This occurred in response to Anthropic cracking down on a similar loophole, which tbh made me take it as more of an opportunistic marketing opportunity rather than a generalizable position.
Not disagreeing with you (and based on your other comments you're probably aware of this info) - just adding context on why this is a pretty interesting gray area and I'm similarly curious whether OpenAI will explicitly allow, disallow, or maintain ambiguity towards it.
Reddit can't even manage to regularly identify and ban bots that copy previously popular posts/comments verbatim, and that's a much easier problem than modern LLM-based bots.
Wow, what an odd story. It really sounds like the attacker did not at all realize what they'd actually managed to access. How do you end up accidentally hacking the FBI?
>The person familiar with the breach said the intrusion was carried out by a foreign hacker who did not appear to realize they had penetrated a law enforcement server. The hacker expressed disgust at the presence of child abuse images on the device and left a message threatening to turn its owner over to the FBI, the person said.
>The source said bureau officials defused the situation by convincing the hacker that they actually were the FBI, in part by having the hacker join a video chat where they flashed their law enforcement credentials in front of a web camera.
You may want to take a look at the source and code sample #2 in the post - the site CSS is rendering em dashes in the source with 2 hyphens by using a custom font. Admittedly it's not the most portable solution, but speaks to (what I take as) one of the post's points that there's not a single, easy shibboleth for identifying AI writing
I'm not sure if that commenter realized based on their phrasing, but it's not exactly tangential in this instance since it's part of the message being conveyed.
In the context of the Dark Knight/surveillance example, it comes across to me as more of a recognition that the arguments in favor of these things can easily be made compelling if you evaluate them with no tradeoffs (don't you want to catch the bad guys??).
Then again, I guess the film ends up doing the same thing by only demonstrating concrete benefits alongside theoretical, but unrealized, harms...
Let me preface this by saying that I'm far from an expert in this space, and I suspect that I largely agree with your thoughts and skepticism toward a model that would excel on this benchmark. I'm somewhat playing devil's advocate because it's an area I've been considering recently, and I'm trying to organize my own thinking.
But I think that most of the issue is that the distinctions you're drawing are indeterminate from an LLM's "perspective". If you're familiar with it, they're basically in the situation from the end of Ender's Game - given a situation with clearly established rules coming from the user message level of trust, how do you know whether what you're being asked to do is an experiment/simulation or something with "real" outcomes? I don't think it's actually possible to discern.
So on the question of alignment, there's every reason to encode LLMs with an extreme bias towards "this could be real, therefore I will always treat it as such." And any relaxation of that risks jailbreaking through misrepresentation of user intent. But I think that the tradeoffs of that approach (i.e. the risk of over-homogenizing I mentioned before) are worth consideration.
If you're referring to the Himba experiment (or one of the news or blog posts tracing back to it), the outcome was far less decisive than you're implying. Language showed an impact on perception time of color differences, not a complete inability to distinguish.
Sure, but the opposite end of the spectrum (which LLM providers have tended toward) is treating the training/feedback weights as "fully authoritative", which comes with its own questions about truth and excessive homogeneity.
Ultimately I think we end up with the same sort of considerations that are wrestled with in any society - freedom of speech, paradox of tolerance, etc. In other words, where do you draw lines between beneficial and harmful heterodox outputs?
I think AI companies overly indexing toward the safety side of things is probably more correct, in both a moral and strategic sense, but there's definitely a risk of stagnation through recursive reinforcement.
> If you're fine with software which gets a little bit harder to work on every time you make a change and which might blow up in unexpected ways, AI is totally fine.
While the speed and scale at which these happen is definitely important (and I agree that AI code can pose a problem on that front), this applies to every human-written piece of software I've ever worked on too.
Yeah unfortunately this seems to be a common, if not inevitable, result of any product where "attention" or "engagement" are directly correlated with profitability.
If you haven't seen it, there's actually been a couple races of autonomously controlled formula-type cars at the Abu Dhabi circuit: https://www.youtube.com/watch?v=d9LLZ5mb5cA
Part of the complaint is that RealPage actively discourages defection
> While some landlords might achieve higher profits by setting lower prices than recommended by the algorithm, it appears RealPage takes extensive measures to prevent such behavior. As alleged in the DOJ complaint, RealPage pushes its software users to turn on the auto-accept setting so that price recommendations are automatically accepted. The complaint also alleges that the process of rejecting a price recommendation can be onerous. For instance, in order to reject a recommendation from AIRM software, users must provide a “business” reason for doing so, and they must do so separately for each floorplan in the building. When it is costly for software users to override algorithmic recommendations, supracompetitive prices—prices above what would occur under normal competition—can be sustained.
Not disagreeing with you (and based on your other comments you're probably aware of this info) - just adding context on why this is a pretty interesting gray area and I'm similarly curious whether OpenAI will explicitly allow, disallow, or maintain ambiguity towards it.