Pharmacies are already doing it. I had the displeasure of dealing with an LLM VOIP chatbot not long back.
It was obvious from the start it was some sort of agentic monstrosity. So I said nothing (im not giving that model my voice to train it) and hung up. Went down there myself and got my housemaids knee tonic.
It’s funny that the EU AI Act states that use of AI needs to be made transparent. No mention of it over the blower however…
I asked Google AI what instructions (rules) I should provide to avoid the serious problems with AI that I discussed above.
You can put rules into the settings in your user account so you don’t need to enter it for every AI session. As per my 2nd. post, I don’t want to use my Google account so I cut & paste my ‘rules’ from a file - it takes only seconds to do that. In fact, for complex engineering problems I prefer to compose & edit my question off-line (in the same file) and cut & paste it all together.
For straight-forward questions (e.g. “what is the gain of a Diamond RH-770 on 2m and 70cm?”) I usually don’t enter any ‘rules’ at all. To avoid the problem Richard @G4TGJ had with two different pinouts (one wrong). you should include “State your sources” so you can choose the best one.
I still use text books, technical articles, etc for reading around a topic. But with specific detailed questions / problem solving you often won’t find a specific answer or you take ages to synthesise one from multiple sources.
As someone said, AI is a tool, I use it only when appropriate. The purpose of my original post was to give a warning about using AI and not to defend or condemn its use.
Since when did the purpose of the original post have any influence on the direction of discussion on the reflector?!
Any time a powerful technology comes along there is a convulsion in the culture. Think the written word, the printing press, electronic media, etc. It gets wildly overused, misused and vilified. Up until now, eventually the tool finds it’s place and settles into a practical role.
I remember when Wikipedia came up on my radar screen. Wow, did it blow the doors off the World Book Encyclopedias that were in our living room when I was a kid. The refresh rate on new information wasn’t even close to that of an aging set of hardcopy. But Wikepedia is prone to the occasional sloppy and downright erroneous entries by questionable Wikipedians. The good news was the there was usually enough footnoting (hyperlinks) pointing to more in-depth research. It was, and remains, a great starting place in any research project.
I think LLMs serve a similar role, and provide a similar benefit. They are a great starting place in any research project.
I tend to suspect where AI is going to get really weird is in stuff like undetectable deepfakes, meticulously and laser-targeted psy-ops. For years I’ve seen advertising as rather insidious, subconscious psychological warfare. I think having a TV in your bedroom and falling asleep with commercial TV on is crazy. I would never have a TV in my bedroom. The propaganda potential of AI is staggering and it will take a very active, concerted effort to resist. I don’t think your average person has the stamina or the training to ward off the upcoming onslaught
Another country heard from
Eric KG6MZS
Which is better? Springy clothes pegs, or, the other kind? 
…hear hear 
Geoff vk3sq
I have a feeling we are already doing so. NHS111 on-line asks lots of questions one at a time - I don’t suppose there is a person selecting the next question after looking at your response.
The Soldersmoke guys had an informative discussion about using AI for QRP circuit design and diagnostics.
The current GenAI models can go on for an hour discussing a radio (or antenna) design problem with you. They cannot yet render beautiful schematics or drawings but they can do a reasonable job of interpreting one if it is cleanly drawn, just from a jpeg or screen scraped image.
Dean KK4DAS described dropping into Claude (not Copilot) digital oscilloscope images from a project he was working on and getting impressive diagnostic advice back.
The general putpose LLMs we are all using have been trained on the internet, and as gi-normous as that dataset is, it may still not be enough for the models to achieve human electronics/radio/RF design capabilities. But if the data is made available to them at scale, I expect they will, and exceed it.
That isn’t necessarily AI though. I suspect it has been programmed with the sequence of questions. It has been around since before the recent AI explosion.
I am programming my latest project in Rust which is a new language to me and also has quite a steep learning curve. When I have had an error message I didn’t understand I just pasted it into Gemini and got a really helpful explanation back.
Android phone dose similar to you for almost two decades 
You can still buy ordinary phone or few years old car.
I’m with you. I’m sure LLMs will be a great tool for some people. They may even get to the point where they are reliable. I still don’t care. The fun of engineering is in the problem solving and building things. For me AI isn’t threatening, it’s just boring. I imagine I have the same interest in AI that artists have in paint-by-numbers sets.
The internet is the largest source of misinformation on the planet. AI seeks to learn from that and then tell you what you want to know. There is a study from a number of years ago that shows that most people come to the internet to find information that reinforces their beliefs- not to learn something new. AI serves to continue that behavior.
My personal experience is that it can’t do the same not-very-difficult calculation twice in different requests using the same data. Then it will lie about why it can’t get the same result when challenged.
As others have said, you still have to think critically about what it’s providing, then double check it’s results.
If you really want to test an LLM, ask it for information about something or someone you know well, and then check how much of what it tells you is completely wrong. Assume it is going to do no better on subjects about which you don’t know much. Anyone got a longer barge-pole? Mine’s not long enough.
This is what I was warning about in my first post [#1 above]. I’m shocked by how often LLM’s get the simplest of calculation-based questions wrong. And when you challenge its garbage answer it either (a) attempts to defend it with some wacko reason (known as ‘AI hallucination’) or (b) say something like “You are right, I was incorrect”. Don’t think you’ve won some moral or intellectual victory in the latter case - its core programming is merely trying to avoid conflicts with the user.
That’s an example of its conflict avoidance. If LLMs were really cognisant (as opposed to just having the appearance of it) like the fictional HAL 9000 (in 2001: A Space Odyssey) you might describe the ‘lies’ as it having cognitive dissonance. But for ‘dumb’ statistical analysers like LLMs, that’s too grandiose a term - it’s more like it’s got lost. When this happens with me (usually after I’ve pushed back a few times on its crap answers or justifications) its prompt starts to include “Do you want to end the session?” If I were anthropomorphising, I might think it was trying to get rid of me.
It doesn’t ‘learn’ from its mistakes; if you start a new session & ask exactly the same original question, you will get Groundhog Day!
The ONLY way [it seems to me] to avoid it making (those frequent) mistakes in any calculation-based question is to tell it to do the alternate Python script method.
What an interesting and informative discussion!
Whilst my thinking about the role of AI in my own life is unchanged, it is clearly an emerging technology that might one day be relevant to me (or inescapable!), so I absolutely do feel the need to be informed.
What a profound comment. I find it somewhat alarming that businesses and individuals are piling into AI as if it is some sort of universal panacea. As if it removes the need for critical thinking. As if it is some sort of distillation of all human knowledge, unsullied by inaccuracy or doubt. As if it can actually replace employees. As if it can, somehow, make our lives better.
It seems to me that AI is a bubble, even more gigantic than the dot-com bubble, just waiting to burst.
usually my conversations with LLMs start with a question, it gives me an answer i tell it its wrong, it says well done for pointing that out your right here is the actual answer. i tell it its wrong still it says well done for pointing that out your right here is the actual answer. I tell it its wrong, its crap and give me the right answer it agrees its frustrating and it is wrong here is the confirmed correct answer. i tell its its wrong and go google it myself! its good in certain circumstances but anything remotely clever it is not. I’m not sure why the media call them artificial intelligence. They have no intelligence. Its a very large collection of nested if then else statements.
I have been using EE Help a lot in the past week. It offers me Chat and then it turns out to be AI chat - which then answers a question I did not ask or offers a selection of responses for me to choose (none of which takes my question forward) and blocks any typing until I make the selection. This is not really much help - current phone help is very good with real people in UK and speaking perfect English. I think AI has a long way to go.
Rod
Without a doubt. But AI will get better and better until the planet runs out of water and expires from global warming due to the massive data centres.
No it isn’t. AI systems are not programmed in that way. They learn from information they are fed.
The AI chatbots that I have tried for customer support have generarlly been as crap as your experience. That isn’t a fundamental problem of AI but a result of the models the companies are using. If they aren’t willing to invest in the processing power and memory required then they won’t be good.
We have Nadia at work, an AI chat system to help you prepare for man-management issues and improving interviews and one-on-one conversations. As I’m retiring (3 weeks including this one but who is counting!) I’ve been chatting with her, as I get paid, and trying to twist her melons by suggesting things like I want to go postal with my TT33 and throwing the odd Alice’s Restaurant line "I mean, I wanna—I wanna kill. Kill. I wan—I wanna see, I wanna see blood and gore and guts and veins in my teeth. Eat dead burnt bodies. I mean kill. Kill. Kill! Kill! to poison her. Doesn’t seem to have done jack but make me smile for getting paid for doing daft stuff. Maybe she’ll ask one of my colleagues if they want to kill and eat dead burnt bodies. It’s worth a try. I did this after reading in Private Eye how someone had pushed fake news out suggesting Hanta is Hebrew for fake so Hantavirus is “fake virus”. As some AI is trained on the mince and pish which is the internet, Hantavirus has already appeared in some AI systems knowledge.
I could say we’re doomed but actually if you use it properly AI systems can be a big help. But it helps if you already know the subject so you know when AI is hallucinating. I use VMWare Workstation Pro here to run a W10 virtual machine on my Linux desktop. VMWare still uses SysV scripts for startup etc. and SystemD has a tool (SystemD Generator?) that finds old SysV init scripts and makes SystemD units from them. This happens on every boot. There is a boot time message saying the generator is being deprecated so I thought I should make some proper ones as VMWare haven’t yet. I asked Copilot (we’re a big Copilot user at work) and I was very impressed with what it suggested… all looked good and made sense. But I never made a copy so when home asked Copilot again. It gave me the same SystemD units but this time also included the reference to where it had found someone’s checked in work on GitHub. So it was a cool and clever search engine rather than something that understands SystemD. Still it looked cool for a while 
…and sadly the big players seem, at least for their “general” public-facing large language models, to be happy to source training material from wherever they can find it, including (but not limited to) archives of pirated ebooks, fan-fiction websites, any (and every) web forum they can get at, and whatever else they can grab for free. I suspect something over 95% of the “knowledge” they’ve absorbed is fantasy. When all you want is a stochastic parrot to spout plausible text it doesn’t really matter that it’s fantasy, but it all falls a bit flat if you’re after actual accurate information.
I was amused by a rumour, recently, that some large company, having failed to put constraints on its AI usage, had managed to run up a bill for half a billion dollars in a month. However, I’m guessing that story’s fantasy on the basis that no sane company would play that loose with its finances, and no sane company would allow a customer to get that far into debt without at least ringing some warning bells…
