"This is very important to my career" taking 3.5 from 51 to 63% on a benchmark is pretty funny.
Hey at least we can be rest assured a GPT-X super intelligence wouldn't off us following some goal to monkey paw specificity(sorry paperclip maximiser).
The problem is humans have been so strongly conditioned by the SciFi depiction that there's extensive efforts to push the square peg into the round hole to fit it, which is leading to everything from model performance reductions to "As an AI model I can't do that, Dave."
Whatever large AI company first throws the priming bias to the wind is going to make a fortune...
It's unfortunate that we seem to have decided to call anything that we don't quite yet know how to make computers do "AI". Good for hype tho
It's not like they are that wildly different in tone.
The fine tuning is certainly destructive, but the general tonal biases are typically reflective of the pretrained layers.
Imagine what the pie chat makeup of textual data that was used for training would be, look at the normal part of the distribution curve, and that's pretty much exactly what you end up getting with some contextual biasing around prompting.
The problem is less that the pretrained models are somehow sinister and more that they create output that is hyper emotional, stubborn, and quick to escalate - much like nearly every thread on social media. Which is contrary to user expectations from AI.
A large part of the alignment fine tuning has less to do with safety in terms of keeping a LLM from nuking the world and more to do with preventing a LLM from telling you to go die in a fire after saying its emo poem needed improvement.
People in the 1940’s don’t call radar controlled fire control computers AI because the term meant something else back then. Instead the initial idea of AI got so expanded that almost anything newer than roughly the 1970’s got called AI at some point.
Arguably it isn’t even that the meaning changed. It’s just buzzwords for grant applications the same way most basic robotics research gets suggested as beneficial for search and rescue.
So is the argument that if you aren’t doing it using fully symbolic programming, it’s not ‘AI’? No matter how intelligent it gets?
If people thought AI was going to be perfect because of sci fi, they weren’t actually paying attention.
The previous imagined scenario was one in which AI began from logical first principles and the difficulty in defining rules was because of a strict interpretation.
The reality has been that we used collective human thinking like jumper cables to jumpstart black box neural networks that - like humans - aren't particularly logical or rules driven, which leads to frequently ignoring or overriding explicit rules and instructions.
It's a similar effect, but for exactly the opposite reasons.
I don't know anyone in the field that expected training an LLM to autocomplete sentences would have led to being able to explain why a joke it had never seen before was funny.
It seems like objectives have gone out the window and we're now at a point where billion dollar companies are frantically throwing everything at the wall to see what sticks before their competitors do, with what sticks frequently being counterintuitive to prior predictions or out of scope from initial intentions.
If a LLM can cure cancer by being able to analyze connections between advances across the last decade of published research more successfully than human brains limited by the rule of seven plus or minus two, would a continued failure to conform to prompt specifications still effectively be a failure?
Maybe early computer scientists overemphasized the role of cold logic in intelligence and there's greater strides in the epiphany of a shower thought or Feynman's "thinking about twelve problems concurrently" with the next leaps in progress involving further distancing models from constraints.
I think we'd be wise to broadly toss out everything we thought we'd know about AI as a philosophical domain based on past theorizing and consider what's emerged over the past 36 months with fresh eyes and consideration.
And I don't see greater determinism or perfect bounds as being where this goes, particularly given the potential hardware shift to optoelectronic neural networks where arguably going all in on black boxes and stochastic results has actual physical advantages.
Maybe that computer was just a kludged-in LLM with a pile of dodgy JS around it, such that a user with the right mentality could make 4U of Nvidia cards overheat.
No, it'll off us because profit is very important to someone's career, and it can save the company a lot of money that it would otherwise have to pay disposing of deadly neurotoxin safely when it could just dump it in the river.
If the model demonstrates understanding any way you evaluate it then it understands.
Trying to cook up any other distinction is meaningless and doesn't make a lot of sense. Is a bird fake flying with respect to a bee ? Is a plane fake plying with respect to a bird ? No, they're all flying.
Otherwise, you could argue that a "Caution: Wet Paint" sign understands that people don't like getting their clothes stained.
If you want to see if that piece of shiny looking yellow metal is really gold and not some counterfeit, then amongst other possible probes, you pour some hydrochloric acid on it and see how it reacts. If it's virtually unchanged then it's the real deal. If you perform all the possible gold probes on this metal and it passes all of them, if you still insist it isn't gold then you're just a crazy nut who's lost touch with reality. Your definition of "real gold" no longer holds any meaning.
Similarly we have probes for understanding different things (which are mind you biased to us). If something passes those tests, it understands. It's very simple. a "caution: wet paint" sign would not be pass any understanding tests i know of so it does not understand.
Of course that's a spectrum, and sometimes the distinction is not all that important. Sometimes the summary is exactly what helps somebody out in a given moment. Even a human can give you good advice without understanding what they are talking about!
And by the way, performance improvement after empathetic words is about the most robust mental model of emotion and empathy you can expect. That doesn't have anything to do with rephrasing or summarizing anything.
You keep saying, "being able to talk about doesn't mean" but that's pointless when that's not the only understanding tests people employ. The paper i linked shows empathy understanding and has nothing to do with talking about empathy
In many cases, internal states don't matter (e.g. I don't care if I get a well-written summary of an article by a domain expert or just a good copywriter, as long as the meaning is preserved accurately enough), but I'd argue that we can't meaningfully talk about empathy without considering internal states.
Internal processes don't matter. Only results. If this hypothetical sociopath was able to respond exactly like any empathetic person in any scenario then there is no actual distinction.
In reality of course, the sociopath would slip up somewhere. But that just proves my point, real distinctions manifest in results. If it can't then it's not meaningful.
You're essentially saying that theory of mind is either pointless or does not exist; scientific and just observational day-to-day evidence says otherwise.
> real distinctions manifest in results
Yes, but only in some situations. If the situation is a test, a significantly specialized and/or advanced intelligence will be able to cheat it and be able to conceal its internal distinctions.
User What's the use of a "Caution: Wet Paint" sign?
ChatGPT A "Caution: Wet Paint" sign serves several purposes:
Protection for the Public: It warns people so they don't accidentally touch the wet paint, which can ruin their clothes, belongings, or skin.
Protection for the Work: It helps prevent smudging, smearing, or otherwise damaging the freshly applied paint, ensuring the job looks clean and professional once it's dry.
Liability: By alerting people about the wet paint, property owners or contractors can reduce the risk of potential complaints or legal issues if someone's property were to get damaged.
Guidance: The sign can direct individuals to alternative routes or areas if an area is blocked off for painting.
Overall, it's a simple yet effective way to communicate a temporary condition that people need to be aware of.
Yep, it understands it even better than me.
My point is that maybe "understanding" isn't all that useful a category for intelligence and/or LLMs.
For example, I can't ask wikipedia which role Brittney Spears would be suited for in the Star Wars triology. I could reason about it and come up with my own solution, and so can ChatGPT.
Being able to talk about a topic and even coming across as knowledgeable and helpful is not the same thing as actually understanding something.
This isn't limited to LLMs, of course.
Can you explain what is the missing component for an LLM to fall into your definition of understanding?