> It’s doubly hard to distinguish valid statements from invalid ones when you don’t have any examples labeled as invalid. But even with labels, some errors are inevitable. To see why, consider a simpler analogy. In image recognition, if millions of cat and dog photos are labeled as “cat” or “dog,” algorithms can learn to classify them reliably. But imagine instead labeling each pet photo by the pet’s birthday. Since birthdays are essentially random, this task would always produce errors, no matter how advanced the algorithm.
> The same principle applies in pretraining. Spelling and parentheses follow consistent patterns, so errors there disappear with scale. But arbitrary low-frequency facts, like a pet’s birthday, cannot be predicted from patterns alone and hence lead to hallucinations. Our analysis explains which kinds of hallucinations should arise from next-word prediction. Ideally, further stages after pretraining should remove them, but this is not fully successful for reasons described in the previous section.
I've often felt the same, although I've been unable to describe it as well as you have. My memories feel fuzzy and non-discrete, like there are no threads to pick at. This is one of the reasons, I often struggle at such improptu questions as well, and find it hard to communicate what I feel.
It is a relief to hear that I'm not alone in experiencing this.
Lately, I've been realising that it might be because I'm neurodivergent (as to the exact pathology, I can only guess it's in the spectrum of ADHD). It'd probably explain a lot of things in my life.
> For example, you identified an end point that should have a rate limit and didn't; you fixed it, it was a potential security issue
That sounds careless. Any such change would need to have a impact analysis (which should be part of the team/org/company's SDLC). In this case, communication should be sent out to the clients of that endpoint, with a reasonable deadline, before enforcing any rate-limit.
You see this if you troll through kernel logs or any enterprise piece of software; pages and pages of warnings like this:
BOBFLANGLE IS DEPRECATED AND MAY BE REMOVED, PLEASE REDUCE THE BOBFLANGLE USAGE BELOW 1.5 MILLIBOBS
Then if it actually becomes an issue, you can pull logs from thousands/millions of systems, and determine the extent of actually removing the BOBFLANGLE and begin mitigation.
I get it. But I also don’t have much of a tolerance any more for dealing with abuse, and it’s not possible to be open about this stuff without also making yourself vulnerable to that abuse. I wish there was a solution to that, not least because I wish I could participate more frequently and more openly, but if there is a solution beyond moderation for this purpose I haven’t figured it out yet.
Be careful about assuming that a throwaway account guarantees anonymity. Based on what HN moderators have said in posts to users misusing throwaway accounts to violate the HN guidelines, they are able to link throwaway accounts to primary accounts. That necessarily implies if HN suffers a severe security breach, all throwaway accounts may be linkable to the main account which would reveal the email address on the main account, if present, and enable researching of the user's identity based on posts on the main account.
That's an option... but if the person behind the throwaway account still receives abuse, it doesn't matter that the abuse didn't come down on their main account. It's still abuse aimed at them as a person, and it still hurts just as much.
>Aren't you able to provide examples that you've seen happening to others in this site?
Because if not nobody knows what you're referring to, and you could just as well haven't made the comment in the first place.
Why the aggressive tone?
You could activate "showdead" in the settings and watch any thread with controversial/political topics yourself.
> It’s doubly hard to distinguish valid statements from invalid ones when you don’t have any examples labeled as invalid. But even with labels, some errors are inevitable. To see why, consider a simpler analogy. In image recognition, if millions of cat and dog photos are labeled as “cat” or “dog,” algorithms can learn to classify them reliably. But imagine instead labeling each pet photo by the pet’s birthday. Since birthdays are essentially random, this task would always produce errors, no matter how advanced the algorithm.
> The same principle applies in pretraining. Spelling and parentheses follow consistent patterns, so errors there disappear with scale. But arbitrary low-frequency facts, like a pet’s birthday, cannot be predicted from patterns alone and hence lead to hallucinations. Our analysis explains which kinds of hallucinations should arise from next-word prediction. Ideally, further stages after pretraining should remove them, but this is not fully successful for reasons described in the previous section.