Unfortunately LLMs don’t have actual “intelligence” as such, they are basically statistically driven word generators. Because their job is to generate responses they descend into waffle and bullshit when they don’t know the answer rather than saying “I don’t know”.
They also have no intrinsic understanding of good/bad, right/wrong, etc.
Yep I’ve tried to explain much the same to son in law but despite being bright enough to be an actuary, he is prone to some bizarre magical thinking when he feels like it.
Don’t get me wrong, LLMs have their uses and it can be a very good tool to generate some stuff that is onerous to produce by hand. But it’s not a replacement for true expertise (yet).
I have some sympathy here - up to now, computer systems have by and large been deterministic. For sure, bugs exist in software (and occasionally hardware), but they are normally consistent, repeatable things.
AI is very much non-deterministic, which is both its strength and its weakness. Where it becomes a challenge is we’re very used to being able to rely completely on putting 2+2 into a calculator or spreadsheet and getting the answer 4, but that’s not actually something an AI can do 100% reliably.
Tell me you don’t understand how LLMs work, without telling me you don’t understand how LLMs work (that’s directed at Michael Samadi, not you Kev!)
The human desire to anthropomorphise is proving a real headache when it comes to people talking about LLMs.
It’s not something that a fair few people can do 100% reliably either, so maybe the AI is more like us than we’d like to admit 
I think that’s the point. It’s a tool and suitable for certain use cases, but not others.
An interesting case is self-driving cars. People get fixated on a zero error rate. But if the AI (or whatever you want to call it) is better than a human driver overall and the accident rate drops (albeit not to zero) then everyone’s better off, no? It may kill a few people, but so do human drivers.
True.
But then the (non-computable ?) delusion of personal bias kicks in. On average, human drivers kill 0.000x people per mile. However, I am so far above average that I am happy to trust myself in that environment. I am so careful that I will never hit anyone ever and I am so capable that if I come across some nutter then I will always be able to get myself out of whatever trouble he causes. So I’m happy to drive myself but I won’t trust myself to a machine which is ‘only’ better than average.
Yes, delusion.
Do self driving vehicles believe they can “accelerate their way out of trouble?”
While it’s reasonable to be irritated at AI’s patchy handling of facts and logic, more concerning is the way the algorithms seem to be skewed to be a ‘people-pleaser’. I guess there’s a fair bit of DNA either derived or learned from social media which helps shape it into an echo-chamber, but some of this has to be deliberate since people are always more inclined to accept as fact anything that reinforces their pre-existing notions - no matter how ludicrous, or awful …
Truth is AI could potentially be almost anything - good, bad or ugly - but all it really needs to do is simply persist long enough for humans to become habituated to it, and then it will become as hopelessly entangled with our lives as (e.g.) the internet is. Once anything becomes universal and mundane, it becomes invisible.
In terms of manipulating and propagandising us, the scope is as vast as it is alarming. It looks ever-more a totalitarian wet dream.
My wife insists on saying please and thank you when using Chat GBT. When asked why she said: “one day they will run us all and I want them to at least know I am polite” … couldn’t think of a non expletive response - still can’t.
One day quite soon if enough of us fall for it.
There’s a mix of things. One of the major elements is that when when models are being evaluated, a chunk of that evaluation basically translates to random folk being asked to choose which option they prefer between two, without much critical evaluation. That’s been quite reinforced with the release of GPT5 - they deliberately made it less sycophantic etc, and a lot of folks got very vocal about how they didn’t like its “personality” compared to GPT4. To the degree that OpenAI is, I believe, tweaking the personality back somewhat. It’s pissed me right off because I hate the way they do drop this shit in.
I’m not quite so convinced of its ability to persist though. All the AI players have been burning through VC cash at the most unbelievable rate, and none of them are even remotely close to breaking even (I’m not sure any of them are covering even 50% of their ongoing costs via actual sales revenue). Anysphere, the makers of Cursor, a programming IDE with AI tools built in, is I think turning a profit. But crucially, they don’t make any of the actual models - they build software that accesses the models, gives them access to certain functionality within the software, and presumably a degree of prompt engineering, but they leave the crucial bit to others.
The wife’s work think that AI is going to do their jobs for them, so she’s responsible for setting up the back end.
This morning she has delivered her judgement, and I quote “I was prepared to be surprised, but no, Copilot is just fucking Clippy on steroids. Shut the fuck up. I don’t give a shit.”
Bringing you infra with attitude since 1991.
Meanwhile in “try to shoehorn an existing wish on the latest buzzword”, the TUC is saying that AI should mean that workers have a seat on the board according to the news on Radio 3.
Still no idea how that works for owner ma aged businesses (i.e. most of the employers and employment in the country).
I use Copilot at work.
While I agree with Mrs G, a lot of people at work would have no idea what Clippy was.
To them Copilot is a new tool and in some cases works quite well.
I have used Copilot to write minutes of Teams meetings, and even though you have to check/amend them afterwards it is still quicker than writing from scratch.
I have also found it useful to turn a very factual email into something our Customer Service team is happy to send out (it is all about the tone).
Like any tool it has it’s place.
This is another pet peeve that results in more f-bombs. People doing exactly this, it Makes Shit Up or misses the basic point, but they don’t review the results.
As a complete ignoramus (OK, maybe I should just go and read up if I care) how does that work in terms of confidentiality ? Presumably Copilot eavesdrops on the meeting. Can it run stand-alone on a non-internet-connected machine or is it doing its job back at MS HQ (I’m guessing it is) ? Does this bother anyone ? Maybe if you’d written the minutes yourself you’d have stored them in the cloud so they’d be just as accessible ?
Come to think, maybe I’ve missed the obvious point that if you’re using Teams then MS could be listening in with or without Copilot.
