Discussion of AI tools for development

[hjtrbo]

Damn this AI shit is the future.

After Claude made me a 5" esp32 dash mounted can gauge I got it onto this bad boy. Full blown GM data logger inspired by the HP Tuners logger. Claude can't claim all the credit - I already had the transport / protocol layer done and bench tested and most of the ui done, but I got it to refactor the whole code base up to higher standard and it fully implemented the plot library and interfacing with the other display modules on it's own. I also threw it GMW3110 and had it verify my handling of the can frames etc was correct. Using Avalon Dock, Material Design and OxyPlot Skia wpf libraries. Supports reviewing files in csv format from pcmtec and hpt. Implements data logging global A via J2534 pass thru. Supports mode $22 (polling) and mode $AA (rapid packet) via mode $2C / $2D (read by address). Logs 50 channels at 50Hz very well and garbage collection seems non-intrusive in the simulated environment, as we all know, real life is not simulated.

Perfect for the ghidra users wanting to know what ram addresses of interest are reading. Star Trek is going to be real.

[8SecSleeper]

Damn this AI shit is the future.

After Claude made me a 5" esp32 dash mounted can gauge I got it onto this bad boy. Full blown GM data logger inspired by the HP Tuners logger. Claude can't claim all the credit - I already had the transport / protocol layer done and bench tested and most of the ui done, but I got it to refactor the whole code base up to higher standard and it fully implemented the plot library and interfacing with the other display modules on it's own. I also threw it GMW3110 and had it verify my handling of the can frames etc was correct. Using Avalon Dock, Material Design and OxyPlot Skia wpf libraries. Supports reviewing files in csv format from pcmtec and hpt. Implements data logging global A via J2534 pass thru. Supports mode $22 (polling) and mode $AA (rapid packet) via mode $2C / $2D (read by address). Logs 50 channels at 50Hz very well and garbage collection seems non-intrusive in the simulated environment, as we all know, real life is not simulated.

Perfect for the ghidra users wanting to know what ram addresses of interest are reading. Star Trek is going to be real.

You can take it a bit farther even. I have made my own ai harness that does a pretty good job reverse engineering ecu firmware, My harness grabs the top 5 functions that have the most variables learned, but function is still generic. It goes through and looks where each variable is named and it will rename functions, summarise them, after the 5 are processed, file is updated, git commited, it then wipes the ai memory, feeds it 5 again, repeat 24/7 on it's own. Starting with just giving it the dtc table it's made major progress, It auto identifies tables now, using the lookup functions and how they work.

As for the csv parser, take it a step farther than that. I gave mine a exported csv log, then the raw binary, 57 mins later it had a working csv parser, saves you a step loading a software that's 20 years old and is extremely slow to open and export a log.

I've spent 10-12 hours a day dealing with ai for the last month. Learned a lot, it's a pretty good bs artist at times, you gotta be careful and double check everything. But it's a tool and has a purpose. You need to do most of the ground work with normal scripts to reduce how much thinking it must do, keep the context window under control.

I may or may not have had it crawl a site or two for 100 pages in a single thread and create full summaries and a large instruction guide with quiz's after every section to make it easier to consume data, vs just reading a thread directly.

[antus]

That is a cool process!

Yep, its those site crawlers that have been a major PITA for me maintaining this site. It is expensive for an individual myself to support the kind of bandwidth needed when many people do that, and people are doing it. It's the new AI world we live in. I am not going to spend too long complaining about it though. AI is here, for better or worse, and complaining isn't going to change anything. Might as well just take the wins from it and have to spend the time to keep the information up and the community alive. I can only hope that those who scrape the site feed at least some of their learning back to other members to keep it fair.

My 2c for the discussion, I noticed Claude got really dumb about a week ago for me. Searching it lots of people have been talking about it since February. It lost the ability to think properly and wasted 2 days before I was absolutely sure that it was nerfed, not just problems it couldnt comprehend for some meaning. So I had to switch up to the top models and then get it revisit the last few commits and undo/fix some of the work the nerfed 4.6 version had done.

I'm not sure if they're nerfing it region by region, or maybe they do it about a week after your second billing cycle. So do watch out for that. The quota seems to be OK on 4.8 anyway. My fear was it'd drink through it in 2 seconds and become unusable at the base paid tier, but that has not been the case.

[8SecSleeper]

That is a cool process!

Yep, its those site crawlers that have been a major PITA for me maintaining this site. It is expensive for an individual myself to support the kind of bandwidth needed when many people do that, and people are doing it. It's the new AI world we live in. I am not going to spend too long complaining about it though. AI is here, for better or worse, and complaining isn't going to change anything. Might as well just take the wins from it and have to spend the time to keep the information up and the community alive. I can only hope that those who scrape the site feed at least some of their learning back to other members to keep it fair.

My 2c for the discussion, I noticed Claude got really dumb about a week ago for me. Searching it lots of people have been talking about it since February. It lost the ability to think properly and wasted 2 days before I was absolutely sure that it was nerfed, not just problems it couldnt comprehend for some meaning. So I had to switch up to the top models and then get it revisit the last few commits and undo/fix some of the work the nerfed 4.6 version had done.

I'm not sure if they're nerfing it region by region, or maybe they do it about a week after your second billing cycle. So do watch out for that. The quota seems to be OK on 4.8 anyway. My fear was it'd drink through it in 2 seconds and become unusable at the base paid tier, but that has not been the case.

I think if you look into what meta is using, it will solve the crawling problem. I'm not sure on the costs, but they use something good. Cloudflair is very easy to bypass, it used to be a great solution, not so much anymore.

I've noticed when using provider that offers multiple models, that when you go straight to the source and use the raw api and skip a provider. That the model is so much smarter. I'm thinking the middle man providers that offer you to use multi models are somehow nerfing to reduce their cost.

[8SecSleeper]

Attached is a guide I had it make, (edit this is after I told it to include details to get security access, flash it. I think it made up some stuff. I need to have a new one generated from the crawled data), note I didn't flood request, made it wait several min's between request, vs being like my site is crawled by bots in an instant.

Also attached, python script that you can use with a exported ghidra file, it scores based on known variables, returns the top 5 functions that have yet to be named, Makes it easy for AI to consume. Pair that with a good prompt and you can get alot done, like telling it to provide a summary and stop after 5 and request permission to continue. To keep the ai's job easy. I make it produce a summary this run a rename script like attached.

[antus]

The problem in preventing or limiting scrapeing is the time cost. Its not free to configure servers, DNS, anti-scrapeing tech and so on. And even when the software is free, the time cost can still be significant. Thats why there is few places like this left. Most he small players who are not driven by money are being pushed out. Anyone can start yet another facebook group but its not the same. The number of original web sites and forums must be well on the way down now.

Personally I put more time in to this one which is why it is still up but there is another enthusist one that has had less of my time and its off-line right now and will take more to get it back up as its older to begin with and needs more work. If I cant or dont find the time for that, a lot of cool content from the last 15 years for that particular topic will be gone for everyone.

[8SecSleeper]

I understand all of that and I agree about the Facebook deal. I've been around on the Internet since the 90s. And this is one of the last forums online. Aside from ones that have heavy sponsors. Facebook has zero flow, they could easily fix it.

You could make it so attachments can't be downloaded unless you have 25 posts. You could use some basic logic to disable serving the site based on crawl rate. I haven't setup cloudflare but would assume they have something like that. I can try setting it up for my website today. I only avg about 30 bots a day. And I have it setup to identify bots and parse out legit looking sessions.

I assume your getting thousands of guests a week though. Ai is only just getting started now too . These open source models are great but are too expensive to run. That won't be the case in a few years.

[antus]

There is cloudflare, time spent and it is enough, plus moving the DB to my own server away from high cost. The other site, needs time.

[hjtrbo]

I'm growing very tired of the proliferation of ai generated images, sentences and text style.

em dashes.

This isn't merely an oversight, nor is it simply an error of any kind—this is, quite frankly, a precisely calculated, surgically executed knockout blow of absolutely devastating proportions. A pinpoint strike of pure, unadulterated accuracy that leaves no room for doubt, ambiguity, or further deliberation. This is a masterclass in tactical excellence, a flawless demonstration of methodical precision that resonates with undeniable, irrefutable finality.

[antus]

Yeah, the world and future generations are worse off with AI, that is for sure. But its here, so its not like we have a choice.

You can tell it to never use em dashes, or emojis and remember that, though. But the internet has really become a trash heap in the last year and especially the socials now AI content is a thing. And watching all these masssive AI data centres be built, I know its because those artificial brains will be doing most the work in the next few years. I reckon the kids today are likely to see the shit hit the popper fan and possibly the collapse of old school society. It'll probably be during their lifetime where there are so few jobs left we need to have proper universal basic income with only 5% of the people doing the actual work, but otherwise running the bots, both logical and physical. Can we do that transition cleanly? not so sure.

As a side note I think we're making the same mistake naming things now as they did at the dawn on radio. HF - High frequency transmissions, are really damn low on the scale of frequencies we run now. So too I believe we'll laugh at the definition of LARGE language models in another 50 years, as they'll be toys compared to whatever future hardware can do. Assuming we don't have something like the "dark dark days of the data krash" as the cyberpunk 2077 game called it between now and then (and also the RPG, before the game.. https://cyberpunk.fandom.com/wiki/DataKrash )