You don't need to be an 'investor' to invest in Singletrack: 6 days left: 95% of target - Find out more
ChatGPT seems to be highly rated and AI is seen as the future, however how can it be used if it can't answer simple questions correctly?
I asked v5.0 to name the US states ending in the letter o.
It states that 'There's only one U.S. state that ends in the letter "o" - Ohio. It's got the distinction all to itself'.
Do Colorado, Idaho and New Mexico not exist as US states anymore?
How is this the future?
ChatGPT said:
Only one U.S. state name ends with the letter “o” — Ohio.
You said:
ChatGPT said:
You’re right — Colorado also ends in “o.”
So, the complete list of U.S. states ending with the letter “o” is:
-
Colorado
-
Ohio
You said:
ChatGPT said:
Exactly — Idaho makes it three.
US states ending in “o”:
-
Colorado
-
Idaho
-
Ohio
You said:
ChatGPT said:
Yep — New Mexico belongs on the list too.
Final list of U.S. states ending in “o”:
-
Colorado
-
Idaho
-
New Mexico
-
Ohio
You said:
ChatGPT said:
Here’s the full list of U.S. states whose names end with the letter “o”:
-
Colorado
-
Idaho
-
New Mexico
-
Ohio
progress....
Google's AI answer, includes disclaimer 🙂
"how many US states end in the letter O
- Colorado
- Idaho
- New Mexico
- Ohio
At least they did not claim that spaghetti-o belongs on the list.
I dunno, some people are very much more concerned about it. Stealing copyrighted artwork for example. Or how about 24 hour surveillance (and data for training) via wearable AI tech?
Or that the very name of OpenAI alluding to openness and sharing of knowledge (especially if you've ever been involved in the open Source Software community) is far from the reality of the corporation.
I’ve been using it quite a bit lately to help translate some French legal documents and mostly it’s been very useful but occasionally it gets the translation completely wrong. Not in some detailed grammatical way - more like it is looking at a completely different text.
It amazes me how often it produces completely false information. When I've played around with it to verify some facts over some particular interest of historical knowledge which are very familiar to me, and of which countless books and documents/ reports exist online... a huge dataset, and it inexcusably gets dates wrong, gets characters in events mixed up, not even close to what is a well defined historical artifact. Amusingly - I've told it it's wrong and it admits it made a mistake, tries again and still gets it wrong again...and again...and again. It rewrites history, gets it completely wrong and presents it as fact.
The more worrying bit is that without a deep historical knowledge, ie: a normal awareness... the first incorrect answer would be taken as fact. It may create a new generation who do the standard 5 mins of research online and declare themselves an expert who don't realise how wrong they're getting it...
Not convinced as of yet... it needs to be more transparent on verifying validity of sources, and provide an estimation of certainty of its output for the end user.
I'm using Claude to generate code and graphs for investment purposes and I'm very impressed. First attempt told me what Python modules to install and the graphs produced looked OK, but on closer inspection were not correct so I told it:
That code ran ok but the figures produced are different to the ones I get from the Trading View website. Your code for AAPL produced a MACD of -1.3071 and a Signal of -0.5633 but Trading View produced a MACD of -7.83 and a signal of-5.61
and the reply was:
You're seeing different values because TradingView uses Simple Moving Averages (SMA) for MACD calculation by default, while my code uses Exponential Moving Averages (EMA). This is a common source of confusion!
Let me update the code to match TradingView's default calculation
Job jobbed.
Not convinced as of yet... it needs to be more transparent on verifying validity of sources,
Quite often it just makes up its own sources. There are several examples I've seen where it's cited references that simply don't exist.
It may create a new generation who do the standard 5 mins of research online and declare themselves an expert who don't realise how wrong they're getting it...
I remember my uni days of combing through the archives, finding obscure German chemistry experiments from the early 1900's, having to verify and cite the reference (obviously the lecturer knew it and expected you to cite it - failure to cite it, or getting it wrong was a clear case of failing to do the research or copying someone else for example).
Now, everyone spends 3 minutes on Google and produces an answer which is unquestionably wrong (sometimes dangerously so) on a worryingly large number of occasions.
You actually have to be quite good at what you're researching to know that it's wrong and question it again. Anyone without that knowledge is just going to go "ah well AI says it's [thing]..."
The mis-information and dis-information potentials are pretty horrifying.
It may create a new generation who do the standard 5 mins of research online and declare themselves an expert
That came long before the AI stuff.
My company has introduced its own internal Chat GPT. Its the ChatGPT LLM using RAG to answer queries based on our internal documents.
We are actually measured on how often we use it - essentially we are training the AI model for the company so no doubt some of us can be replaced. Pretty dystopian.
essentially we are training the AI model for the company so no doubt some of us can be replaced.
They won't replace staff. They will have the staff do something else. This has been demonstrated over the last 200 years. Most things that people used to do have been replaced by machines, but most of us still have jobs.
I think the main risk is that low paid low skill staff could be replaced, but someone will think of something else for them to do. The question is, will it be worse than what the AI now does?
I tried.
What's worrying / annoying is that it gives objectively wrong answers with such utter conviction.
I've been doing that for years!
Ask it to answer the question before last, that should be interesting
I've asked ChapGPT (that's the posh, English version 😀 ) and it still says there is only one state. So I got it to work out that there are 4 and told it to remember this as people are mocking it. It says it will remember in future.
That's a thing in itself. It's supposed to be machine learning - why isn't it learning from its own previous mistakes?
Granted there's a (probably fairly high) risk that people will deliberately contaminate it, but in the example above I didn't give it any new information other than telling it it was wrong.
I've just done the same and once it got the right decision I asked it why it got it wrong:
Because you'd need to retrain the model itself which is the most expensive part of running them.
I wonder how long before we'll see paid ads in the all the answers from chatgpt. I mean Google's already gone massively downhill with al the paid sponsorship.
All LLMs do is determine the next word on the basis of probability. The word apple could follow the, but so could orange. It’s a bullshit machine. Sometimes it guesses right, often it doesn’t. If you believe Orwell 2+2=5
I just used it to fix a SQL script and it did a great job 😀
All LLMs do is determine the next word on the basis of probability. The word apple could follow the, but so could orange. It’s a bullshit machine.
Exactly. An LLM isn't intelligent in the sense of understanding the meaning of language, it's just very good at identifying highly probable responses to prompts.
They won't replace staff. They will have the staff do something else. This has been demonstrated over the last 200 years. Most things that people used to do have been replaced by machines, but most of us still have jobs.
I think the main risk is that low paid low skill staff could be replaced, but someone will think of something else for them to do. The question is, will it be worse than what the AI now does?
They've already started, thinned out management as each manager is now supposed to be able to handle more direct reports because of AI.
My fear is that they'll looked to replace the expensive tenured staff who actually know the answers to most scenarios without having to rely on AI and replace them with cheaper less experienced staff and let AI cover the gaps.
I wonder how long before we'll see paid ads in the all the answers from chatgpt. I mean Google's already gone massively downhill with al the paid sponsorship.
IMO it will be about 0.00001 seconds after OpenAI figure out how to do it and sell the ads.
OpenAI is haemhorraging money and has no realistic path to profitability, unless Mr Altman can convince enough government bodies to embed ChatGPT into their infrastructure and then yank up the price.
They'll be desperate to monetize it in any way they can. For regular users, it's just a commodity - people will happily jump ship to Gemini or Claude if they feel they offer better value.
GPT 5 was supposed to be the big reveal, the gamechanger. But it doesn't look like it's much of an improvement on 4, although I believe it is cheaper if you're a power user, and has smaller versions than can run locally on (relatively) accessible home hardware (ie computers that cost about £3k). So it's a step forward, but it won't have the cash flooding in.
If you consider the volume of erronous shite on the Net it's not surprising that code developed to skim it (AI) produces unreliable answers. My browser AI clearly skims the first results and summarises them when the truth may lie a few pages beyond.
They've already started, thinned out management as each manager is now supposed to be able to handle more direct reports because of AI.
IME the opposite is happening. The thinning is occurring at the bottom because who needs junior staff anymore?
I can't say as I'm overly shocked either. A company is unlikely to thin the herd by getting rid of managers when it's the managers who are deciding who gets made redundant. Also IME.
Do Colorado, Idaho and New Mexico not exist as US states anymore?
According to some of the people I meet Colorado actually ends in an 'a'
A company is unlikely to thin the herd by getting rid of managers when it's the managers who are deciding who gets made redundant. Also IME.
Its a big company (multinational IT vendor) its got a lot of management layers. To their credit, a lot of staff reductions focus on management bloat first but I get the distinct impression they will start to go further.
A company is unlikely to thin the herd by getting rid of managers when it's the managers who are deciding who gets made redundant. Also IME.
Its a big company (multinational IT vendor) its got a lot of management layers. To their credit, a lot of staff reductions focus on management bloat first but I get the distinct impression they will start to go further.
As with all current AI related things, it's less about the AI and more about how you phrase your query. There is a real skill involved in formatting a query correctly to get the answers you need, when done properly it is a very useful tool for research and many other uses.
Mrs H1ghland3r heads up the AI development team at a major IT firm, and they have their own models that are being trained to be used for document review, coding and technical documentation. They are currently aggressively hiring people specifically to correctly format AI queries. As was said earlier, the jobs won't be lost, they'll just change.
In this instance, if you ask 'Can you methodically check and list all US states that end with the letter O' then it gets the right answer first time.
The key word in the query is 'methodically' as this instructs the AI to bypass the various shortcuts it uses to reduce processing time (and cost) and do a full review of the data to answer the question.
As with all current AI related things, it's less about the AI and more about how you phrase your query.
Which is fine, if you understand that, but the average user of Facebook or Google? no chance. That LLMs are bullshit machines really is an issue. In the right context they can be useful tools. Though an observation at the current costs I can see a lot of companies walking away. The pricing models are a joke.
Which is fine, if you understand that, but the average user of Facebook or Google? no chance. That LLMs are bullshit machines really is an issue. In the right context, they can be useful tools. Though an observation, at the current costs I can see a lot of companies walking away. The pricing models are a joke
Which is also part of the perception problem, no-one in AI gives two figs what the 'freemium' users and public are doing with the AI models which are generally 2-3 versions behind the cutting edge. They are all focussed on the corporate development as that's where the money is. After that, then it will filter down to consumers. Not even the models Google are shoehorning into every search query are where their focus is, which is part of the problem, they want everyone to be thinking about incorporating AI into everything, but the current state of the public faced AI is so bad it's ruining the perception of it as a tool.
I also have issues with them even using the term 'AI' as it's nothing of the sort, it's just a natural language search engine based on a snapshot of data from a given point in time, as the US state question highlights, it can't even update its own parameters when a mistake is pointed out to it, mistakes have to be reviewed and updated by real people at OpenAI or wherever the models are trained.
Possibly a case of user error here but the article highlights some very real risks with AI.
https://www.independent.co.uk/news/health/chatgpt-medical-advice-hospital-b2803992.html
Possibly a case of user error here but the article highlights some very real risks with AI.
Yep, none of this sort of thing is helping, but I'd be curious to see EXACTLY what he asked to get that answer. I suspect the answer he got is more to do with what he asked than anything else. Like I said before, businesses are hoovering up people trained to ask properly formatted queries. It's more like writing an SQL query of a database than chatting with a person.
I've mentioned before that I work for Microsoft, and somewhat relevant to part of the discussion on corporate/enterprise use we've just published a story with my customer SSE.
Compared to public ChatGPT there's a huge amount of extra work to build something that's reliable enough to use in a corporate environment, particularly when it's public facing. Some companies get this, some less so.
https://www.microsoft.com/en/customers/story/24657-sse-microsoft-copilot-studio
Compared to public ChatGPT there's a huge amount of extra work to build something that's reliable enough to use in a corporate environment, particularly when it's public facing. Some companies get this, some less so.
My employer has only just granted access to copilot to a small subset after many hours chatting to MS because of this. Still along way to go though before it gets rolled out to users at any level. And even longer until every possible complication is understood. There are business rules around sales and until everyone is 100% certain that users will not see information that they are not allowed, copilot stays off.
My employer has only just granted access to copilot to a small subset after many hours chatting to MS because of this. Still along way to go though before it gets rolled out to users at any level. And even longer until every possible complication is understood. There are business rules around sales and until everyone is 100% certain that users will not see information that they are not allowed, copilot stays off.
Absolute opposite at my place, CoPilot is being touted as the greatest thing since sliced bread. We're being encouraged to use it for all sorts, "try it out". Lots of breathlessly exciting articles (which sound exactly like they've been written by AI) on the intranet about how it can summarise documents, write your meeting minutes, organise your calendar...
It all *sounds* great:
oh why bother reading that long document, I'll get Copilot to summarise it...
oh why bother listening to the meeting, I'll get Copilot to write it up...
Until you realise that Copilot doesn't know the relationships of the people in the meeting, doesn't understand the technical stuff, mishears and misunderstands words, phrases and sometimes whole contexts, can't do sarcasm and is literally just providing a live speech-to-text. With occasional wild errors as it misunderstands a regional accent.
Possibly a case of user error here but the article highlights some very real risks with AI.
I've shared something similar before, but "AI psychosis" is also quite troubling.
crazy-legs - All you've said is true, it's an accelerator, not an answer. Verify the outputs, check the references.
mrmo - it can only see what you could see anyway. The main issue is that it can find things you didn't know you had access to. Most organisations aren't great at classifying information and controlling who has access.
mrmo - it can only see what you could see anyway. The main issue is that it can find things you didn't know you had access to. Most organisations aren't great at classifying information and controlling who has access.
We primarily use D365 F&O and there is paranoia around the permission models in place, a mix of Entra/AD on and off premises, and security within D365, PowerBI and SQL. So you can imagine how something might fall through a gap.
mrmo - it can only see what you could see anyway. The main issue is that it can find things you didn't know you had access to. Most organisations aren't great at classifying information and controlling who has access.We primarily use D365 F&O and there is paranoia around the permission models in place, a mix of Entra/AD on and off premises, and security within D365, PowerBI and SQL. So you can imagine how something might fall through a gap.
Gulp. Good luck! I'd be recommending an external specialist or MS services for that one.
I wouldn't ask ChatGPT how to solve it.
As with all current AI related things, it's less about the AI and more about how you phrase your query.
This.
I've been using it (Copilot) a fair bit at work. The most successful instances are when I drip feed requests and slowly build up to what I need.
My wife has been using AI to check maths problems and sometimes it makes terrible mistakes.
AI responses may include mistakes."
Hallucinations without the use of psychedelic substances! How can I do that?
The more worrying bit is that without a deep historical knowledge, ie: a normal awareness... the first incorrect answer would be taken as fact. It may create a new generation who do the standard 5 mins of research online and declare themselves an expert who don't realise how wrong they're getting it...
Sounds like your average Republican voter…
What's worrying / annoying is that it gives objectively wrong answers with such utter conviction.
I can do that, after significant amounts of alcohol.
There was talk that Apple was considering buying Perplexity, but it now appears that the AI is just as guilty of stealing information from copyright holders as all the others, so either the deal is off, or Apple buys it, then uses the technology to train it internally.
I don’t really use AI apps; I’ve found Perplexity handy occasionally for giving me answers that would otherwise be full of random garbage or links to advertising sources, but I can’t be arsed to sit playing around creating random content or art - I’d rather read a book.
Quite often it just makes up its own sources.
Manchester Uni did some work looking at how they can detect plagiarism by students who've used Chat GPT and other LLM to produce essays, and they realised that they don't need to check the actual essay, they just go directly to the references and citations and check those instead. Students don't tend to make up books or authors, so it's pretty much 100% accurate. Although I understand now that the software that looks for plagiarism can now detect LLM use anyway.
So, it got there in the end without me giving it any answers. What's worrying / annoying is that it gives objectively wrong answers with such utter conviction. It might not be the future of scientific research any time soon but there will be a general election in a couple of years.
This is the greatest problem, it’s being foisted on an unsuspecting public as something it’s not.
There’s no intelligence behind it but to your average user it generates the appearance there is or someone who thinks they can lose staff/increase productivity gets excited.
Hallucinations without the use of psychedelic substances! How can I do that?
Non AI search says
https://www.sciencealert.com/how-to-create-hallucinations-without-drugs-surprisingly-easy-science
As with all current AI related things, it's less about the AI and more about how you phrase your query.
This.
I've been using it (Copilot) a fair bit at work. The most successful instances are when I drip feed requests and slowly build up to what I need.
My wife has been using AI to check maths problems and sometimes it makes terrible mistakes.
Ahh prompt engineers are so last weeks bandwagon 🙂
It’s a great tool but unfortunately still requires programming skills and experience to get what you want although it can save on a lot of typing and sometimes you do get some good ideas from it.
I find it’s useful as a really good help system but tbh as soon as your doing stuff that there’s not much material to plagiarise it can go well off the reservation.
Without already having the skills it’s pretty much back to the good old days of that monster departmental access db app that someone coded years ago which everyone’s too scared to touch as it’s the flakiest code thang ever 🙂
What's worrying / annoying is that it gives objectively wrong answers with such utter conviction.
I can 100% guarantee that it does not.
I can 100% guarantee that it does not.
How so? I can understand the semantics of querying the truth relationships between predicted words... but, as in my example, if it fails to get the date correct within a few decades of a well known and extensively digitally documented historical event, something that occurred without questioning. Literally hundreds of books written on the subject, and reams of web available reports from the time- then how can we describe it as anything other than 'objectively wrong'? It is very odd that it makes this mistake given the quantity of reliable consistent data available from archives.
If it's getting confused over solid events from history then what else is it hallucinating? In my mind I call it the F.I. 'Fake Intelligence'.
Think you misunderstood the previous post.
It seems the Altman bubble is slowly deflating,
Think you misunderstood the previous post.
Oh, I thought we were getting in to some philosophical semantics about how we determine what is real. Maybe it was just a joke.
The thing is... the misplaced Techno Optimism amongst the silicon avant-garde is so pervasive it is hard to differentiate - but yeah, that's pretty much what I look like.
The thing is... the misplaced Techno Optimism amongst the silicon avant-garde is so pervasive it is hard to differentiate
Yep, it's pretty hard to tell if a lot of stuff on the net is a parody or someone who is genuinely an idiot.
oh why bother reading that long document, I'll get Copilot to summarise it...
This irks me , as the some of the online newspapers seem to be using this at the top of every story.
Its hardly like they need it as the stories are hardly that long.
Why bother writing a long article in the first place.
Seems to just be a way of making people dumber or un-necessary use of ai.
I used to love Apple News but now find it disheartening as it’s so bad, I’m just finding all the online news seems to be getting worse.
Seems to just be a way of making people dumber or un-necessary use of ai.
But ‘summarize this’ is the killer app of LLM’s few-trick pony! You’ve got to use it, otherwise what have the billions been spent for? 😉
Why bother writing a long article in the first place
As Blaise Pascal is alleged to have written
Je n'ai fait celle-ci plus longue que parce que je n'ai pas eu le loisir de la faire plus courte.
[I would have written a shorter letter, but I did not have the time].
Constructing a concise and informative article would take skill and time. Many organizations are unwilling to pay enough for the former and ‘pushing content’ limits the latter. These pressures have always been present in journalism but seem accentuated in the current ‘attention economy’.
there are some good articles out there still.
Its all going to be Ok, Sam Altman introduced Chat GPT 5 a few days ago - which claims to now be able to make mistakes at PHD level- with an image likening it to the Death Star... everything's going to be just fine.
[img] 
[/img]
Me: "How many teeth on this coupling?"
Gloried Simon Says: 60
Me: hmmm...{counts 50}. ****ing waste of time.
But it is good for wasting megawatts of energy making useless images
[img] 
[/img]
[img] 
[/img]
So the cut to the chase, is anyone here using AI for commercial gain. Either to increase revenue/profit, reduce costs or reduce risk?