I wonder whether they are as shitty as the first ones and whether they allow to disable the annoying “battery low” warning.

Faster at being a shitty platform does not make it a better platform. Just a faster one.

Pregnant people must scare you. As do synthetic bacterial cells, artificially synthesized viruses, cloning, organoids, etc.

god-like isn’t religious mumbo jumbo. That’s standard sci fi talk.

I am talking rather about the Sci than the Fi.

Define “god-like” if you do not mean it in a religious manner.

Regarding your question: it makes sense to continue research and educate people, yes.

Are you really requiring that we allow research in one field only after we have advanced research in another entirely different field? That seems arbitrary.

Humanity can already create life.

Creating conscious entirely artificial things is a different pair of shoes.

I would cut out the whole religious “god-like” mumbo-jumbo though, because that does not necessitate any progress in such research directions.

The fruit fly may feel pain as a stimulus. Even if not: you are interrupting it’s normal biological function and disabling it of living its life normally. This may even be a disturbance of local eco-systems and thereby affect more forms of life, both feeling and not-feeling.

EDIT: Just realised that the following might be beside your point, misread your comment, lol. Will leave it as is now, because I think it’s worth considering when thinking about consciousness, experience, etc. in the context of LLMs.

But more importantly: LLMs are trained to model patterns and make predictions. This is indeed something our brains do as well. The difference lies in a state of mind and ‘understanding’ though. I will detail this:

Reminiscient of the chinese room thought experiment: Consider you are imprisioned in the chinese room and have to figure out yourself which chinese characters you have to slip back under the door depending on which ones you receive. (Or in any other language that you have absolutely no understanding of. I will assume chinese for this now.) If you provide a correct answer, you will get food. If not, you won’t and maybe even get additionally punished somehow.

Over time you will start to see patterns and remember which characters to correctly provide. However, you will never hear someone speaking, and you will never really associate the characters with things you know from life. For example, you wouldn’t notice you were talking about an apple when slipping 苹果 through. You would also not know how to pronounce it. You would just know this is the desired answer to draw and slip back.

Now, a native chinese person is sitting in a neighbouring room. And some experimenter who doesn’t know in which room you or the native speaker sits, will slip the same ‘input’ characters to both of you and may ask how the acidic-sweet, fruit is called that is round and grows on trees. Assuming you have trained long and well enough, both of you will provide the correct reply, 苹果, and from the perspective of the experimenter it will be impossible to distinguish who the real native speaker is and who you, the prisioner is.

Now, the chinese person would have had an image in his head when thinking about the apple, would know how it tastes like, etc… You on the other hand wouldn’t even have noticed that you were talking about apples. The 苹果, is just some symbol to you that seems appropriate.

For the experimenter though, they might think they were speaking to two real chinese speakers although you were just ‘faking’ it.

Next, we place a computer in the third room that uses our current LLM technology. Those are trained in a manner that is in principle similar how you were trained: roughly speaking, negative feedback for bad replies, positive feedback for good ones. A mathematical optimization algorithm then modifies the model such that the desired replies become more likely to happen.

After a while the experimenter returns and now has to distinguish behind which door the computer is, the prisioner, and the native speaker. It will be impossible. It may be easier to test with more difficult ‘questions’, there the native speaker might perform much better than you or the computer. But in theory, we assume the computer and you have learned the patterns perfectly. Then it really is impossible to tell who is faking it and who the ‘true speaker’ is.

But, there is still a difference. You are not the LLM. You got hungry, you may have even felt pain, you have wondered what those symbols mean, you may even have remembered some symbols that you found especially beautiful. And of course, you wondered when those maniacs would release you.

All of that does not happen in LLMs. They do not have a thought-loop. They do not feel anything. No pain, no hunger. They have no awareness of their situation and no sense of being or something that might resemble consciousness. They do not strive for anything, have no motivation or aspirations. They are fire-and-forget machines without any state of mind. Receiving input on one side, providing output on the other without any reflection or retention. Their only memory is how they should answer provided some input, since they, same as you, just modeled the patterns.

This can appear eeringly authentic. If the experimenter would have asked you on chinese what your life goals are, both you and the LLM would provide an answer that would be likely accepted by the experimenter. But both you and the LLM actually have no idea what you have just said. It would be fakery.

If the experimenter would have asked about what your deepest, darkest fears are, same thing: you would provide an answer that would be acceptable when someone asks about deep dark fears, but you would again have no concept of what you have just talked about. This doesn’t make you or the LLM good chinese speakers. It makes you convincing fakers.

So there are two levels of difference here between a human and a LLM as we currently have them:

1. The conversation is convincing fakery.
2. There is no experiencing subject behind the conversation. The LLM does not know that it is having a conversation, does not understand the meaning of what it says in the experiential sense described above, and does not feel or experience anything while producing its answers. There is no inner perspective to which the conversation is happening.

That may change in the future. But currently, there is no established evidence of an experiencing subject behind the conversation. Moreover, it is technically unnecessary, because nothing about how current LLMs operate requires subjective experience, awareness, or an inner perspective.

And this is precisely where the convincing fakery becomes deceptive, since language is normally one of the strongest ways humans infer another person’s state of mind. If someone tells you that they are scared, lonely, excited, or in pain, it is usually reasonable to infer that there is a person behind those words actually experiencing something. An LLM can produce exactly the same statements simply because they are appropriate responses in that context. The statement itself therefore cannot tell us whether there is an experience behind it.

A recording of your parents telling how disappointed they are. /j

Maybe just uneducated in how LLMs work. If you have no sufficient understanding of the technology it is easy to be impressed how human-like LLMs talk to you, almost magical.

It foreshadows an issue that will be a reality one day nevertheless. The work on AI continues. We may one day very well have conscious AI. This will be the start of a whole bunch of ethical and legal problems. It might be meaningful to regulate / guardrail such research directions.

Be kind. Absolutely. But anthropomorphising lifeless and consciousness-less objects is not always meaningful, sometimes even detrimental. As in the case with our current LLMs.

This guy is a SI. Super Idiot.

Given how many Anthropic fans there are, I am not entirely sure. Marking sarcasm in text can help.

New hires and comments in interviews show that the AI giant is using its capabilities against dissenters.

…

Read the damn article.

How this can become public:

You are dumb enough to use the share feature.

No leaks so far. But it is of course always a possibility. As it is with almost any other web service.

How this can get accessed by others without being explicitly public to everyone:

• Law enforcement, the justice system, and intelligence agencies
• Your conversations will be used for training LLMs of at least the company whose LLM you’re using. Data deals with others are a possibility.

I do not have the time to work through every part of the example, but imo the main claim is still overstated. Showing that a result would be extremely unlikely under a particular null model is not the same as showing that it is “statistically impossible” for a human to produce. It also does not give a guarantee how the text was written. A tiny p-value is still a probability under assumptions and not a proof of provenance.

Furthermore, a human does not even need to know the secret key. By pure chance a human written text can display an unusually high alignment with the detector’s secret partitioning.

The published watermark work, which is also cited by the article, appears to be much more careful about this (based on a quick skim). It reports false positive/negative rates, thresholds, length requirements, and more. Those can be very strong results, provided the assumed conditions apply. They do not turn a detector into an infallible test. Moreover longer text only helps if the assumptions and watermark signal actually remain intact, which can fail in general.

In such controlled settings, sure, I do not have much issues there. But the claims of “a human cannot replicate this” or “we can guarantee the text was generated with the watermark” are much stronger than the statistics, and especially the cited literature, actually appear to support.

Interactive demonstrations are not the same as a formal proof or experimental validation. So we shouldn’t attribute more to this technique than the available evidence can really support.

I found some time to quickly skim through the sources they have listed. And from that it became pretty clear that this is not realiable in detecting LLM generated versus human output in general. Under very tight assumptions specific error rates were reported that appeared rather low. However, these assumptions do not hold in general, even with more text if no relevant signal remains. There is currently no scientifically validated general purpose way of reliably detection.

More importantly in the context of Claude, the production watermarking scheme is undisclosed. Therefore, the cited experiments on known watermarking schemes can neither establish how reliably text generated by Claude can be detected, nor how reliably the technique described in the article removes the actual watermark.

It can be treated as an indicator at best, but not as validated proof.

But how reliable? With which guarantees? What are the prerequisites for this to work at all? Telling people from each other apart is one thing, the other is telling them reliably apart from a machine generated text.

But it does not show a sufficient formal proof and no experimental validation. Many important questions to evaluate the concept are left unanswered, which limits the interpretability and condenses it to “just trust me, bro, it’s a good idea, because I say so”.

Yeah, still, I wouldn’t claim “as no human would end up falling into that”, given that it may not be that unlikely to find at least one human who displays similar writing the more humans you involve.

Until a formal analysis is presented and an experimental study is published, which covers the most important influencing factors, the reliability of this concept is limited.

Although this looks like a clever approach, a kind of stochastic key, I do not see how this guarantees to distinguish text written by big babble machines versus humans. Humans also have a certain pattern of writing, a given distribution of how some words are more likely to appear than others. How can one tell them really apart?

As an indicator, yeah, might be usable. But I wouldn’t read too much into it before seeing results of a study that runs actual tests.

with enough text, you will see that certain words are used more often

Which is also a thing humans do.

do to identify AI text — it has lots of tells anyway

I see what you did there.