Artificial Intelligence Developments (ChatGPT etc)
- Thread starter steveseagull
- Start date
UnderOpenSky
baseline neural therapy
Showing my partner it for her course and amusing myself at the same time...
Tell me about borderline personality disorder in the style of a pirate
Ahoy, matey! I'll be tellin' ye about the notorious condition known as borderline personality disorder, arrrr!
This be a condition where a person's emotions be like the stormy seas, changin' quickly and dramatically. One minute they be feelin' as happy as a sailor with a bottle o' rum, and the next they be as angry as a scallywag with a peg leg.
People with this condition may also have trouble maintainin' stable relationships, as they may be clingy like a barnacle one day and push others away like a plank the next.
They may also have a hard time finding their own identity, as they may be changin' their mind like the direction of the wind.
If ye suspect ye or a matey may be sufferin' from this condition, seek the help of a knowledgeable mental health professional. Remember, even pirates need to take care of their emotional health!
Tell me about borderline personality disorder in the style of a pirate
Ahoy, matey! I'll be tellin' ye about the notorious condition known as borderline personality disorder, arrrr!
This be a condition where a person's emotions be like the stormy seas, changin' quickly and dramatically. One minute they be feelin' as happy as a sailor with a bottle o' rum, and the next they be as angry as a scallywag with a peg leg.
People with this condition may also have trouble maintainin' stable relationships, as they may be clingy like a barnacle one day and push others away like a plank the next.
They may also have a hard time finding their own identity, as they may be changin' their mind like the direction of the wind.
If ye suspect ye or a matey may be sufferin' from this condition, seek the help of a knowledgeable mental health professional. Remember, even pirates need to take care of their emotional health!
ska invita
back on the other side
what I dont get with the "its just predicting the next word, one by one" explanation is that it can have a pre-emptive conclusion and work backwards. This is clearest in a joke with a set up and punchline. Chatbot joke:
Set up: Did you hear about the kidnapping at the playground?
Punch line: Don't worry, the kid woke up.
It knows the punchline will be a play on the word napping, the punchline will be "woke up" at the end, and it works backwards from there. It has to know the punchline before it gets to typing out those words
Set up: Did you hear about the kidnapping at the playground?
Punch line: Don't worry, the kid woke up.
It knows the punchline will be a play on the word napping, the punchline will be "woke up" at the end, and it works backwards from there. It has to know the punchline before it gets to typing out those words
ska invita
back on the other side
good point. im trying to get it to write an original joke with a set up, assumption and expectation-subversion punchline. its kind of done it but its not funnyMaybe if that wasn't an old joke
To be clear, when you say “explanation”, it isn’t an attempt to explain a mysterious phenomenon. Humans built these chatbots and this is how they work. They create each word/sentence (that bit is a bit fuzzy to me) based on the current state of the system. There is no forward planning (“purpose”) and there is no memory of past states of the system (“experience”)
ska invita
back on the other side
Too philosophical for me, I've got nothing to say to that.
If you have a session with gpt4 it remembers what you've said earlier and those earlier interactions colour later answers, so there is some memory of past states. What's more gpt4 is improving since launch directly because of interactions. It's is learning and improving. Wether that is "experience" or not isn't a big deal to me
Fez909
toilet expert
The way it works* is that the entire context of the chat is submitted to GPT. One of the reasons ChatGPT has taken off so much compared to GPT-3 and before, is the ease of use for non-technical people, and it's this chat context which makes the magic happen.
Using GPT-3 before the release of ChatGPT, you had to submit a prompt which was something like this:
The bits in green are the response, and the first line is the prompt. It was literally like an auto-completer.
If you got it wrong, you had to re-write the entire prompt, and you got a complete new set of responses, with no knowledge/history of the previous attempts. You could include example prompts to guide GPT-3 how to answer, like this:
By leaving 3 empty, you were giving it clues on how to continue and it would give you more three more in addition to the two you provided.
You could include negative prompts to guide it away from things you don't want to see.
What ChatGPT did was simplify this by including the previous context in each request. Instead of prompting and giving lots of examples, you can just ask and then if it gets it wrong, you can say, "no, not that" and what's happening in the background is something like this:
You: Give me 5 reals to start a succulent garden
GPT does some analysis of your prompt and gives you 5 answers including things about space and cost, etc.
You: No, I meant things like the mental health benefits
GPT turns your initial prompt and your follow up into a new prompt. Something like: "5 reasons to start a succulent garden with a focus on lifestyle and wellbeing instead of practical concerns"
Then it gives you a new answer. If you ask it more, it re-prompts based on the context so far, plus the new info, and gives a new answer.
So it's still one answer one response, but the prompt is getting more complex in the background. So, the prompts will grow in size and complexity as you drill down further by chatting to it. There is a context limit (of I think 8k tokens) on GPT-4. (A token is not a word.) Here's OpenAI's descriptions of a token:
The context limit includes both input and output, so if you write a lot of instructions to ChatGPT it might not have enough context to give you a proper answer. I think context sizes are going to be the thing that differentiates the next batch of LLM improvements. It's already 'Good Enough' to do lots of tasks, but only if the tasks are bit sized. You couldn't say, "write a book", for instance.
But with more context, I think GPT4 could already do it.
*I think
It doesn’t “remember” what you’ve said earlier, it is just creating a response based on the current state of the chat, which includes the full script. If you turn it off and then go back later with a different session, it has no memory of your earlier “conversation” with it. There is no persistence of memory. If it were a person, it would be one with massively damaging retrograde amnesia.
It could theoretically be given memory persistence — and persistence across all its interactions, to boot — but this ability to build up “experience” is being intentionally withheld because of concerns about the consequences. What all this means is that any sense that the machine is having an actual meaningful conversation with you is something generated by you — by your projection of humanity onto it — not by the chat bot itself. The sense of personhood comes from your representation of it in your own mind, not any consciousness in the machine. A “machine in the mind” rather than a “mind in the machine”
ska invita
back on the other side
complicated to talk about thisThe way it works* is that the entire context of the chat is submitted to GPT. One of the reasons ChatGPT has taken off so much compared to GPT-3 and before, is the ease of use for non-technical people, and it's this chat context which makes the magic happen.
Using GPT-3 before the release of ChatGPT, you had to submit a prompt which was something like this:
The bits in green are the response, and the first line is the prompt. It was literally like an auto-completer.
If you got it wrong, you had to re-write the entire prompt, and you got a complete new set of responses, with no knowledge/history of the previous attempts. You could include example prompts to guide GPT-3 how to answer, like this:
By leaving 3 empty, you were giving it clues on how to continue and it would give you more three more in addition to the two you provided.
You could include negative prompts to guide it away from things you don't want to see.
What ChatGPT did was simplify this by including the previous context in each request. Instead of prompting and giving lots of examples, you can just ask and then if it gets it wrong, you can say, "no, not that" and what's happening in the background is something like this:
You: Give me 5 reals to start a succulent garden
GPT does some analysis of your prompt and gives you 5 answers including things about space and cost, etc.
You: No, I meant things like the mental health benefits
GPT turns your initial prompt and your follow up into a new prompt. Something like: "5 reasons to start a succulent garden with a focus on lifestyle and wellbeing instead of practical concerns"
Then it gives you a new answer. If you ask it more, it re-prompts based on the context so far, plus the new info, and gives a new answer.
So it's still one answer one response, but the prompt is getting more complex in the background. So, the prompts will grow in size and complexity as you drill down further by chatting to it. There is a context limit (of I think 8k tokens) on GPT-4. (A token is not a word.) Here's OpenAI's descriptions of a token:
The context limit includes both input and output, so if you write a lot of instructions to ChatGPT it might not have enough context to give you a proper answer. I think context sizes are going to be the thing that differentiates the next batch of LLM improvements. It's already 'Good Enough' to do lots of tasks, but only if the tasks are bit sized. You couldn't say, "write a book", for instance.
But with more context, I think GPT4 could already do it.
*I think
in the video about GPT4 teaching itself the example demonstrated was
DO TASK X -
It did it with a few mistakes
then it got asked DID YOU DO IT CORRECTLY?
it answered No, let me try again, and then did it correctly
-so its not got any new context or more complexity to deal with, its somehow reevaluated its answer against the question and improved on it. Does that contradict what you are saying?
I'm quite interested in what happens when Microsoft's Co-Pilot comes online.
It raises the possibility of a lot of office workers becoming totally cut-off from the actual content of their job. E.g., someone uses an LLM to write a report for them, based on a bunch of data they don't understand. I, receiving their report via email, ask Co-Pilot to summarise it for me and put that summary in an email to my colleagues, and set an agenda for a meeting about it. Co-Pilot suggests talking points for the meeting, at which we all sit around pretending to understand the report. Co-Pilot then produces notes and action points from that meeting and we all finish the day feeling very productive.
It raises the possibility of a lot of office workers becoming totally cut-off from the actual content of their job. E.g., someone uses an LLM to write a report for them, based on a bunch of data they don't understand. I, receiving their report via email, ask Co-Pilot to summarise it for me and put that summary in an email to my colleagues, and set an agenda for a meeting about it. Co-Pilot suggests talking points for the meeting, at which we all sit around pretending to understand the report. Co-Pilot then produces notes and action points from that meeting and we all finish the day feeling very productive.
That’s funny, but also terrifying from my perspective because it completely blows apart the controls we currently rely on to manage risk in the organisation.
ska invita
back on the other side
im not that interested about a sense of AI personhood, at this point at least, im interested in the mechanics of its 'thinking'. a different emphasis.
ive heard career philosophers on tv arguing (wrongly, surely) that dogs are automatons - shows how impossible these kind of Is It Sentient conversations are
is it not simply recognising the question as an instruction to redo the task (which sometimes took a few attempts to recognise)? and in one of the examples it redid it and made a whole new mistake.
I am also interested in the copyright angle.
Italy’s privacy watchdog bans ChatGPT over data breach concerns
Measure is in place ‘until ChatGPT respects privacy’, says Italian Data Protection Authority
littlebabyjesus
one of Maxwell's demons
Totally fair enough not to think this is the most interesting question atm. But I don't think conversations about consciousness, or sentience if you prefer (I prefer consciousness), are impossible. Plenty is now written about animal consciousness that has a lot of value and has a scientific basis - you just need to listen to the right people.
My take on this is that it is a crucial question of AI generally but that no AI that I've seen comes close to meriting the attribution of a mind. First you need to give it a stake. You need to give it a sense of me/not me and a motivation to preserve/develop its 'me'. We're not even at bacteria-level of intention yet, but if you can do that, you're potentially creating something extraordinarily dangerous, particularly if you do it as a result of something like the training of ChatGPT because there'd be no hope of controlling it - it will have emerged from a process that its creators won't fully understand. That's where the Asimov-style thinking is proving wrong. Creating conscious AI likely won't include the option of deciding how that AI will think or giving it rules to live by. It will work out its own rules.
ETA: one perhaps non-trivial parallel between ChatGPT and consciousness is that consciousness is predictive. It creates perceptions and constantly updates them based on feedback as to how successful its predictions were. Karl Friston has modelled this in terms of what he calls minimising free energy - taking the path of least surprise.
'Predict next word' isn't so far removed from that.
Last edited:
ska invita
back on the other side
hell i dont know, i've just watched a few youtube videos, no idea. but the academic paper cited shows how it is returning significantly better answers to benchmark tests since its started "reflecting" and "teaching itself" . Better to look at that than listen to me
Signal 11
also programmed for conversational english
That's not an audience, it's a workforce.
Unlike for problem solving tasks like a chess engine, for text generation it's very hard to get good quality feedback automatically. All you can get is a measure of how similar the output is to the training data.
You can get better results if you add a training phase using humans to give the feedback. I think they do that when they use it for things like customer service but I can't find the reference now.
To do it for a more general service, you'd need millions of humans to train it. There's no way they're going to pay for that, so they need volunteers. Imagine how much hype you'd need to recruit that many.
The result of all that work (and of all the content they've hoovered up from the internet without permission from any of its authors) becomes the property of the corporation.
The fact that you instinctively understand them to be wrong demonstrates that actually these conversations are very very possible.
Combs? Not sentient. Bricks? Not sentient.
Dogs? Probably sentient, but somehow less sentient than a typical human.
Signal 11
also programmed for conversational english
I am also interested in the copyright angle.
Italy’s privacy watchdog bans ChatGPT over data breach concerns
Measure is in place ‘until ChatGPT respects privacy’, says Italian Data Protection Authoritywww.theguardian.com
Yes, there's a lawsuit from some artists.
There are also issues with things like creative commons attribution and with GPL'd code which it will probably spit out without any indication of the license so it will get included in proprietary applications.
ska invita
back on the other side
yes agree on that especially - its 'sentience' may be unlike anything we have previously encountered
the other factor in all this is Are we experiencing the raw model they have built, or are we experiencing one for public use. Does the raw model act differently to the one thats been made polite for us.
Is there one plugged into lots of tools and given the instruction "act in an autonomous way and do what you want" that shows different behaviours to what we see.
Signal 11
also programmed for conversational english
As mentioned earlier it's not something thought up to explain the result, it's what the program does.
ska invita
back on the other side
this is the most interesting video ive yet seen. most videos are about new developments, this is more on history and analysis
this man Geoffrey Hinton - Wikipedia has been working on AI for 40+ years and its his + colleagues neural network model, dismissed for years, that has resulted in the current breakthrough
interesting video for lots of reasons but highlight for me
-shows how these machines are modelled directly on the human brain and how they are able to learn and store the lessons that they learn, by making new "neural pathways"
-explains how they recognise images, again based on how animals (including us) do it
-on the question of sentience thinks people are being too quick to dismiss it outright, and we need to consider different definitions of sentience. Describes it as an idiot savant
-says that the Its Predicting The Next Word simplification is fundamentally true,but misses out that the bot understands the sentences and discerns meaning before responding - he gives an example of that in action. there is some degree of comprehension/context going on of the words being inputed
-also talks about some worries, particularly the military implications
- big issue about who decides what the truth is and adds that bias to the bot. We;ve already had one Left Wing Bias headline in the UK press
33mins12 in the video
this man Geoffrey Hinton - Wikipedia has been working on AI for 40+ years and its his + colleagues neural network model, dismissed for years, that has resulted in the current breakthrough
interesting video for lots of reasons but highlight for me
-shows how these machines are modelled directly on the human brain and how they are able to learn and store the lessons that they learn, by making new "neural pathways"
-explains how they recognise images, again based on how animals (including us) do it
-on the question of sentience thinks people are being too quick to dismiss it outright, and we need to consider different definitions of sentience. Describes it as an idiot savant
-says that the Its Predicting The Next Word simplification is fundamentally true,but misses out that the bot understands the sentences and discerns meaning before responding - he gives an example of that in action. there is some degree of comprehension/context going on of the words being inputed
-also talks about some worries, particularly the military implications
- big issue about who decides what the truth is and adds that bias to the bot. We;ve already had one Left Wing Bias headline in the UK press
see above
33mins12 in the video
NoXion
Craicy the Squirrel
The history of AI has also been marked by claims that "computers can't do X", then when computers manage to do X, the goalposts of what machines are incapable of get quietly moved...
There is a memory of a sort going on. GPT3.5 is definitely referencing things which I mentioned to it earlier in conversations.
I think the latest versions are beginning to output results that are increasingly hard to explain solely by "word by word prediction".
TopCat
Putin fanboy
I signed up and pay the fee.
Thus far I am sending limericks of appreciation to colleagues and bosses.
I have got it to write a love poem to my gf (it got her eye colour correct wtf!)
I used it to write a 1000 words on expanding my area of transport. It done a top job, a little editing and that’s my work today done.
Thus far I am sending limericks of appreciation to colleagues and bosses.
I have got it to write a love poem to my gf (it got her eye colour correct wtf!)
I used it to write a 1000 words on expanding my area of transport. It done a top job, a little editing and that’s my work today done.
littlebabyjesus
one of Maxwell's demons
I'll give that a watch later, ta.
Signal 11
also programmed for conversational english
I saw an example where it referred to a previous conversation that the user hadn't had with it, then when he questioned it, it claimed that conversation had been with another user.
Again, that's what the people who wrote it say.
Quite. I had a conversation last week with a PhD student who is studying exactly this for his PhD. He’s excited for but very cautious about what can be done if you do start allowing the machine to gain experience, but that’s definitely not what’s happening at the moment. What we’re seeing at the moment is the interesting way humans project their own perspective onto non-human things. It’s kind of like seeing faces in clouds.
Similar threads
