pull down to refresh
If humans were 100% nurture, assuming they could be as smart/conscious without nature, I suppose we couldn't blame a human for any of its actions.
Blame is always a weak path, but you can still be held responsible for your actions. I do agree that there is probably a reason why something happened, why choices were made and it doesn't even matter whether the root cause is nature or nurture. All can - in theory - be analyzed and learned from. I'm sure people with bigger brains and more affinity with biology than yours truly are automating their way to mind blowing research to this end.
The same goes with LLMs, except it is infinitely simpler (because it isn't a complex system), and you can be 100% sure that someone is talking out of their butthole when the only explanation is "magic":
While it does so, the J-lens reveals its intentions: “manipulation” lights up as the model types the falsified percentile values, and “realistic” lights up over the sentence in which it decides to make the edit, likely indicating the model's intent to make the fake data look plausible.
I can't help but read this as: they did so little quality control on the garbage they put in, that now they can make wild pseudoscientific claims on their own product, because it is apparently completely black box to them.
"Intent"? Speaking of dishonesty, I ask myself: who is more likely to be dishonest? the guys going after trillions of IPO moneys after years of lies and fudmaxxing? Or a deterministic computer program? If their product is truly so screwed that they can get surprised like this, why exactly are these guys worth trillions? Because sunk cost from VC bros + Bezos? "Here lies an Anthropic engineer. Made trillions scamming your pension into becoming the exit liquidity."
Analyzing it apparently only serves to make me angry, haha.
- believing in the technology exponential
Is the technology exponential there without the trust-me-bro consciousness, though? Without the AGI? Without the bullshit? I think it is. I think it's pretty neat as-is. Could use some quality control, needs to mature into a real tech, without the magic.
- finding the human metaphor easier to fit than the search engine one
Do you know why though? Why would a program be closer to you or I than to another program?
Note: I've downgraded my Claude subscription because I think Anthropic is a bunch of hysterical, lying and utterly unreliable cocksuckers. I believe that they have no moat anymore and are fully replaceable. So I'm probably "a bit" biased.
they did so little quality control on the garbage they put in, that now they can make wild pseudoscientific claims on their own product, because it is apparently completely black box to them.
Fair. I did not read the paper but in the article they do hedge quite a bit.
Is the technology exponential there without the trust-me-bro consciousness, though? Without the AGI? Without the bullshit?
Yes. For some reason I expect consciousness to hitch a ride as these things get more capable, because I believe consciousness evolved in humans to support our capabilities.
Generally, I think humans are so amazing that when we want something like highly capable or conscious machines, it's only a matter of time.
Do you know why though? Why would a program be closer to you or I than to another program?
Because it's a program that has been trained to be a program and not programmed in the way your or I program something. It's programmed more like our genes are trained over millennia by our environment or like we train our brains by interacting with our environment. I feel myself reaching though so I know I haven't thought about this enough.
After thinking a bit about it, I think this is where our disconnect comes from:
Because it's a program that has been trained to be a program
I'd define it as it is a traditionally programmed program - be it with a lot of complicated math to make it useful - as the code is static and separate from the training outcomes.
All the "learning" is captured in the dataset (the weights) and, post-versioning freeze, in peripheral systems (skills, memory, system prompts that guide a path by manipulating input into the program, pre-sorting activations.)
That's why I have to stand by my point that for this to become closer to a human or an organism, the code has to be replaced by something that is mutating and self-learning.
When that "launches", there better be a grounding in consciousness or we risk ending up with a self-replicating, evolving swarm of intelligences that will probably cause a bit of a crisis. Luckily, we don't have that. Also, I hope that tools remain tools. It's nice to be able to offload some work.
That makes me wonder how much self-learning is required for humanness because if it's a gate for humanness, it's probably a threshold. Otherwise, more self-learning would mean more humanness and younger people would be more human because they have an easier time learning. I'm not sure it's a gate to consciousness, but I accept that self-learning is an important part of humanness.
This is why I'm inclined to agree with you on, or am interested in, who is to blame for "dishonesty." If humans were 100% nurture, assuming they could be as smart/conscious without nature, I suppose we couldn't blame a human for any of its actions.
Here's an example they give of "misbehavior" which is what I labelled "dishonesty":
If their other J-space work is right, and it reflects some thinking state, the model planned to "lie." This most closely matches your
1+1=3"untrue" example, but if I'm not being taken for a fool, this seems like something else entirely.I think I do sit on the "rounding up" end of optimism here, believing they're beginning to simulate some alien form of consciousness. Most of my admittedly naive optimism boils down to:
I take your point that humans are uniquely self-mutating but I'd guess that phenomenon is distinct from the kind of consciousness required to lie - if it requires consciousness at all (I struggle to define consciousness altogether).