I think to some extent it’s a matter of scale, though. If I advertise something as a calculator capable of doing all math, and it can only do one problem, it is so drastically far away from its intended purpose that the meaning kinda breaks down. I don’t think it would be wrong to say “it malfunctions in 99.999999% of use cases” but it would be easier to say that it just doesn’t work.

Continuing (and torturing) that analogy, if we did the disgusting work of precomputing all 2 number math problems for integers from -1,000,000 to 1,000,000 and I think you could say you had a (really shitty and slow) calculator, which “malfunctions” for numbers outside that range if you don’t specify the limitation ahead of time. Not crazy different from software which has issues with max_int or small buffers.

If it were the case that there had only been one case of a hallucination with LLMs, I think we could pretty safely call that a malfunction (and we wouldn’t be having this conversation). If it happens 0.000001% of the time, I think we could still call it a malfunction and that it performs better than a lot of software. 99.999% of the time, it’d be better to say that it just doesn’t work. I don’t think there is, or even needs to be, some unified understanding of where the line is between them.

Really my point is there are enough things to criticize about LLMs and people’s use of them, this seems like a really silly one to try and push.

We’re talking about the meaning of “malfunction” here, we don’t need to overthink it and construct a rigorous proof or anything. The creator of the thing can decide what the thing they’re creating is supposed to do. You can say

hey, it did X, was that supposed to happen?

no, it was not supposed to do that, that’s a malfunction.

We don’t need to go to

Actually you never sufficiently defined its function to cover all cases in an objective manner, so ACTUALLY it’s not a malfunction!

Whatever, it still wasn’t supposed to do that

The purpose of an LLM, at a fundamental level, is to approximate text it was trained on.

I’d argue that’s what an LLM is, not its purpose. Continuing the car analogy, that’s like saying a car’s purpose is to burn gasoline to spin its wheels. That’s what a car does, the purpose of my car is to get me from place to place. The purpose of my friend’s car is to look cool and go fast. The purpose of my uncle’s car is to carry lumber.

I think we more or less agree on the fundamentals and it’s just differences between whether they are referring to a malfunction in the system they are trying to create, in which an LLM is a key tool/component, or a malfunction in the LLM itself. At the end of the day, I think we can all agree that it did a thing they didn’t want it to do, and that an LLM by itself may not be the correct tool for the job.

Where I don’t think your argument fits is that it could be applied to things LLMs can currently do. If I have an insufficiently trained model which produces a word salad to every prompt, one could say “that’s not a malfunction, it’s still applying weights.”

The malfunction is in having a system that produces useful results. An LLM is just the means for achieving that result, and you could argue it’s the wrong tool for the job and that’s fine. If I put gasoline in my diesel car and the engine dies, I can still say the car is malfunctioning. It’s my fault, and the engine wasn’t ever supposed to have gas in it, but the car is now “failing to function in a normal or satisfactory manner,” the definition of malfunction.

It implies that, under the hood, the LLM is “malfunctioning”. It is not - it’s doing what it is supposed to do, to chain tokens through weighted probabilities.

I don’t really agree with that argument. By that logic, there’s really no such thing as a software bug, since the software is always doing what it’s supposed to be doing: giving predefined instructions to a processor that performs some action. It’s “supposed to” provide a useful response to prompts, anything other than is it not what it should be and could be fairly called a malfunction.

Haven’t digital price tags been used for decades? I’m sure these will be more high tech, but I remember ones like this at least 20 years ago

Yeah, I just did a quick test in Python to do a tcp connection to “0.0.0.0” and it made a loopback connection, instead of returning an error as I would have expected.

I’ve heard the sentiment that change and convenience are killing society before, and I’m sure I’ll hear it again. I prefer to shop online. I get no sense of community from stores where every interaction has a hanging financial incentive around it, I get it from local organized runs, other frequent visitors of the dog park, etc. To me, that line of reasoning feels almost like lamenting how good the pipes in your house are, because you don’t need to call a plumber and get to interact with them.

Shopping online gives me more options, more reviews, easier ways to look up additional technical details without feeling weird taking space in an aisle while researching on my phone. It’s also more efficient in terms of total driving; one person making deliveries for everyone in a neighborhood requires less total driving than all those people making individual trips to a store. And it frees up more time for me to do things I actually want with the people I enjoy.

Looking completely realistic and being able to discern between real and fake are competing goals. If you can discern the difference, then it does not look completely realistic.

I think what they’re alluding to is generative adversarial networks https://en.m.wikipedia.org/wiki/Generative_adversarial_network where creating a better discriminator that can detect a good image from bad is how you get a better image.

These commercials https://youtu.be/qfv6Ah_MVJU?si=7uaDPCCRXXJ50_ES

Nice try, agent

In my experience, it has not generated results in real time. I’ve either gotten the exact same response, or a prompt asking “would you like to generate an AI response to your search?”

So it seems like, and would make sense, that in a given time period they only generate a response once per given search, and reuse that response in the future, since that’s far more efficient

I think that would work, and that’s essentially what I was trying to say when I’d said

What you would want instead is for everyone to post a (salted) hash, and after the hashes are posted, reveal what the original numbers were and then publicly add them. Everyone could verify everyone else’s numbers against those hashes.

comment, as well as my other https://beehaw.org/comment/3531769

How do you do fair random pairing, though? If you are able to safely do that randomly, you might as well use that same method to do the random flip.

Edit: And even ignoring collusion, there’s still the issue of lying (or lying about lying). Only one of a pair would need to be a cheater for the system to fail, if the rest of the group is unable to determine which is the cheater.

I think you run into other issues, depending on OP’s meaning of “untrusted.” If people are paired off, whoever is in the last group to report can control the outcome. Either if there is a risk of collusion within the group or if one member doesn’t like what the outcome is going to be they can claim whichever of them is reporting the group outcome is lying, or the person reporting actuality could lie.

I think this vulnerability will come up most of the time when information is shared with only part of the group and not the entire group.

The last person would still decide the outcome. They could keep choosing values for whatever function until it produces their desired result and then post that.

What you would want instead is for everyone to post a (salted) hash, and after the hashes are posted, reveal what the original numbers were and then publicly add them. Everyone could verify everyone else’s numbers against those hashes.

I would go further to say that if “making someone do something” is the definition, literally any action taken by any government is authoritarian. If a government did not make people do things, it would functionally cease to be.

Specifically one of the imperfections is that the server and players are not trusted. If a player doesn’t like the result, they could claim the server lied about what number they had picked, or the server could actually lie. The remaining players wouldn’t know which one is telling the truth.

I think the responses with an encrypted/committed guess being made public, a public result, and then a reveal of the key, have it right for the scenario of people making guesses as to the result of a flip.

Re-reading your question, though, refers more to there being an agreed result for a group of people as opposed to checking a guess. I think this would require a bit of a variation. The trivial method would be to use the previous method and assign “correct guess” to heads and “incorrect guess” as tails, but this only works if you don’t believe that any two members are colluding with each other.

Another solution would be to have each member generate a random number and encrypt it, and post the encrypted value. After all have been posted, everyone posts the key to decrypt their number, and adds up all the numbers together and takes the sum modulo the number of options (2 in the case of a coin) and matches it with a predetermined mapping. For instance, if 1 is heads and 0 is tails, and the sum of the numbers is 63752, 63752 % 2 = 0 which is tails.

There are a couple gotchas to prevent errors. There has to be an agreed upon maximum number which is one less than a multiple of the number of options. For instance, if random numbers are allowed from 0 to 2 inclusively, there is a bias towards tails (0 % 2 == 0, 1 %2 == 1, 2 % 2 == 0). The other is the encryption algorithm would need to be chosen such that multiple keys can’t easily be created to provide different valid decrypts. This would also likely require some padding to the clear text, which could be achieved by some member of the group posting some arbitrary text first, and then all members appending that text to their number before encrypting it.

Huh… Will this message then get re-ingested by chatgpt? Did it just poison itself?