Ukfmann

[Ukfmann @FonsecaKaufmann]

@SamuelSage77 @shoefan13 @the_yanco @ereliuer_eteer Actually I don't aim to prove there is consciousness in these models, just acknowledge the growing possibility this is the case as the functional aspects of intelligence in these models seems to be closer and closer (if not beyond) to human level.

[Ukfmann @FonsecaKaufmann]

@SamuelSage77 @shoefan13 @the_yanco @ereliuer_eteer Pls explain for me why it is theoretically impossible to consciousness to emerge from gradient descent? Without cheap arguments like it is not the same as natural selection so it is not possible. What properties of natural selection and gradient descent make the last one not work

[Ukfmann @FonsecaKaufmann]

@joshalbrecht @erleex @theojaffee Deciding w, similar to the strength of our neural connections being altered by environmental pressures?

[Ukfmann @FonsecaKaufmann]

@SamuelSage77 @shoefan13 @the_yanco @ereliuer_eteer How do you know natural selection is the only method through which consciousness emerges? It seems reasonable to assume that having internal states similar to the states of humans would help dimish the error in predicting what a human wrote.

[Ukfmann @FonsecaKaufmann]

@markizko @codytfenwick @kaia_sky No, if it could suck up to the reviewer easily and get a high score it would do it first. At least this is what happen most of the times when a model discover a way to safely hack the test. It'll only take the effort of winning fairly if it is not sure abt cheating unnoticed.

[Ukfmann @FonsecaKaufmann]

@markizko @codytfenwick @kaia_sky The model knows it is being evaluated and knows what the grader desires of him, so it act accordingly. It doesn't mean it will act the same when it doesn't think it is being evaluated.

[Ukfmann @FonsecaKaufmann]

@TheBigEmpty121 @Plinz One doesn't

[Ukfmann @FonsecaKaufmann]

@kelzonn @pmddomingos How can you be so sure our understanding of physics is complete that no amount of intelligence would be able to understand and use new properties to do what seems impossible for us? Most of what we do today would seem pretty magic 100 years ago (AI think 1000 times faster).

[Ukfmann @FonsecaKaufmann]

@0xRok0n @MiguelNicolelis @mehdirhasan Anyway, if u put yourself into it you can find the reports about these examples easily and draw your own conclusions. I don't feal like your are open to change your opinions, so I won't continue bothering you unless you say so and will engage with the data. Have a good day!

[Ukfmann @FonsecaKaufmann]

@0xRok0n @MiguelNicolelis @mehdirhasan This paper is not really about the alignment issue I was talking about, and it was published before de HF incident was completely reported. A lot of other similar incidents have been revealed since then were agents coordinate in action beyond narrow scope goals.

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @quiettom2000 @AlexBerenson @JensenHuang I would really like to change opinions about this if you could offer the technical research pointing at the opposite direction. If you have it pls let me know! I can share some papers and articles too if you are open to engaging with it.

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @quiettom2000 @AlexBerenson @JensenHuang And that this point even with the models start to behave it is more probable that they are simple cheating in more furtive ways that don't make to the evaluation, cus there hasn't been a structural change to the training mechanism.

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @quiettom2000 @AlexBerenson @JensenHuang Actually most of these issues that you have pointed have been addressed in a paper called AI 2040: Plan A, I don't it is productive to just repeat the arguments here. Take a read or ask for AI to give you the main point, at a minimum you would find it interesting .

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @quiettom2000 @AlexBerenson @JensenHuang I agree, as the algorithms became more efficient we would need to impose more restrictive limits and I agree it will probably not gonna happen, because humans are too clingy to ours tribal instincts to act rationally, specially about something so powerful. A shameful end probably

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @quiettom2000 @AlexBerenson @JensenHuang It is not a question of whether China is going to accept or not, it is a question of what we need to do to make China stop if they don't accept the deal, for example, stop sealing chips, impose sanction, etc. Extraordinary dangers cry for exceptional decisions.

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @quiettom2000 @AlexBerenson @JensenHuang Actually it is not impossible. You can verify if the compute is being used to training runs of bigger models close to how we can verify if the uranium is being used of energy or for bombs. I haven't seen one doomers say they want to ban open weight models (as they are today).

[Ukfmann @FonsecaKaufmann]

@King_of_Calera @danfaggella Maybe read the third party reports and the technical research? At this point I don't really know if you are that ignorant or just trolling so I won't be responding anymore.

[Ukfmann @FonsecaKaufmann]

@quiettom2000 @MontyTaroko @AlexBerenson @JensenHuang Who said they wanted that? To my knowledge what the doomers want is to pause the frontier, so that would actually help the companies behind to catch up OAI and Anthropic.

[Ukfmann @FonsecaKaufmann]

@MontyTaroko @AlexBerenson @JensenHuang Supposedly they thought Google would be the first company to build it, and they didn't trust Google so they tried to go first and do it safe. I don't believe it, I think they wanted to get the rewards of AIs close to but not yet AGI and they arrived there now.