Useful Applications for AI

[mpw]

While still working in software R&D I needed a simple webserver returning a JSON file to test my C++ client code. I asked the very first free Chat GPT to write sich a server using Python. The 1st implementation was wrong, the 2nd almost good, the 3rd worked flawlessly. It took me less than half an hour. Although I knew Python I had no knowledge about the libraries/packages I would need for this job, so it saved me a lot of time searching and trying.

AI is also quite good in translating text and searching technical information. It stumbles in fields where the internet is full of wrong data, like audio. Therefore one needs to have at least a basic knowledge in those fields to be able to check the validity of its answers.

I used Google Gemini to write a simple python program to run local LLM's that I modified to my liking. If one is a programmer then a basic template of an application is good then we validate it if makes sense. If one has no programming skills how would a person know if code produced by AI was any good ?

 

[mglobe]

Great ! Have you given your Claude a name yet ?

Sorry for the slight diversion, but for whatever reason, this comment made me think of this:

 

[kemmler3D]

what kind of home automation are you building? i struggle to understand what kinds of things people are actually using ai for. to me it is just a technology to get a picture of jim varney and charlie brown in a canoe-- just silly stuff that has no real value.

If you have any HA setup where custom code might be useful, you can probably get pretty far asking the AI to code up the solution.

For example if you have an ESP32 and you want your garage door to open when you get close enough to the house, you'd explain that to the LLM and more than likely get a working solution within an hour or two.

 

[GS430man]

I find that LLM AI is very good at regurgitating medical, scientific and engineering results where there are good internet-published documents.

It's not bad at helping with product functionality questions when the manuals and faqs are online.

It's great at helping with complex math operations.

Most of the remainder of cases I need to know enough BEFORE I use AI to tell if it is giving me good data, nonsense or some mixture.

 

[Power Pop 23]

A.I. is one factor credited with fueling an increase in the number of start-up businesses in the USA since the pandemic. (gift link)

America’s Enterprising Spirit Is Booming After Decades-Long Slump

Hundreds of thousands of people are creating companies each year, ending a decades-long drought.

www.nytimes.com

 

[fivepast8]

Here is another useful application. I am reviewing a quality agreement with a Chinese multinational that has too many unfavorable clauses in it but we have next to no leverage in negotiating changes. Thus, I wanted to make only one change and include an aggregated liability cap. I asked ChatGPT to write me said clause. Then fed it to Co-Pilot for review. This went back and forth twice until both models were happy. Now let's see if my customer accepts it...

We routinely upload entire agreements and ask it to point out all one-sided or dangerous clauses.

 

[ephemere]

Today I asked ChatGPT to review the last 3 months of my gmail, find all messages from recruiters head-hunting for robotics companies, and forward each message to my son. Before AI this would have been effectively impossible — way too much email to wade through manually, and even writing a script wouldn't really have worked. But using agentic ChatGPT Work, it took all of 60 seconds to write the prompt.

For fun, I asked ChatGPT to read this thread and report: "How many of the posts are actually responding to the OP's prompt? Among all the others, categorize them into buckets by type of response."

I audited all 106 posts currently in the thread: one OP and 105 replies.

Using a strict standard—“describes a concrete positive AI use the poster used or observed, with enough detail to show its usefulness”—44 of the 105 replies respond to the prompt: 42%.

Including the OP’s own example, 45 of 106 total posts contain a prompt-compliant use case.

Other response type	Count	Post numbers
Near-misses: vague, hypothetical, mixed positive/negative, or general impact rather than a demonstrated use	9	30, 32, 44, 55, 63, 88, 90, 103, 105
Criticism and broader debate: accuracy, failures, risks, jobs, education, privacy, copyright, and negative experiences	22	21, 27, 28, 36, 46, 56, 64–66, 70, 72–76, 80–82, 89, 92, 93, 95
Follow-ups and tangents: questions, corrections, technical side discussions, or comments on another person’s example without offering a new use	18	14, 34, 35, 47, 50, 52, 53, 58–61, 78, 83, 84, 86, 87, 91, 98
Low-substance/meta: thanks, moderation, jokes, sarcasm, link-only replies, or pasted AI output	12	10, 17, 29, 38, 39, 41–43, 68, 79, 100, 102
Total not directly responsive	61

If the nine near-misses are judged generously, the responsive count rises to 53 of 105 replies, or 50%. So the defensible conclusion is: 42% clearly answer the OP; at most about half do under a looser interpretation.

 

[ephemere]

Another example. When buying a house, under time pressure of an offer deadline, I dumped many years of permit docs — dozens of disorganized files — into ChatGPT and asked it to summarize the history of the house. Among other things, I learned all about a major mold issue (not disclosed by the seller) and what was done to mitigate it.

 

[Jeromeof]

So some of you might be horrified or intrigued by this but the speed that we can generate software with AI is always impressing me (and I have been using / programming against AI tools since the first llm / API that OpenAI released).

So for fun I thought I would see how long it might take to create a new viable audio website and fully publish it with 'useful' content. So here is a little timeline.

Note: I have my Pragmatic Audio blog already created and I knew what worked well with that so some of what I will describe below will seem is sort of cheating as I made upfront decisions to most generate what I now works for me.

But I will more highlight the AI tools used and mostly how long I spent with them

1. Domain Names ( 20 minutes - mostly thinking more deeply about what I wanted to build )
Last night, I thought it might be nice to have a second website - less about 'Pragmatic' stuff - but more about well engineered audio products at any price. So, using claude code I more / less asked for a list of possible websites that I could registered with ideally one .com but I did briefly ask for .audio domains. Given this is a bit of fun I didn't want to pay too much for the domain so as well as asking for good relevant domain names I wanted the annual fee to be cheap. I got probably 20 semi usual domain names and in the end I settled for https://www.informedlistening.com as I felt it had the right concept for what I wanted to publish on the website and $15 a year seems reasonable. Using AWS route 53 !!

2. Architecture ( 10 minutes - I already had an idea of how and what I wanted to use to put everything together )
So, I originally was planning on looking at it over the weekend but when I got up this morning I thought why not use Codex to create a skeleton architectural plan for what to build - this involved me create a detailed 'prompt' describing mostly how I had build PragmaticAudio.com originally as a Hugo 'static CMS' Website hosted on gitlab pages. The advantage of this is you mostly get free worldwide hosting and gitlab pages when setup correctly will manage certiifcates and other 'difficult' aspects of website hosting.
3. Infrastructure setup ( 10 minutes )
This was basically giving codex readonly access to by private gitlab - so it could give instructions both for Gitlab but also AWS route 53 instructions to get the DNS and certificates all setup correctly - this took me most of a day to get fully right when I did it by hand for PragmaticAudio.com.

4. Build And Preview (30 minutes - mostly the back and forth refining what was generated )
So this involved reviewing the high level plan from 2 above, describing the main attributes I wanted this new website to have. Which was 3 things - Reviews ( just come cherry-picked ones I had already written copied over), 'Informed Articles' and News stories. These 2 extra features were subtasks of the overall plan which I will get too below.

So, which this expanded prompt I got codex to generate a complete Hugo website structure. There are lots of Hugo websites already on GitHub so it can easily 'Vibe code' a suitable structure with sample content. After a bit of reviewing ( basic back and forth with codex) I got mostly something locally that looked good and I briefly published it to make sure the gitlab setup from 3. above was all correct.
Note: I probably should have spent more time reviewing as tweaking this with 'real' content later too most of my time.

5. Subtask 1 - News and Editorial 'tool' ( 2 hour )
Since I wanted the 'Informed' name as part of the website - I feel that having up to date audio news might be the most useful aspect but while I don't want to have 'AI' scrap and generate crap new stories, I also am mostly busy with lots of other projects. So I thought of a middle ground - I still like RSS feeds and there are many useful sources of what I feel are 'informed' audio information - obviously AudioScienceReview being one of the most informed - so I got claude this time to generate a scheduled feed monitoring tool in python that would look at a bunch of websites that I would configure to find new news stories and update a small local DB of candidate stories for review - this tool looks something like this:

Because I don't want this to cost me money I have a local LLM Gemma4:12b running locally on my MacBook Pro - and the schedule tasks takes each story and asks this Gemma model to extra the website and summaries the contents so I can preview as per the image above. 'Approve' moves the generated summary (with the link to the stories, article or review) into my 'live' news to publish on the main website content - Hugo content is published by pushing content to gitlab.
I ended up spending probably an extra hour, refining the summarisation prompt used to generate suitable 'news' content from the various sources I picked. I also wanted to 'pin' certain stories e.g. the one against 'Capra Audio' that is now pinned on the website. So it got a little more complex.

6. Some interesting 'Informed Articles' for the website ( 2 Hours ).
So this will be the most controversial thing I feel in this article. So one tool I have been separately playing around with is thread summarisation - basically finding a good way to get LLM's to understand how to properly extract useful information from long threads - and I had written a extraction tool and summarisation tool for threads on ASR and some other websites. So for today I decided to create 'articles' based on those threads - so this is done via a 'Skill' which tries to structure the article into something relevant extracting the main topics from the longer thread into hopefully good paragraphs ( to be honest this is still a work in progress but I thought I would use it this morning ). The initial version of these was too short so I expanded the prompt and then expanded the skill to ask Codex to generate images and some diagrams to include in the content ( automatically without me prompting ) and to flesh out these still fairly short articles a little. Basically 4 of the classic threads from ASR and one from the headphones.com forum were used as examples ). The most fun was seeing the images generated from the topics extracted from the threads !!

7 Final tweaking ( 1 hour )
So, after all this AI fun - I decided to review what I had and change lots of things around - this is where I probably should have spend more time originally but for example, the new suggested layout for reviews didn't match my PragmaticAudio layout so I had to ask Codex (OpenAI) to generate the missing bits etc and the layout of the different sections needed again lots of tweaks. But in the end as a quick experiment in what could be done in a few hours I thought it was pretty good.

I will hopefully be keeping this update to date - that was the idea of the daily news edition tool being developed - the reviews should be shared with ones that I was doing on pragmatic audio and the articles will be updated as I make that thread summarisation better and some up with other articles or topics that would be good for a website called 'Informed Listening' !!

Here is the 'final' website after 6 hours:

Informed Listening

Audio reviews, measurements, news, and technical analysis judged by engineering and listening performance.

www.informedlistening.com

For those who won't want to click here is what it looks like:

And I made that banner stories be dismissible - so the reviews and news become more visible:

 

[Jeromeof]

Ok maybe this is another example of useful AI - so I have a few friend 'groups' that chat on Signal ( and WhatApp) about various stuff (and occasionally meet in person for some 'pints') - but what we talk about in general is interesting and it always fascinated me, as some people were very engaged in politics, current affairs whiles others like myself are way more interested in History or Technology and obviously futuristic stuff like AI) - so I thought a visualisation might be interesting for the group chat but then after about 10-15 minutes of prompting I got a nice visualisation which with a extra few minutes got converted into a simple 'website'. I feel this is really an early example of 'Generative UI' but I was fairly impressed with the overall simplicity of the process - though I feel OpenAI will 'own the intellectual property' of this content as that has to be the goal of 'free hosting' these result of this 'chat session':

The Time Horizon of Curiosity

Where does your curiosity point in time: the urgent present, the deep past, or possible futures?

past-present-future-disciplines.jeromeof.chatgpt.site

 

[eduardw]

At work, we have a private agreement with an AI language model that guarantees our company data remains confidential. As a network and firewall engineer, it acts as a valuable co-pilot—helping me troubleshoot and navigate multi-vendor platforms efficiently. I use it to generate specific CLI commands, analyze system logs, interpret diagnostic output, and even parse Wireshark capture files. It significantly speeds up my root-cause analysis and bridges the gap when I’m working on less familiar platforms.

It helped me to fix some network traffic problems related to mtu and mss problems.

I also used AI to make the above text better.

 

[janbth]

Ok maybe this is another example of useful AI - so I have a few friend 'groups' that chat on Signal ( and WhatApp) about various stuff (and occasionally meet in person for some 'pints') - but what we talk about in general is interesting and it always fascinated me, as some people were very engaged in politics, current affairs whiles others like myself are way more interested in History or Technology and obviously futuristic stuff like AI) - so I thought a visualisation might be interesting for the group chat but then after about 10-15 minutes of prompting I got a nice visualisation which with a extra few minutes got converted into a simple 'website'. I feel this is really an early example of 'Generative UI' but I was fairly impressed with the overall simplicity of the process - though I feel OpenAI will 'own the intellectual property' of this content as that has to be the goal of 'free hosting' these result of this 'chat session':

The Time Horizon of Curiosity

Where does your curiosity point in time: the urgent present, the deep past, or possible futures?

past-present-future-disciplines.jeromeof.chatgpt.site

Do I understand it correctly that OpenAI agents can get access to encrypted chats on Signal and WhatsApp, which you agreed to, and this agent produced a sort of summary?

 

[Jeromeof]

Do I understand it correctly that OpenAI agents can get access to encrypted chats on Signal and WhatsApp, which you agreed to, and this agent produced a sort of summary?

No - what I meant by this is I want to create an AI image that would I thought be interesting for these chats about politics and history etc ( so as content to post into the chat) but in the end ChatGPT offered this 'Site' option so the content I prompted to create the image was then used to generate the little webpage / website and I shared that link with the private chat - just to get people's feedback.

It was just an experiment like lots of things with AI and I imagine this content will disappear at some stage.

 

[janbth]

No - what I meant by this is I want to create an AI image that would I thought be interesting for these chats about politics and history etc ( so as content to post into the chat) but in the end ChatGPT offered this 'Site' option so the content I prompted to create the image was then used to generate the little webpage / website and I shared that link with the private chat - just to get people's feedback.

It was just an experiment like lots of things with AI and I imagine this content will disappear at some stage.

Ok, got it!

 

[DVinylfan]

"Claude users are creating public share links, but probably don't realize that means their chats are now ending up in Google searches where anyone can dig through them."
"Anthropic did not immediately respond to a request for comment." Lol

Tons of Peoples’ Claude Chats and Creations are Exposed on Google

Claude users are creating public share links, but probably don't realize that means their chats are now ending up in Google searches where anyone can dig through them.

www.404media.co

 

[Power Pop 23]

Researchers examining the cognitive skills of Capuchin monkeys in Costa Rica rely on customized A.I. to recognize the monkeys amidst other animals.

gift link

In This Costa Rican Forest, Monkeys Come Face-to-Face With A.I.

CapuchinAI, a portable testing station equipped with a touch screen and facial recognition software, could help scientists study primate intelligence in the wild.

www.nytimes.com

 

[Power Pop 23]

The manufacturer of Pringles is deploying sensors at a production line in Poland to recognize differences in batches of potatoes and relying on A.I. that has been trained to recommend tweaks to production parameters to ensure a more consistent product from batch to batch.

gift link

https://www.wsj.com/tech/ai/inside-the-long-ai-powered-quest-to-perfect-pringle-making-ab37a231?st=Lt2Dvv&reflink=desktopwebshare_permalink

 

[Somafunk]

s to ensure a more consistent product from batch to batch.

Consistently utterly shite and only fit for lighting the fire, , for this they are very suitable

 

[Steve Dallas]

I am an IT guy old enough to be aging out of the game. It looks like my last hurrah will be widespread AI implementation at my company. The bad news is I have replaced 84 people so far with things like automated AP, automated demand and supply planning, etc. with more eliminations in the pipeline. The good news is the remaining people (around 300) are more productive and efficient.

For example, I needed to create and implement ITIL-compliant change management and new demand management processes to fit within the tools and people we have available. Normally it would take about 2 weeks to design and document and another week to implement. It took about 2 hours with Claude and Copilot Enterprise. I spent roughly 1 more hours editing the documentation and executive summary slides before presenting it to the ELT. Then I used Zendesk's built-in AI tools to create the fields, forms, views, queues, triggers, and workflows, which took about an hour.

Another example is report writing. Our ERP is MS Dynamics, which has built-in AI tools. People just ask it to create reports for them and it does. (It does a ton of other useful things as well.) That eliminates the entire report writing lifecycle.

Then there is my software engineering group. They barely write code anymore, and their output is approaching 400% of what it was. No loss in headcount there yet, but I have been able to avoid hiring, which is bad for workers, but good for specific cells in spreadsheets that my bosses like to be as small as possible so they can get bigger bonuses.

 

[Osvaldo de Garay]

It would be very useful to build a AM FM receiver with AI embedded capable of muting or silencing the speaker while ads are being emmited, both recorded or read live. There are ads from 10 or more years old still playing on some radio stations. UNBEARABLES.