> And this thread is seemingly full of people claiming AI can read it while simultaneously sharing that AI could not read the actual message, only the decoy as demonstrated in TFA.
That’s 100% on the authors for failing to make the default main “hidden” text and the decoy easily distinguishable. The way this is set up is incredibly confusing.
I would be very willing to pay more! The choice between “you may get a correct answer, or you may get lied to, without a clear way to distinguish between the two” and “you may get a correct answer, or a clear indication that the answer was not found” is pretty clear. One is a much more useful tool than the other. I don’t see any real incentives for companies making LLMs to keep their AI factually unreliable. (Full disclosure: I work for one, but I’m definitely not in the rooms where such decisions would be made.)
Maybe that’s because I work with agentic AI in my day job, but this seems utterly obvious to me: no reasonable person would ever claim that LLMs are better at keeping secrets or enforcing rules than human employees.
This notice is not about comparing humans and LLMs. It seems that the system was designed in the only reasonable way: with a deterministic permissions layer separate from the agent. But that layer failed to work properly.
So the notice is comparing the difference between how the system was supposed to work and how it actually worked in reality. Normal post-mortem stuff.
I like to dunk on Meta as much as the next guy, but I think this makes sense: deterministic verification like this is not, and should never be, the LLM’s job. The tools it has access to should enforce the permissions layer, ensuring that the LLM can never perform actions the user themselves should not be allowed to perform. In this case, the tool failed to do that.
Is it? Both supervised learning and reinforcement learning are ways of training the model, and the difference between them is not that big. I would say that innate means "in the weights", while non-innate means things the model learned during inference, during its "lifetime".
I think this is exactly it, but let me ask another question (which is not rhetorical, I really don't know). Does the fact that one can describe what consciousness is and where it came from in humans help them to detect it in non-human and/or non-biological entities?
It’s obviously not a new model capability. But using this well-known, existing capability to solve this particular issue is only obvious after the fact.
It’s a useful trick to have in one’s toolbox, and I’m grateful to the author for sharing it.
The top comment categorized scraping as abuse ("abuse such as [...] scraping") - that's precisely why some accuse its author of lack of self awareness.
As a childless OMSCS graduate, I also can’t imagine doing it while having kids, because it took basically all of my free time. That said, I met quite a few people in the program who were in situations similar to yours. I have no idea how they managed it, but they somehow did.
> In that case the winning strategy would be to switch hedge funds every 3 years.
When you flip a coin, you can easily get all heads for the first 2-4 flips, but over time it will average out to about 50% heads. It doesn’t follow from this that the winning strategy is to change the coin every 3 flips.
Okay? I specifically responded to your comment that the parent comment implied "if you make a mistake and say sorry you are also a psychopath", which clearly wasn’t the case. I don’t get what your response has to do with that.
I think the point of comparison (whether I agree with it or not) is someone (or something) that is unable to feel remorse saying “I’m sorry” because they recognize that’s what you’re supposed to do in that situation, regardless of their internal feelings. That doesn’t mean everyone who says “sorry” is a psychopath.
That’s 100% on the authors for failing to make the default main “hidden” text and the decoy easily distinguishable. The way this is set up is incredibly confusing.