific source and header file, i.e. I make sure to split the project as much as makes sense, and all ChatGPT is made aware of, is the specific source/header and perhaps a general outline of what the purpose is. I find ChatGPT works nicely when scope is limited in this way. Also, very descriptive function names helps, it then copies your style for any new functions it creates.

Recently I found a magazine called Code, and got a subscription. It's not bad. The latest issue which I received a couple of days ago, has six main articles, of which three are AI-related. I tried one of the articles, which teaches how to use Gemini. I decided to try it for a MSPM0L1105 microcontroller project, which is to create an I2C-to-RS485 converter. Gemini is being used in a mode where it takes almost full control. It has access to a folder you choose, and it takes its instructions from a text prompt. Here are some screenshots.

When started up, this is the text prompt:

According to the magazine article, it's suggested to create a GEMINI.md file containing all your preferences/desires/coding style etc. I decided to type in the text prompt for gemini to create that GEMINI.md file itself, to contain the detail regarding the project I wanted. This is what I typed:

It went and did it, and then prompted me to accept that change. 

Now I figured it would be good to get to a baseline, where I could compile at least something, before adding in the new functionality:

Gemini went off and downloaded the appropriate code and made any necessary changes to the makefile, and it couldn't find my make.exe, so I then told it where that was in the text prompt, and then it built the code. Now the previously empty folder in which I had run Gemini originally, was full of this content:

Then I told it to make a change to the makefile to fix a 'make clean' issue (I pasted in the error I was seeing) and it fixed it, but it was quite trigger-happy, it went ahead and made the code changes to implement the new desired functionality.

This screenshot shows some of the things it did by itself:

It ended like this:

I've no idea what bugs are in the code, but the files content looks plausible. I've confirmed I can build the code using make myself, and it does build. I'm sure I will need to troubleshoot and so on, but it's a start, and the volume of code isn't so much that I will get lost. At the end of all this, I've used 39% of my allowance quota (it's visible at the bottom-right of the screenshots), I'm using free Gemini for this test.

Finally, to try to understand things, and for a couple more percent of quota:

After it thought for a bit:

The resulting documentation looks quite good, this is a snippet:

I then asked Gemini to rename the project to i2c_uart_expander, and to create a README and upload it all to GitHub. All the documentation and housekeeping stuff took about 14% more quota, so I'm currently at 55%, I think this will reset to zero tomorrow (not sure). Anyway, if anyone is interested in browsing it, the project is here: shabaz123/i2c_uart_expander

I'll try to run it at some point, I'm sure it will need some troubleshooting as mentioned before, but I think it's a good start, allowing me to concentrate on the more interesting stuff, rather than spending time with Makefiles, datasheets, documentation, etc.

All the above took a couple of hours (I was the bottleneck; I was going slowly, and doing other things in the background).

I found Gemini a little scary in that it has a lot of control and might be trigger-happy, but it's concentrated into the folder/sub-folders from which it is run from, so it hopefully can't do a lot of damage. Plus, with the repo, at least now I have the backup if it later messes up things too much. I think I still prefer the slower per-file method I use with ChatGPT, but I can see there's potentially a huge speed advantage with the method described. 

I agree with BigG that there is a big risk that it's going to get too expensive; what if one day, it costs say a sizeable proportion of a typical salary to use it : ( I don't think the issue is that it makes people become lazy, I think the issue is, it moves efficiency to a level unattainable without AI; and that will likely have a massive value, so AI firms will feel justified in pricing accordingly. 

Indeed, it saves my time. But someone needs to be good at it. Otherwise, they would skip important details of the code and technical side of that. 

It indeed saves a lot of time. 

For example, I have posted a project for spring clean where I need a heatmap display with LVGL with MicroPython or C/C++ and the quick prompt to create such code helped me a lot. There were some errors but then I found the solution soon. That makes it work. 

The code is not too generic always but for sure, there is some mix up between the code and platform the AI engine is using or it is trained on. For specific platform to improve the quality, they need a large code base for that particular platform and then train the model for that. Then it would be more accurate. 

That's the best part. When you're developing something new for work no one quite knows what the right answer is.  People only know about the "surely this SHOULD work".

AI is merely an input into the "how about you try this" or you ask AI "would this work" and you get an answer. The beauty about AI is that you don't have to feel embarrassed asking those stupid questions and you can blast it full of questions (until your tokens run out). You'll never get the silent treatment. 

Basically, it's a sounding board, who's trigger happy to provide a positive answer. Of course you have to take it with a pinch of salt those bullish statements of "You are absolutely ready to test. Flash this to the board and you'll have it working...". So the trust is that it's also trying to converge at a solution. It just won't get there in one go.

I know some people often say to me "I just want you to listen. You don't have to always offer an opinion". Unfortunately AI doesn't do that either.

“Faster garbage is still garbage” sums up the risk pretty well. It sounds like AI would need to prove that it can be checked, traced and trusted before it becomes useful in a work context - interesting, thank you!

Thanks, this is a really useful perspective, especially from someone using AI heavily across both work and personal tasks.

Your point about becoming too dependent on it is interesting. It sounds like AI can speed things up massively, but there’s also a risk that people lose the habit of doing the deeper work themselves.

When you’re using AI for work, what makes you decide whether to trust the output enough to act on it?

It sounds like AI is most useful when it helps make sense of a lot of information or gives people a starting point rather than a finished answer...interesting!

Thanks!

When you use ChatGPT for learning Fusion, do you usually get enough detail to complete the task straight away, or do you still need to cross check with videos/docs?

Thanks. There’s a lot of hype around AI being able to design circuits or write code, but your example shows the gap between something that looks convincing and something that actually works.

It’s interesting that it was more useful for explaining commands than generating working code. That distinction feels important, AI might be helpful as a support or help tool, but not something you can rely on to produce usable output on its own.

It sounds like AI is becoming hard to avoid in search now, even for people who aren’t deliberately choosing to use it.

The Python example is interesting too, AI summaries seem useful when you’re trying to remember a method, library or concept, but not as something you’d trust blindly. That “useful for direction, but needs checking” theme is coming through a lot.

When you do use AI summaries in search, what makes you decide whether the answer is good enough to follow up on?