News for hackers. Stuff that matters.
Re:Gemini "duck" demo was not done in realtime or with voice by brookst on Thursday December 07, 2023 @03:15PM (#38561246)
Is it possible for humans to be wrong about something, without lying?
Re:Gemini "duck" demo was not done in realtime or with voice by windowshopping on Thursday December 07, 2023 @03:20PM (#38561330)
I don't agree with the argument that "if a human can fail in this way, we should overlook this failing in our tooling as well." Because of course that's what LLMs are, tools, like any other piece of software.

If a tool is broken, you seek to fix it. You don't just say "ah yeah it's a broken tool, but it's better than nothing!"

All these LLM releases are amazing pieces of technology and the progress lately is incredible. But don't rag on people critiquing it, how else will it get better? Certainly not by accepting its failings and overlooking them.

Re:Gemini "duck" demo was not done in realtime or with voice by stocknoob on Thursday December 07, 2023 @04:15PM (#38561911)
“Broken” is word used by pedants. A broken tool doesn’t work. This works, most of the time.

Is a drug “broken” because it only cures a disease 80% of the time?

The framing most critics seem to have is “it must be perfect”.

It’s ok though, their negativity just means they’ll miss out on using a transformative technology. No skin off the rest of us.

Re:Gemini "duck" demo was not done in realtime or with voice by rkeene2 on Thursday December 07, 2023 @03:36PM (#38561479)
If a broken tool is useful, do you not use it because it is broken ?

Overpowered LLMs like GPT-4 are both broken (according to how you are defining it) and useful -- they're just not the idealized version of the tool.

Re:Gemini "duck" demo was not done in realtime or with voice by freejazz on Thursday December 07, 2023 @04:00PM (#38561755)
Maybe not if its the case that your use of the broken tool would result in the eventual undoing of your work. Like, lets say your staple gun is defective and doesn't shoot the staples deep enough, but it still shoots. You can keep using the gun, but it's not going to actually do its job. It seems useful and functional, but it isn't and its liable to create a much bigger mess.
Re:Gemini "duck" demo was not done in realtime or with voice by rkeene2 on Thursday December 07, 2023 @07:09PM (#38563660)
So to continue the analogy, if the staple gun is broken and it requires you to do more than a working (but non-existent) staple gun BUT less work than doing the affixment without the broken staple gun, you would or would not use it ?
Re:Gemini "duck" demo was not done in realtime or with voice by windowshopping on Thursday December 07, 2023 @07:31PM (#38563849)
But nobody said they wouldn't use it. You said that. You came up with this idea and then demanded other people defend it.

I don't know why "critiquing the tool" is being equated to "refusing to use the tool."

I don't like calling something a strawman, because I think it's an overused argument, but...I mean...

Re:Gemini "duck" demo was not done in realtime or with voice by rkeene2 on Friday December 08, 2023 @01:38AM (#38566059)
I didn't come up with it nor ask anyone to defend it. I asked a different question about usefulness, and about what it means to him for something to be "broken".

My point is that the attempt to critique it was a failure. It provided no critique.

It was incomplete at the very least -- it assigned it the label of broken, but didn't explain the implications of that. It didn't define at what level of failure it would need to be to valuable.

Additionally, I didn't indicate whether or not he would refuse to use it -- specifically because I didn't know, because he didn't say.

We all use broken tools built on a fragile foundation of imperfect precision.

Re:Gemini "duck" demo was not done in realtime or with voice by freejazz on Thursday December 07, 2023 @08:22PM (#38564204)
I think you are missing the point. If I do use it, then my result will be a broken and defective product. How exactly is that not clear? That's the point. It might not be observable to be, but whatever I'm affixing with the staple gun will come loose because its not working right and not sinking the staples in deep enough...

If I don't use it, then the tool is not used and provided no benefit...

Re:Gemini "duck" demo was not done in realtime or with voice by rkeene2 on Friday December 08, 2023 @01:41AM (#38566082)
It's not clear because it is false and I believe I can produce a proof if you are willing to validate that you accept my premise.

Your CPU, right now, has known defects. It will produce the wrong outputs for some inputs. It seems to meet your definition of broken.

Do you agree with that premise ?

Re:Gemini "duck" demo was not done in realtime or with voice by freejazz on Friday December 08, 2023 @12:44PM (#38571798)
One has nothing to do with the other. There's no rule about all broken tools because they can be broken in different ways. What's so difficult about my hypothetical? I laid it all out for you.
Re:Gemini "duck" demo was not done in realtime or with voice by rkeene2 on Friday December 08, 2023 @02:19PM (#38572975)
I assumed you understood we reached the end of the usefulness of your hypothetical to the original analogy since, as you said, the tools can be broken in different ways. I tried to introduce a scenario that was more applicable and less theoretical so that we could discuss those particular points.

If we do somehow try to apply your analogy, it would indicate that the LLM output is flawed in a way we cannot scrutinize -- the hidden failures that we aren't detecting (why? It's not specified, I am assuming because we didn't check to see if the tool was "broken" and not meeting some unspecified quality-level; that is, it's an unknown unknown failure mode).

This doesn't really comport with the LLM scenario, where the output is fully viewed, and the outputs are widely understood (that is it is a known failure mode).

This is more closely related to a computing service -- of which you are an active user of. You are using a "broken" computer right now, according to your definition of broken correct ?

Re:Gemini "duck" demo was not done in realtime or with voice by freejazz on Saturday December 09, 2023 @12:59PM (#38584030)
>This doesn't really comport with the LLM scenario, where the output is fully viewed, and the outputs are widely understood (that is it is a known failure mode).

Yes, it does. People are regularly using LLMs to brief themselves on topics they are ignorant about. Are you actually serious?

>This is more closely related to a computing service -- of which you are an active user of. You are using a "broken" computer right now, according to your definition of broken correct ?

Doesn't really matter, because I'm not relying upon the computer for the verity of its functions

Re:Gemini "duck" demo was not done in realtime or with voice by freedomben on Thursday December 07, 2023 @03:48PM (#38561626)
I think you're reading a lot into GP's comment that isn't there. I don't see any ragging on people critiquing it. I think it's perfectly compatible to think we should continually improve on these things while also recognizing that things can be useful without being perfect
Re:Gemini "duck" demo was not done in realtime or with voice by freejazz on Friday December 08, 2023 @12:47PM (#38571831)
I don't think people are disputing that things can be useful without being perfect. My point was that when things aren't perfect, they can also lead to defects that would not otherwise be perceived based upon the belief that the tool was otherwise working at least adequately. Would you use a staple gun if you weren't sure it was actually working? If it's something you don't know a lot about, how can you be sure it's working adequately?
Re:Gemini "duck" demo was not done in realtime or with voice by bee_rider on Thursday December 07, 2023 @05:04PM (#38562464)
I think the comparison to humans is just totally useless. It isn’t even just that, as a tool, it should be better than humans at the thing it does, necessarily. My monitor is on an arm, the arm is pretty bad at positioning things compared to all the different positions my human arms could provide. But it is good enough, and it does it tirelessly. A tool is fit for a purpose or not, the relative performance compared to humans is basically irrelevant.

I think the folks making these tools tend to oversell their capabilities because they want us to imagine the applications we can come up with for them. They aren’t selling the tool, they are selling the ability to make tools based on their platform, which means they need to be speculative about the types of things their platform might enable.

Re:Gemini "duck" demo was not done in realtime or with voice by lxgr on Thursday December 07, 2023 @03:20PM (#38561331)
Lying implies an intent to deceive despite, or giving a response despite having better knowledge, which I'd argue LLMs can't do, at least not yet. It just requires a more robust theory of mind than I'd consider them to realistically be capable of.

They might have been trained/prompted with misinformation, but then it's the people doing the training/prompting who are lying, still not the LLM.

Re:Gemini "duck" demo was not done in realtime or with voice by hunter2_ on Thursday December 07, 2023 @03:50PM (#38561639)
To the question of whether it could have intent to deceive, going to the dictionary, we find that intent essentially means a plan (and computer software in general could be described as a plan being executed) and deceive essentially means saying something false. Furthermore, its plan is to talk in ways that humans talk, emulating their intelligence, and some intelligent human speech is false. Therefore, I do believe it can lie, and will whenever statistically speaking a human also typically would.

Perhaps some humans never lie, but should the LLM be trained only on that tiny slice of people? It's part of life, even non-human life! Evolution works based on things lying: natural camouflage, for example. Do octopuses and chameleons "lie" when they change color to fake out predators? They have intent to deceive!

Re:Gemini "duck" demo was not done in realtime or with voice by og_kalu on Thursday December 07, 2023 @03:28PM (#38561410)
Not to say this example was lying but they can lie just fine - https://arxiv.org/abs/2311.07590
Re:Gemini "duck" demo was not done in realtime or with voice by lxgr on Thursday December 07, 2023 @03:34PM (#38561467)
They're lying in the same way that a sign that says "free cookies" is lying when there are actually no cookies.

I think this is a different usage of the word, and we're pretty used to making the distinction, but it gets confusing with LLMs.

Re:Gemini "duck" demo was not done in realtime or with voice by og_kalu on Thursday December 07, 2023 @04:02PM (#38561779)
You are making an imaginary distinction that doesn't exist. It doesn't even make any sense in the context of the paper i linked.

The model consistently and purposefully withheld knowledge it was directly aware of. This is lying under any useful definition of the word. You're veering off into meaningless philosophy that has no bearing on outcomes and results.

Re:Gemini "duck" demo was not done in realtime or with voice by vkou on Thursday December 07, 2023 @03:29PM (#38561418)
Most humans I professionally interact with don't double down on their mistakes when presented with evidence to the contrary.

The ones that do are people I do my best to avoid interacting with.

LLMs act more like the latter, than the former.