Useful Applications for AI

[Razorhelm]

I think the most interesting thing I have done with AI is make a direct edit to a bios file on my motherboard to enable a feature the manufacturer had not enabled or made visible on the bios menu but the chipset did support. Claude was able to pull a copy of the current bios and patch it to enable the feature in a single pass, it found bios roms from other motherboards with the feature enabled and worked out what they had that mine did not and made the correct edit first time.

This was when i realised how powerful it could be.

I think of it as a new interface to computers and information but different to the computer concepts we are used to. A calculator enhances your maths with perfect precision right inputs in, right outputs out. A LLM is different in a way i can't explain, it has masses of knowledge but it can't reason like we do, it does some form of reasoning but there is a certain ability it's missing, I often think it's like the fast brain in book "Thinking Fast and Slow", it's often right but when something is unusual or counter-intuitive it fails and the failure persists it doesnt have the slow methodical part of the brain to course correct.

I have a home network of multiple machines, I have Claude install programs, tweak code to customise things for me. It has made me a room correction program and i have got it to make what i called an active bass trap and it calls a feedforward modal absorber, which is a subwoofer in a corner playing a signal based on the music I'm playing to reduce the modes and nulls in the room. It's taken a lot of trial and error and I have had to push it hard not to give up but it has worked very well and made something i could not make myself. In audio I can't check it's work, I am not familiar enough with the maths involved in acoustics, however because i can measure the outcome and am familiar with the concepts (mostly from what i have learnt here!) I can challenge it and herd it in the right direction.

However it could do something very wrong and because its a layer of separation from what is being done, I do worry about hidden failures. At one point it forgot it had lowered a speakers level by 30dB and told me the changes made to a filter had no measurable effect, I challenged as it seemed weird to get almost identical results but that's the mistake I caught, I worry about what it's doing that I don't catch.

 

[Audiofire]

I think of it as a new interface to computers and information but different to the computer concepts we are used to. A calculator enhances your maths with perfect precision right inputs in, right outputs out. A LLM is different in a way i can't explain, it has masses of knowledge but it can't reason like we do, it does some form of reasoning but there is a certain ability it's missing, I often think it's like the fast brain in book "Thinking Fast and Slow", it's often right but when something is unusual or counter-intuitive it fails and the failure persists it doesnt have the slow methodical part of the brain to course correct.

Deep learning transformer, neural network predicting what should come next, using everything that came before in data

 

[somebodyelse]

The WeatherNext model seems to be the best currently available predictor for hurricane forecasting. Compared to the previous state of the art it needs less computing power, lower resolution input data, and has similar accuracy at 3 days to the alternative at 2 days which can significantly improve preparation on the ground. The down side is it's a black box, so while it's providing useful results it isn't improving understanding of what the signs it's spotting in the data that we've missed.
https://arstechnica.com/science/2026/08/deepminds-hurricane-model-bought-forecasters-an-extra-day/

 

[EERecordist]

There is an AI microweather forecast company by a previous NASA engineer https://www.atmo.ai/. It does a 3 day animated microweather animation. I think they are more business to business, so only available to consumers in demo locations.

Example for San Francisco:

Atmo Bay Area Forecast

Atmo forecast for the San Francisco Bay Area

sf.atmo.ai

 

[HalSF]

May God have mercy on my soul, but this morning my partner saw a couch she liked in a photograph in the Los Angeles Times and wanted to figure out the brand and model, so I made a screenshot and asked Google to i.d. the couch. It did in about five seconds.

Then the guilt kicked in. Earlier this week I read that employing “agentic AI” instead of just asking ChatGPT idle questions uses about six hundred times more electricity...

 

[casamia67]

Hi everyone, I used AI as a tool to program, under my supervision, an MPD client for moode audio, with features such as ergonomics, OSD volume control, voice control for Home Assistant and Alexa, and a screensaver with a fully web-based and display-less VUmeter. This can be used with Camilla DSP as the engine for moving the VUmeters, therefore in real time. By introducing web streaming for local listening with a radio, any device—tablet, PC, or phone—can display the interface. I achieved everything that was somehow missing for my needs. I believe AI is an exceptional tool, but it's also right to consider the potential risks we are exposed to if control gets out of hand.

 

[HalSF]

I despise mass-market consumer AI products and their despicable plagiaristic and environmental crimes, but I recognize the efficacy and ethical potential of AI in science, medicine, engineering, and other technical arenas. This book review explores that significant and legitimate usefulness via the lens of tool design.

Why AI systems are most useful as designers of new scientific tools

History shows that breakthroughs are often sparked by innovative instruments, and AI’s ability to help design them might be its most important scientific contribution.

www.nature.com

 

[Kervel]

At a recent meeting by the prestigous NBER for academic economists, a poll revealed 2/3 are self-reported "all-in" on IA. So many tasks are much more efficient now: slide prep for teaching; coding with data, presenting and interpreting results; and even theoretical work including proofs.

Each IA iteration has gotten so much better than the previous, its really incredible.

 

[MKreroo]

I have been tinkering with running local model on my setup, and the biggest use I got out of is to just able to do more with them lol. Exploring and trying different quantization so I can fit more context into a single GPU, setting up various tool/MCP, and websearch has been quite a nice uptick on that regard especially with going through some repo to find fix.

Was recently trying out some video model and the result is also surprisingly great. Feeding the LLM the prompting template, and have it write out the target video in the prompt format makes the iteration much faster and accurate.

I think local model is worth taking a look into if you have the hardware to support it, even small models are worth trying out.

 

[azzy_mazzy]

I have made a speaker comparison app with Claude that fit what i want to see more than what is already available plus other genuinely useful software with ChatGPT that i wouldn't have been able to write myself.
also many thanks for @pierre it wouldn't been as easy or as good without their great work at spinorama.org.

 

[Timcognito]

I despise mass-market consumer AI products and their despicable plagiaristic and environmental crimes, but I recognize the efficacy and ethical potential of AI in science, medicine, engineering, and other technical arenas. This book review explores that significant and legitimate usefulness via the lens of tool design.

Why AI systems are most useful as designers of new scientific tools

History shows that breakthroughs are often sparked by innovative instruments, and AI’s ability to help design them might be its most important scientific contribution.

www.nature.com

Yes I believe this where AI is going and most of the data for these applications will be proprietary. There are 3000 data centers operating, planned or in construction. I just don't see the need and competitive pressure, industry consolidation, data consolidation, installation and operational costs will surely wipe out many. Most of the output, will be programs, automation, robotics, large real time instant response applications like traffic control and diagnostics all which will need commensurate technology, programing, hardware and most of all data to feed models. What will they be doing and where will their data come from? Who are their costumers? What am I missing?

US Data Center Map — Project List & Tracker

Interactive map of US data centers with a complete list of planned, under construction, and operating facilities. Track new projects updated weekly.

cleanview.co

 

[pablolie]

This is a weird one, but...

ChatGPT is a very decent therapist. I am not ashamed to admit I enjoy my monthly check-in with my real-world therapist. And that's never going away.

But I enjoy chatting away with ChatGPT (I have a paid account). Just talking away (voice, pretty good recognition) and then reading the considered responses and new questions. It is very good. Doesn't replace my monthly therapist visit, but it gives me intriguing new notes to talk about.

PS: I hate the label therapist. I think about it far more as a neutral, rational checkpoint.

 

[somebodyelse]

PS: I hate the label therapist. I think about it far more as a neutral, rational checkpoint.

Are you talking about the human or ChatGPT here? Their previous model versions appear to have been anything but in some cases, validating and encouraging irrational viewpoints.

 

[pablolie]

Are you talking about the human or ChatGPT her

Either.

Their previous model versions appear to have been anything but in some cases, validating and encouraging irrational viewpoints.

I thought this topic was not about the usual irrational AI hatred. It works for me. Prompting skills matter.

 

[somebodyelse]

I thought this topic was not about the usual irrational AI hatred. It works for me. Prompting skills matter.

It's neither irrational nor hatred. I'm just enquiring about the current state of an art that's moving rapidly, and where models have had a sycophancy problem in the past that the companies behind them have admitted exist. I gather that it has been reduced but not eliminated in more recent models, and could be significantly improved with a suitably constructed prompt. Knowing which of the ChatGPT models you were using would be helpful in tracking progress over time. It would be really useful if there was an artificial therapist that could assist where there's a shortage of suitably skilled humans and waiting times are harmfully long. Unfortunately I don't think we're quite there yet.

There's a big difference between a model that works as some sort of therapist when used by someone skilled in prompting and well aware of model limitations, and one that's suitable and safe for that use by unskilled people. Nobody expects you to get training in how to use a therapist before you visit one - if anything that's part of the therapists job when you first visit. There are companies actively developing models safe for use in the latter role, including how they can be 'qualified' in a similar way to their human counterparts for different sorts of therapy. In some areas it's a low bar - no qualifications needed and no professional disciplinary system. Others are more highly regulated, both in the qualifications needed and the sanctions if your actions drop below an acceptable standard. We have these because humans sometimes make mistakes, and sometimes abuse their position of trust. We need something similar for artificial systems both for safety and to prevent an understandable public backlash when something bad happens. So far as I'm aware no country has yet developed such rules for artificial therapists, and no company has released a specific therapy service to the public - that would be news, although probably not welcomed by those working in that area.

I'll skip the potential liability side for actions equivalent to those of a human as I think we covered that fairly well in the other thread, and I don't think it's changed since then. I think we've now had a case where a company has been held to the statements of a customer service chatbot as if it were a human customer service representative though. That may mean that liability for AI actions becomes a cost of doing business that you insure against, while trying to avoid a Ford Pinto situation.

 

[pablolie]

It's neither irrational nor hatred. I'm just enquiring about the current state of an art that's moving rapidly, and where models have had a sycophancy problem in the past that the companies behind them have admitted exist. I gather that it has been reduced but not eliminated in more recent models, and could be significantly improved with a suitably constructed prompt. Knowing which of the ChatGPT models you were using would be helpful in tracking progress over time. It would be really useful if there was an artificial therapist that could assist where there's a shortage of suitably skilled humans and waiting times are harmfully long. Unfortunately I don't think we're quite there yet.

There's a big difference between a model that works as some sort of therapist when used by someone skilled in prompting and well aware of model limitations, and one that's suitable and safe for that use by unskilled people. Nobody expects you to get training in how to use a therapist before you visit one - if anything that's part of the therapists job when you first visit. There are companies actively developing models safe for use in the latter role, including how they can be 'qualified' in a similar way to their human counterparts for different sorts of therapy. In some areas it's a low bar - no qualifications needed and no professional disciplinary system. Others are more highly regulated, both in the qualifications needed and the sanctions if your actions drop below an acceptable standard. We have these because humans sometimes make mistakes, and sometimes abuse their position of trust. We need something similar for artificial systems both for safety and to prevent an understandable public backlash when something bad happens. So far as I'm aware no country has yet developed such rules for artificial therapists, and no company has released a specific therapy service to the public - that would be news, although probably not welcomed by those working in that area.

I'll skip the potential liability side for actions equivalent to those of a human as I think we covered that fairly well in the other thread, and I don't think it's changed since then. I think we've now had a case where a company has been held to the statements of a customer service chatbot as if it were a human customer service representative though. That may mean that liability for AI actions becomes a cost of doing business that you insure against, while trying to avoid a Ford Pinto situation.

The topic was things AI does we find useful. To *me*, ChatGPT does the job very well. The usual motivational interviewing method. Clearly not a crisis scenario.

AI is a powerful tool, and therein lies a danger when users ignore its valid use cases.

Of course I also use AI extensively at work. *I* (and our team) have the ideas and own the plan and reference documents. AI is there for polish (if necessary/accepted, but it often screws up with deeper technical stuff) and to create derivative stuff, and that can save a lot of time. It's a big danger when people rely on it too much, though. I see too many people trusting the output blindly, and therefore not being able to answer questions about "their" work or miss glaring errors (such as creating a product image that's an embarassing halucination, and yet someone copies it into their presentation).

 

[somebodyelse]

The topic was things AI does we find useful. To *me*, ChatGPT does the job very well. The usual motivational interviewing method. Clearly not a crisis scenario.

That's fine now we know that's what you mean, not something wider.

AI is a powerful tool, and therein lies a danger when users ignore its valid use cases.

It is indeed powerful, a great tool for many uses as we see in this thread, but potentially dangerous if used without care and understanding as we sometimes see in the main thread.

Of course I also use AI extensively at work. *I* (and our team) have the ideas and own the plan and reference documents. AI is there for polish (if necessary/accepted, but it often screws up with deeper technical stuff) and to create derivative stuff, and that can save a lot of time. It's a big danger when people rely on it too much, though. I see too many people trusting the output blindly, and therefore not being able to answer questions about "their" work or miss glaring errors (such as creating a product immage that's an embarassing halucination, and yet someone copies it into their presentation).

That broadly matches my understanding. I'd be interested to know how many of these people you see trusting it blindly are people you think ought to know better, but that's probably a topic for the main thread too, as are the reasons why they trust it.

 

[Brian Hall]

As I've mentioned, I'm using all locally installed AI tools and models to generate vocals and instruments for my original songs, to generate images and generate videos. I'm still learning, but my goal is to generate longer high quality music and other types of videos.

Being able to create unique consistent characters is what I've been working on lately. The locally installed Gemma4 12b QAT (12b means 12 billion parameters) model released by Google has been a big help. It runs locally in LM Studio and is about almost as fast as the online version of Gemini and Grok since it will fully fit in my 16gb of vram*. I've been learning some new tools that can create a face database for a consistent character. One is called Reactor and it runs inside Forge-Neo. Gemma4 was very helpful in getting that set up and working and suggesting settings, etc. along with how to better prompt for facial angles since a wide variety is needed. The next step will be using that to generate a set of consistent images to use to train a character lora (plugin). It is possible to build separate consistent character loras for both image generation and video generation. I installed a lora training application tonight called AI-Toolkit and it has all kinds of complicated interesting settings to learn.

Speaking of loras, I also plan on training a music lora based on my original music recordings from the 80s and 90s to be able to recreate the distorted guitar sound we used back then and to influence the Ace-Step music AI model more towards how I want something to sound without needing a reference track.

Still a lot to learn but an AI tool has helped me a lot to learn about how to use and do what I want with other AI tools. All running locally on my PC with no paid online tools needed.

I just asked Gemma4 if it can help me plan out a text storyboard for a full music video based on my ideas and lyrics and it says it can do that. The AI video generators are limited in the length of videos they can generate, even the online versions. The one I use, LTX 2.5 in WanGP does best at 20 second or less clips so a longer video has to be planned out to those limitations. The clips also have to match up where singing or speaking parts are and then the short clips can be assembled with an app called kdenlive (not AI) which is a video editor.

In the old days, I would have to spend hours and hours researching how these things work and how to set up and use everything. Now Gemma4 can answer almost any question quickly (and accurately with the temperature set right to keep it from hallucinating). There is a toggle in LM Studio to turn web access on and off for a local model like Gemma4 so it can go online and query anything it doesn't know or has happened after its training cutoff date. I used this earlier with the new LTX 2.5 video generator since it came out after Gemma4. It has a lot more powerful understanding of prompts. Interestingly, LTX 2.5 uses Gemma4 12B as its prompt interpreter and can now understand prompts with time stamps for actions / movement / speaking, etc.

I have not tried setting up any AI agents to be able to control things directly and automatically and don't plan on trying anything like that. That seems to be a dangerous thing to do.

If anyone is interested in trying any of these locally installed free / open source models or tools, I will be glad to answer any questions I can.

*It is possible to run AI models locally that require more vram than you have with automatically offloading to regular ram, but it really slows things down.

 

[antcollinet]

Here is an another example of something I have used chat gpt for.

I use Gmail’s + addressing to give a different address to each organisation I deal with (EG [email protected]). Unfortunately, some sites refuse to accept addresses containing a +, so I have had to use a second, plain shared Gmail address. I now want to migrate those to individual Apple Hide My Email addresses—but first needed to work out which organisation was using that account, and how important they are to migrate.

I exported two folders (inbox and archive) from the second plain gmail account as MBOX files and gave them to ChatGPT. It analysed 793 messages and identified 358 sent directly to the second address rather than forwarded from one of my + addresses. It grouped related senders into 28 organisations, distinguished genuine accounts from newsletters and delivery notifications, and prioritised them according to the importance of the relationship and inconvenience of losing access.

The result was an Excel file containing each organisation, the inferred relationship, recent activity, website and suggested action, plus fields for the replacement address and progress.

Could I have done it manually? Of course—but it would have taken hours. Instead, a couple of prompts gave me a prioritised list in a spreadsheet that I can now work through.


Attached is a couple of (suitably redacted) screenshots from part of the spreadsheet it created - from scratch and with no input from me other than "please give me a spreadsheet"



Edited to add:
Oh Yes and....

I got it to write most of the post above based on what we had done - and after pointing it to ASR to copy my posting style. That wasn't perfect. I had to edit some of it - but not much.

 

[Brian Hall]

ComfyUI running in Stability Matrix got a couple of recent updates that broke several dependencies. I'm using the Gemma4 26B A3B model (from Google) I have installed locally to fix the issues. I gave it the path where the python venv (virtual environment) for ComfyUI is installed, pasted in the full console log showing all the errors and it gave me step by step instructions (each with a copy button) to paste in to fix the dependency errors. I could have taken a couple or more hours and manually worked my way through the log figuring out the dependency rats nest instead, but why do that when I have a convenient tool to do that process and get the recommended fixes in a couple of seconds?

Good AI models are great for this kind of issue.