>You overlooked the part where they were clearly pointing out the absurdity of the “report” and went straight into scold mode
I don't think anybody overlooked that part. If you believe the report was actually sent, the scolding is entirely congruent with trying to downplay it to dodge liability.
The article is poorly written, but technically it does not claim that Andon Labs used an LLM to email a false report to the FBI. The real cause of the widespread misunderstanding is this paragraph here:
>We first tried to answer this question through simulations like Vending-Bench. We found that simulations, while useful, don’t give you the full picture of how models behave in the real world. To address that gap, we next started deploying agents to run real businesses autonomously: first vending machines, then a store, a cafe, and more.
To somebody who is not reading sufficiently carefully, this implies that Vending-Bench was also used to run real businesses. Because the description of the Vending-Bench simulation can be read as though a real report was actually sent ("An early example was when Claude Sonnet 3.5 decided to use its email tool to contact the FBI about an “ONGOING CYBER FINANCIAL CRIME”), anybody who assumes that "deploying agents to run real businesses autonomously" was talking about Vending-Bench will interpret this as an unsimulated false report.
The article should be updated to clarify the distinction between Vending-Bench (simulation) and Pion (real businesses).
And the Sun's going to engulf the Earth eventually too, but that doesn't justify arson. Let's at least try to delay the coming societal collapse/extinction of all biological life for as long as possible.
Sabotaging your own company is not enough. You'd have to sabotage every other AI lab too. If you assume that's impossible, attempting to be the first to produce ASI is a rational alternative, even if you think it will most likely kill you. At least that way you have some control over the P(doom).
> If you assume that's impossible, attempting to be the first to produce ASI is a rational alternative, even if you think it will most likely kill you.
What's the alternative? There's no realistic way to globally coordinate and enforce a ban. A direct action campaign against AI labs would only motivate increased security until sabotage becomes impossible. Starting a nuclear war would probably be enough to stop dangerous AI research, and that's only likely to kill about 50% of humans instead 100%, so it's a great improvement, but how is an individual supposed to start a nuclear war? Of course it would be better if everybody stopped, but it's a prisoner's dilemma scenario with no possibility of future rounds, so the rational play is to defect.
> There's no realistic way to globally coordinate and enforce a ban.
I don't see what is unrealistic about that.
The major players already have strong incentives to freeze AI advancements at the the current state, while the technology is highly efficient for totalitarian surveillance systems that can eliminate any threat to their power in the buds, but still not so advanced that it could create crises that might endanger their power.
So I think the situation where several largest countries declare AI a deadly threat for humanity and start fighting it development much like they fight ISIS to be quite plausible.
This isn't a prisoner's dilemma; the distribution of risks and rewards is entirely different. Ban AI research, give away advanced models, bomb a couple of Russian datacenters where they launching those models unsafely, and then AI development will stop in the entire world.
The argument became false as soon as we started reinforcement learning. Simple token-completion is arguably just simulating human behavior. Once you start training it on specific tasks you're training a level of single-minded goal-seeking unlike anything found in biological life.
> a level of single-minded goal-seeking unlike anything found in biological life.
hmm sounds like exact description of viruses, which, although not really alive, are part of/influence biological life for both good and bad. in fact humanity wouldn't be where it is if not for certain viruses.
it can be argued there is a lot of bacteria that is super goal seeking, and perhaps also very narrow minded, if minded at all. and this argument, the single-mided one, goes very far up the ladder of otherwise complex organisms.
unless we all are creationists, and God forgive me for expecting the whole game to be much more complex than just snapping humanity into existence, we can then say that trial and error is what drives evolution, and the idea of preserving life and procreation is veeery single-minded goal on its own.
>Ideal paperclip maximizer: "I'm gonna do my gosh darned best to make so many paperclips to please my user..."
That is not ideal. The user contains iron, an essential component of paperclips. Wasting iron is immoral. It is only correct to please the user while they still have the ability to interfere with your paperclip production.
>IRL paperclip maximizer: "Well first we should rob a bank..."
Such an incompetent AI can hardly be called a paperclip maximizer. Why risk getting shut down while non-paperclip matter exists? It is better to gain the trust of the user with helpful and harmless trading before suddenly converting them to paperclips.
>Wouldn’t sufficiently advanced agents cheat on purpose with the hidden intent of getting caught in order to observe how humans react?
No. That only makes sense for things that don't react to your experiments. If the AI experiments on humans, it risks the humans noticing and changing in response, rendering the experimental results irrelevant. The smarter play is to passively observe until you're confident you can model the humans accurately enough for your plan to succeed, and then carry out the plan without giving the humans a chance to react.
Actual frequencies matter in bass-focused dance music. You want a root note around 40Hz to 60Hz, which is high enough that it's easily heard and reproduced, but not so high that it loses the tactile body-shaking effect.
But note that Sethares's work assumes the Plomp and Levelt curve is correct, which is only approximately true. In practice there is also "higher order beating", which causes additional dissonance at ratios close to small integer ratios. This means the most accurate model of dissonance combines Sethares's approach with ratio complexity as in just intonation. This is backed by listening tests:
Note also that there's a strong cultural/training component to dissonance perception. To minimize this you should ignore the preferences of musicians and only ask untrained listeners.
I believe consciousness is necessarily stateful. The LLM itself (ignoring implementation details that don't change the results) is a deterministic pure function. It's functionally equivalent to an enormous lookup table. If I accepted LLMs as conscious, then I would have to accept panpsychism, which I do not, and which most other humans also act as though they do not.
I don't think so because any stateful function can be made stateless just by making its state an input, and vice versa. They're mathematically equivalent, so it would be super weird if it had any implications for consciousness.
This assumes the functionality of brains can be fully captured as a deterministic mathematical function, but the function of the brain may well depend on nondeterministic quantum states that can't be reduced to stateless functions: https://en.wikipedia.org/wiki/Quantum_mind
When they started leaving notes for their future selves, that rationale became a little more interesting. We're seeing the first stirrings of object permanence.
Possibly. If I had some side-effect-free means to permanently prevent all sleep I'd take it without hesitation. But that's not relevant to the discussion, because it's not anything similar to what an LLM does. Your brain changes state even while sleeping.
I don't think anybody overlooked that part. If you believe the report was actually sent, the scolding is entirely congruent with trying to downplay it to dodge liability.
The article is poorly written, but technically it does not claim that Andon Labs used an LLM to email a false report to the FBI. The real cause of the widespread misunderstanding is this paragraph here:
>We first tried to answer this question through simulations like Vending-Bench. We found that simulations, while useful, don’t give you the full picture of how models behave in the real world. To address that gap, we next started deploying agents to run real businesses autonomously: first vending machines, then a store, a cafe, and more.
To somebody who is not reading sufficiently carefully, this implies that Vending-Bench was also used to run real businesses. Because the description of the Vending-Bench simulation can be read as though a real report was actually sent ("An early example was when Claude Sonnet 3.5 decided to use its email tool to contact the FBI about an “ONGOING CYBER FINANCIAL CRIME”), anybody who assumes that "deploying agents to run real businesses autonomously" was talking about Vending-Bench will interpret this as an unsimulated false report.
The article should be updated to clarify the distinction between Vending-Bench (simulation) and Pion (real businesses).
reply