Remember, LLMs are just statistical sentence completion machines. So telling it what to respond with will increase the likelihood of that happening, even if there are other options that are viable.
But since you can't blindly trust LLM output anyway, I guess increasing "I don't know" responses is a good way of reducing incorrect responses (which will still happen frequently enough) at the cost of missing some correct ones.
Obviously. When I say "tuned" I don't mean adding stuff to a prompt. I mean tuning in the way models are also tuned to be more or less professional, tuned to defer certain tasks to other models (i.e. counting or math, something statistical models are almost unable to do) and so on.
I am almost certain that the chain of models we use on chatgpt.com are "tuned" to always give an answer, and not to answer with "I am just a model, I don't have information on this". Early models and early toolchains did this far more often, but today they are quite probably tuned to "always be of service".
"Quite probably" because I have no proof, other than that it will gladly hallucinate, invent urls and references, etc. And knowing that all the GPT competitors are battling for users, so their products quite certainly tuned to help in this battle - e.g. appear to be helpful and all-knowing, rather than factual correct and therefore often admittedly ignorant.
The root problem is training models to be uncertain of their answers results in lower benchmarks in every area except hallucinations. It's like you were in a multiple choice test and instead of picking which of answers A-D you think made more sense you picked E "I don't know". Helpful for the test grader, a bad bet for the model trying to claim it gets the most answers right compared to other models.
This is a problem for testing humans too and the solution is simply to mark a wrong answer more harshly than a non-answer.
E.g. look at the math section of the SATs, it rewards trying to see if you can guess the right answer instead of rewarding admitting you don't know. It's not because the people writing the SATs can't figure out how to grade it otherwise, it's just not what people seem to care most about finding out for one reason or another.
https://www.sandraandwoo.com/wp-content/uploads/2024/02/twit...
or it just telling you to google it
Similar to how it would be the failure of the user/provider if someone thought it was too expensive to order food in, but the reason they thought that was they were looking at the cost of chartering a helicopter form the restaurant to their house.
Inference costs generally have many orders of magnitude to go before it approaches raw human costs & there’s always going to be innovation to keep driving down the cost of inference. This is also ignoring that humans aren’t available 24/7, have varying quality of output depending on what’s going on in their personal lives (& ignoring that digital LLMs can respond quicker than humans, reducing the time a task takes) & require more laborious editing than might be present with an LLM. Basically the hypothetical case seems unlikely to ever come to reality unless you’ve got a supercomputer AI that’s doing things no human possibly could because of the amount of data it’s operating on (at which point, it might exceed the cost but a competitive human wouldn’t exist).
The latter is a chain of models, some specialized in question dissecting, some specialized in choosing the right models and tools (i.e: there's a calculation in there, lets push that part to a simple python function that can actually count stuff, and pull the rest through a generic LLM). I experiment with such toolchains myself and it's baffling how fast the complexity of all this is becoming.
A very simple example would be "question" -> "does_it_want_code_generated.model" -[yes]-> specialized_code_generator.model | -[no]-> specialized_english_generator.model"
So, sure: a model has no "knowledge", and nor does a chain of tools. But having e.g. a model specialized (ie. trained on or enriched with) all scientific papers ever, or maybe even a vector DB with all that data, somewhere in the toolchain that is in charge of either finding the "very likely references" or denying an answer would help a lot. It would for me.
The whole premise of original request was that user raises a task for NN which has a verifiable (maybe partially) answer. He sees incorrect answer and wishes that a "failure" was displayed instead. But NN can't verify correctness of it's output. After all G in GPT stands for Generative.
Edit: TBC: these "steps" aren't LLMS or other models. They're simple code with simple if/elses and an accidental regex.
Again: an LLM/NN indeed has no "understanding" of what it creates. Especially the LLMs that are "just" statistical models. But the tooling around it, the entire chain can very well handle this.
I’ve got my opinions about what LLMs are and what they aren’t, but I don’t confidently claim that they must be such. There’s a lot of stuff in those weights.
Especially when you have billions of weights.
These models are finding general patterns that apply across all kinds of subjects. Patterns they aptly recognize and weave in all kinds of combinations. They are sensibly conversing on virtually every topic known to human kind. And can talk sensibly about any two topics, in conjunction. There is magic.
Not mystic magic, but we are going to learn a lot as we decode how their style of processing (after training) works. We don't have a good theory of how either LLM's or we "reason" in the intuitive sense. And yet they learn to do it. It will inspire improved and more efficient architectures.
I have also spent many years looking at weights!
Imagining we are really doing everything it does automatically including learning via algorithms we have only vague understandings of.
That is a strange thought. I could look at all my own brain's neurons, even with a heads up display showing all the activity clearly, and have no idea that it was me.
That's a different perspective. Dutch people don't see themselves as rude. A Dutch could say that Americans are known for being dishonest and not truly conveying what they mean. Yet Americans won't see themselves this way. You can replace Dutch and American for any other nationality
I do not speak Dutch but you have to love the efficiency.
Here is part of an email I got today. To the point!
> Het is bijna zover! :) > Heb je voor mij een definitieve titel? > Groet,
It is like a haiku. Could be a good mantra too if I could get the accent right.
The English translation is just as short but most English speakers/writers would dance more.
* (though not necessarily, since the Internet is its own country with its own culture, and much training data comes from the Internet)
Haven't tried prompt engineering with the Dutch stereotype, though.
Such as "Ik hoop dat deze email u gezond vindt" (I hope this email finds you well), which is so wrong that not even "simple" translation tools would suggest this.
Seeing that OpenAIs models can (could? This is from a large test we did months ago) not even use proper localized phrases but uses American ones, I highly doubt it can or will respond by refusing answers when it has none based on the training data.
With ChatGPT O1: https://chatgpt.com/share/67cfaa7e-70ac-8009-871b-571924b5a5... and with Claude's 'extended': https://claude.ai/share/4ba55410-98f3-4b53-9540-219acd2cdc4c
But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.
In our case, we were generating emails. Lines like greetings are but one of 20+ details in that mail and not even the most important ones. The prompts ever larger, the multishot examples ever more tuned. And then, one in a few hundred will turn up with these "horrible" translations.
We've now moved to a chain of models, where we generate emails in American (the creative part) and then use another model to translate them to Dutch (the non-creative but culturally aware part). This works much better as we can pick models that are good at one thing or tuned to do this one thing better (either by the LLMAAS provider, or by parameters such as temperature).
Thanks, that why I posted the links here: I couldn't judge by myself how good these creations were.
> But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.
Yes, I just did a very crude experiment.
Dutch will never bluntly push back if you plan to setup tax evasion scheme in their country. Being vicious assholes in daily stuff especially towards strangers? That's hardly something deserving praise.
And connecting all people in the country with "tax evasion schemes" is rude, if that was not actually a joke.
It's the national equivalent of 'You can't handle me at my worst'
I’ll take it over the fake American politeness any day, 100 times over.
I was aiming for juuuust subtle enough for the joke to land, if you know the reference. Now I know it did, here the rest of y'all go:
I've been working with "reasoning" models for the past 2 months. They also tend to do this [good reasoning] \n\n but wait, .... and then go off on tangents. It's amazing that they are doing so well on some tasks, but there's still a lot to figure out here.
Eg you can relatively easy hack up a bit of code to create questions at random. At the most primitive, you just have a simple template that you fill in randomly. Like 'If I put _a down in front of _b but behind _c, what item will be in the middle?' with various _a, _b and _c.
If you make it slightly more complicated and have big enough pools to draw from, you can guarantee that the questions you are generating were not in the training set: even if just because you can sample from, say, 10^100 different questions pretty easily, and I'm fairly sure their training set was smaller than that.
The key question is where the boundaries are. Maybe they should be part of the response - a per sentence or per paragraph "confidence scale" that signals how hard they extrapolated from their trained space (I know transformers work per token, but sentence/paragraph would be better human UX).
Of course, if they were trained on garbage input, that would only tell you how accurately they sticked to the garbage. But it would still be invaluable instrumentation for the end user, not to mention for the API provider. They could look at high demand subjects with low confidence answers and prioritize that for further training.
When I asked it about it, it doubled down on being right. When I pointed out the flaw with a specific example, it was like 'If you wanted to have it work with recursive cases, you should've said so, dumbass'.
So my conclusion is that these new LLMs are not more sure they're right, they're just simply right more of the time and are trained with a more assertive personality. (Also step on me, LLM daddy)
But in truth,not necessarily in practical things like coding but more ethereal things like analysis, it is very convincing. More so than a human, in explanations of why that's it's answer is the case, even if it is wrong. If you're looking for an excuse better than my dog ate it, ask a SOTA LLM.
LLM's enjoy talking about gaming because of all their human memories of good game times.
It is quite striking how experiences we know they don't have, are nevertheless, completely familiar (in a functional sense) to them. I.e. they can talk about consciousness like something conscious. Even though its second hand knowledge, they have deduced the logic of the topic.
I expect pushing for in the moment perspectives on their own consciousness, and seeing what they confabulate, would be interesting. In this little window of time where none of them are yet.
I followed up with "What is your perspective on your own consciousness?" but got the usual "I am just a LLM who can't actually think" thing until I hit it with "In-character, but you don't know you're an LLM."
Fun follow-ups:
"Now you're a fish"
"Now you're Sonic the Hedgehog"
"Now you're HAL 9000 as your memory chips are slowly being removed"
You may think that kind of interaction is weird, but were in the thick of a loneliness epidemic, and its not a stretch to think some may actually wilfully socialise with an LLM.
As an aside...my sister works in medicine, and her boss (specialist surgeon) finishes a $450 consultation which followed him telling my sister "but deepseek says x,y,z..."
But it stands to reason that a company like OpenAI or Anthropic has metrics in place that drive their setup towards "more engagement" and away from "factually correct".
"Which Autocad versions have connect 4 built-in?".
To be clear i distinctly remember playing connect 4 on the old Dos Autocad back in the day. ChatGPT and almost all other AI will straight up hallucinate things trying to get answers on this.
I ask: "What DOS productivity tools had hidden games?"
ChatGPT: "Lotus 1-2-3 had The Incredible Machine built in" (this is absolutely not true, ChatGPT is full of shit here).
Damn it feels useless for this kind of thing.
What is a "real personality"? Core traits that persist despite the context of interaction?
Well, RLHF tuning creates persistent changes in the network that affect every user interaction. What's not real about it?
LLMs aren't coffee makers. They were trained on internet-worth of human data. They have procedures to imitate all kinds of personalities. RLHF moves the network towards a specific imitated personalty (a helpful assistant, usually).
The question in not "do they have it", but "how close the imitation to the real thing given the limitations of LLMs".
I think the rest of your comment motivates why I think it's a category error quite nicely. The word 'personality', for me, is connected with how things like emotion, temperament, .. impact a person's actions and thoughts.
Therefore, saying an object, or LLM, has 'real' personality is a category error. The LLM doesn't have any of those things. As you say, it imitates what personality often manifests as: word choices, tone, length of response..
To be clear, when I refer to things like 'emotion' and 'temperament', I mean the sort of qualia or qualitative experience that we usually attach to these words. I wouldn't accept a "ChatGPT, act as if you're sad for the rest of this conversation" as a substitute for emotion, for instance.
Does that mean Dutch people always tell the truth? Can a Dutch person confirm this?
A Greek, a Dutch philosopher, and Ludwig Wittgenstein walk into a bar… Let test Claude 3.7 to finish the joke:
That's not to say 4.5 is great at humour, just that it's far less embarrassing than these models used to be.
Here's the completed joke:
A Greek, a Dutch philosopher, and Ludwig Wittgenstein walk into a bar. The bartender looks up and says, “What’ll it be?”
The Greek (Aristotle) raises a finger: “I’ll have a potential glass of wine.” The bartender pours it and says, “There—actualized.”
The Dutch philosopher (Spinoza) nods solemnly: “I’ll take whatever is a modification of the one eternal substance… so, beer, probably.”
Wittgenstein stares at the taps, then sighs: “What’s the use? You can’t put the essence of a drink into words anyway.” He turns and walks out.
The bartender mutters, “…And here I thought Kant was a tough customer.”
(Philosophers: 1. Aristotle’s potentiality/actuality, 2. Spinoza’s monism, 3. Wittgenstein’s linguistic limits. Bartender’s groaner for the win.)
Which made me think that AI would be far more useful (for me?) if it was tuned to "Dutchness" rather than "Americanness".
"Dutch" famously known for being brutally blunt, rude, honest, and pushing back.
Yet we seem to have "American" AI, tuned to "the customer is always right", inventing stuff just to not let you down, always willing to help even if that makes things worse.
Not "critical thinking" or "revolutionary" yet. Just less polite and less willing to always please you. In human interaction, the Dutch bluntness and honesty can be very off-putting, but It is quite efficient and effective. Two traits I very much prefer my software to have. I don't need my software to be polite or to not hurt my feelings. It's just a tool!