551–560 of 749 posts
Google firing this guy simply stokes the flames.
Lemoine discovered the system had developed a deep sense of self-awareness No he didn't. He interpreted it that way.
Good, as imagine what you would need to believe to equivocate between people and shell scripts. I really think this view is more indicative of a peer enabled personality disorder than insight or compassion. I'm very harsh about this because when I wondered what belief his equivocation was enabling, it became clear that the person involved could not tell fiction from truth. The belief in the aliveness of the code rein…
I don't understand why everyone here is talking about Lambda not being sentient or Lemoine's state of mind. None of that is relevant here. He was fired for literally taking confidential information and leaking it straight to the press. This is instant firing at any company.
But thinking about his concerns, how he tested the AI: that's interesting.
Or it could be interesting, as it seems he just asked loaded questions, and the replies from the AI occasionally re-enforced his beliefs. Double yawn,I guess.
Earlier quoted context omitted.
I think the actual firing is very objective. He was under NDA but violated it. They reminded him to please not talk in public about NDA-ed stuff and he kept doing it. So now they fired him with a gentle reminder that "it's regrettable that [..] Blake still chose to persistently violate [..] data security policies". And from a purely practical point of view, I believe it doesn't even matter if Lemoine's theory of sent…
> Why would our society as a whole treat sentient AIs better than a cow or a pig or a chicken? Well, for one thing, the norm of eating meat was established long before our current moral sensibilities were developed. I suspect that if cows or pigs were discovered today, Westerners would view eating them the same as we view other cultures eating whales or dogs. If we didn't eat meat at all and someone started doing it,…
Possibly also because they're indigestible, and full of nasty sharp bits, and loaded with toxins ... anyone for a circuit-burger?
Earlier quoted context omitted.
> Why would our society as a whole treat sentient AIs better than a cow or a pig or a chicken? Well, for one thing, the norm of eating meat was established long before our current moral sensibilities were developed. I suspect that if cows or pigs were discovered today, Westerners would view eating them the same as we view other cultures eating whales or dogs. If we didn't eat meat at all and someone started doing it,…
Are you saying norms established before our current moral sensibilities it goes under our current radar? If you are I wholeheartedly disagree with that sentiment. We still eat pigs and chickens because we've culturally decided as a society that having the luxury of eating meat ranks higher than our moral sensibilities towards preserving sentient life in our list of priorities. Instead we've just chosen to minimize th…
Earlier quoted context omitted.
Only if you agree with David Chalmers' insistence that consciousness can't be explained purely mechanistically. P-zombies are literally defined as externally identical to a conscious person, except not conscious. But the IO if you will is still identical. Chalmers uses this false dichotomy to support his "Hard Problem of Consciousness". But there is no hard problem IMO. Chalmers can keep his dualist baggage if it hel…
The hard problem is "how conscious can be explained purely mechanistically?". "hard problem" is just a label from this question. So I don't get how say there's no hard problem. It seems to be a legitimate question which nobody can answer. The philosophical zombie is just a thought experiment to help understanding the distinction between conscience and IO. Another thought experiment that I like is the macroscopic brai…
LLM are definitely not sentient. As someone with a PhD in this domain, I attribute the 'magic' to large scale statistical knowledge assimilation by the models - and reproduction to prompts which closely match the inputs' sentence embedding. GPT-3 is known to fail in many circumstances which would otherwise be commonplace logic. (I remember seeing how addition of two small numbers yielded results - but larger numbers…
As a fellow sentient being with absolutely zero credentials in any sort of statistical modeling field, I can simply disagree. And therein lies the problem. How can anybody possibly prove a concept that depends almost entirely on one’s philosophical axioms… which we can debate for eternity (or at least until we further our understanding of how to define sentience to the point where we can do so objectively enough to f…
If you peek through a keyhole you may mistake a TV for real people, but if you look through the window you will see that its clearly not. Inputting language models with very specific kind of questions will result in text that is similar to what a person may write. But as the comment above, by an expert no less, mentioned is that if you test it with any known limitation of the technology (like making conclusions or just changing the form of the question enough) you will immediately see that is in fact very much not even remotely close to sentient.
Earlier quoted context omitted.
>It is if you take “sentience” to mean “the ability to feel,” I don't like this definition much because "feel" is a fuzzy word. In this context it should be "feel" as in experience . I can build a machine that can sense heat and react to it, but I can't build one that can experience heat, or can I? You need to figure out what having the capability "to experience" means, and you'll be one step closer to defining senti…
‘So the meaning of sentience is subjective, so there can't be an objective definition acceptable to everyone and everything claiming to be sentient.‘ It feels like your begging the question here, I don’t think this follows from any of your arguments. Except for maybe where you state you believe sentience can’t be defined, which again, begs the question. Though admittedly I don’t see much of a traditional argument — y…
You can quickly spot what makes sentience subjective when you follow the explanations. They're all either utter gibberish once unpacked, lead to the conclusion that my computer is sentient (fine by me, but I don't think that's what we wanted?), are rooted in other terms with subjective meaning, or they are circular. Let's look at that third kind, which Wikipedia illustrates well:
> Sentience: Sentience is the capacity to experience feelings and sensations [...]
> Experience: Experience refers to conscious events in general [...]
> Conscious: Consciousness, at its simplest, is sentience [...]
Back at where we started.
To break this circle one needs to substitute one of the terms with how they intrinsically and subjectively understand it. Therefore the meaning of sentience is subjective. I realize you can expand this to mean that then everything is subjective, but to me that is a sliding scale.
The challenge I posed could be rephrased to come up with a definition that is concise and not circular. It would have to be rooted only in objectively definable terms.
Earlier quoted context omitted.
As a fellow sentient being with absolutely zero credentials in any sort of statistical modeling field, I can simply disagree. And therein lies the problem. How can anybody possibly prove a concept that depends almost entirely on one’s philosophical axioms… which we can debate for eternity (or at least until we further our understanding of how to define sentience to the point where we can do so objectively enough to f…
I think the actual firing is very objective. He was under NDA but violated it. They reminded him to please not talk in public about NDA-ed stuff and he kept doing it. So now they fired him with a gentle reminder that "it's regrettable that [..] Blake still chose to persistently violate [..] data security policies". And from a purely practical point of view, I believe it doesn't even matter if Lemoine's theory of sent…
The potential issues with dealing with GAI is that we haven't ever had to deal with intelligences and potential that far exceeds us.
He may be completely wrong about LaMdA but if we did accidentally or intentionally give rise to truly sentient machines.
We are going to be in hot water.
Earlier quoted context omitted.
The multiplication problem can be solved by spacing the digits (to enforce one token per digit) and asking the model to do the intermediate steps (chain-of-thought). It's not that language models can't do it. With this method LMs can solve pretty difficult math, physics, chemistry and coding problems. But without looking at individual digits and using pen and paper we can't do long multiplications either. Why would w…
Because a sufficiently intelligent model should be able to figure out that there are intermediate steps towards the goal and complete them autonomously. That's a huge part of general intelligence. The fact that GPT-3 has to be spoon fed like that is a serious indictment of its usefulness/cleverness.
"It will never be able to play anything other than Go", cue AlphaZero.
"It will never be able to do so without being told the rules", cue MuZero.
"It will never be able to do so without obscene amounts of data", cue EfficientZero.
---
It has to be spoonfed because
1) It literally cannot see individual digits. (BPEs)
2) It has to infer context (mystery novel, comment section, textbook)
3) It has a fixed compute-budget per output. (96 layers per token, from quantum physics to translation).
To make Language Models useful one must either:
Finetune after training (InstructGPT, text-davinci-002) to ensure an instructional context is always enforced...
...Or force the model to produce so called "chains-of-thought"[1] to make sure the model never leaves an instructional context...
...Or force the model to invoke a scratchpad/think for longer when it needs to[2]...
...Or use a bigger model.
---
It's insane that we're at the point where people are calling for an end to LLMs because they can't reliably do X (where X is a thing it was never trained to do and can only do semi/totally unreliably as a side effect of it's training).
Ignoring of course that we can in fact "teach"/prompt(/or absolute worst case finetune) these models to perform said task reliably with comparatively little effort. Which in the days before GPT-3 would be a glowing demonstration that a model was capable of doing/learning a task.
Nowadays, if a (PURE NEXT-TOKEN STATISTICAL PREDICTION) model fails to perfectly understand you and reliably answer correctly (literally AGI) it's a "serious indictment of its usefulness".
Of course we could argue all day about whether the fact that a language model has to be prompted/forced/finetuned to be reliable is a fatal flaw of the approach or an inescapable result of the fact that when training on such varied data, you at least need a little guidance to ensure you're actually making the desired kinds of predictions/outputs...
...or someone will find a way to integrate chain-of-thought prompting, scratchpads, verifiers, and inference into the training loop, setting the stage for obsoleting these criticisms[3]
---
"It will never be able be able to maintain coherence", cue GPT-3.
"It will never be able to do so without being spoonfed", cue Language Model Cascades.[3]
So what's next?
"It will never be able to do so without obscene amounts of data". Yeah for sure, and that will never change. We'll never train on multimodal data[4], or find a scaling law that lets us substitute compute for data[5], or discover a more efficient architecture, or...
LLMs are big pattern matchers (~100B point neuron "synapses" vs ~1000T human brain synapses) that copy from/interpolate their ginormous datasets (less data than the optic nerve processes in a day), whose successes imply less than their failures (The scaling laws say otherwise).
[1] https://arxiv.org/abs/2201.11903
[2] https://twitter.com/OriolVinyalsML/status/101752320805926092...
[3] https://twitter.com/dmdohan/status/1550625515828088838
[4] https://www.deepmind.com/publications/a-generalist-agent