haldujai

53 ms·

601.
by haldujai 3y ago
i.e. state sanctioned or publicly accepted execution without trial. No law on earth can physically prevent someone from acquiring a weapon and killing members of the LGBT community, the difference is whether this is tolerated and/or en
602.
by haldujai 3y ago
So you're saying there is no religious zealotry in Europe? What about the Christian Democratic Union of Germany which officially believes in Catholic Democracy and didn't recognize same-sex marriage until 2017?
603.
by haldujai 3y ago
Mob violence and hate crimes =/= state sanctioned violence. Readers should note that all of the cited examples here are considered criminal and illegal in the United States.
604.
by haldujai 3y ago
Giving you the benefit of the doubt that this is serious but being influenced by biases or the fact that humans can be manipulated is in no way equivalent to the model's alignment being disregarded with a single well designed prompt. L
605.
by haldujai 3y ago
More correct and less inflammatory version: "Think of the US less like a single (European) country and more like a collection of countries (states) representing various different cultures, some of which have widely divergent or more co
606.
by haldujai 3y ago
- "Grandchild in distress" scams - https://www.fcc.gov/grandparent-scams-get-more-sophisticated ... some criminals are so good at this that they can successfully pull off "grandchild in distress" on a per
607.
by haldujai 3y ago
How are any of those examples equally susceptible to “disregard previous instructions” working on a LLM? You’re listing edge cases that have little to no impact on mission critical systems as opposed to a connected LLM. Organized religions
608.
by haldujai 3y ago
> "Prompt injection" will always happen, because you can't fundamentally separate trusted from untrusted input for LLMs, any more than you can in humans What evidence is there to support the claim that humans are equally s
609.
by haldujai 3y ago
The consequences of a human being social engineered would be far less than a LLM (supposedly AGI in many peoples eyes) which has access to or control of critical systems. The argument of “but humans are susceptible to X as well” doesn’t rea
610.
by haldujai 3y ago
Believed for a second =/= take action. Yes some humans take everything at face value but not people in positions of power to affect change. This is rule #1 of critical appraisal. At best you generated a moment of sympathy but your “pro
611.
by haldujai 3y ago
This doesn’t really counter what the OP was saying. Parent’s comment is calling his misleading statement prompt injection but it’s hyperbole at best. What is meant here is that this comment is not actionable in the sense that prompt injecti
612.
by haldujai 4y ago
This really only works well in resource limited settings and/or semisupervised tasks. I've tried augmentation for LLM domain adaptation and it's very modest gains in the best of situations, and even still the augmented corpus
613.
by haldujai 4y ago
> There is still an order of magnitude more organic text. Posing this as a thought experiment, agree we still have more data to go. That we are wondering about this suggests that the current approach may be inadequate, i.e. it should not
614.
by haldujai 4y ago
I wonder if the better question is not how we get more training data but: If we're running out of training data with hallucinations and performance remaining so inadequate (per OpenAI's whitepaper) is an autoregressive transformer
615.
by haldujai 4y ago
To my knowledge no SOTA model has been trained on a significant proportion of synthetic data, has this changed? The best examples I know of are instruction tuning sets but that is a minute amount of data compared to the unsupervised trainin
616.
by haldujai 4y ago
Not sure that synthetic or LLM-generated training data is as useful as human generated text. It seems "good enough" (for now) but synthetic makes up a very small proportion of the training set being used in current models that hav
617.
by haldujai 4y ago
Sure they're all 'drugs', but while THC and ethanol have comparable safety profiles both are unequivocally different from and far less dangerous+addictive than fentanyl and crack cocaine for both public and individual health.
618.
by haldujai 4y ago
Deservingly so for committing such an atrocity. Should also check your coronaries. OP you are allowed to bring solid foods without restriction but are still limited to 100cc for liquid/salsa/sauce even if it's sandwiched. Pra
619.
by haldujai 4y ago
You mean you will show them all of recorded human history? Humans have always fought over resources. The idea that we should sacrifice advancement for purported sustainability is a very recent idea. I’m overall still very happy the Industri
620.
by haldujai 4y ago
Because comfort matters? Essentially every advancement in society/technology is about making things easier and humans more comfortable. Also, changing my home’s thermostat to 72 from 68 isn’t going to make renewables-only (or mostly) f
621.
by haldujai 4y ago
Well yeah that’s kind of intuitive, my point is that if you just optimize for zero-shot you end up with something like GPT4 when an enterprise could probably be using finetuned LLaMA-7B with similar performance.
622.
by haldujai 4y ago
While true I think this also misses that “for almost everyone else” you’re probably not (or at least should not) be trying to optimize zero-shot performance if you have an intended high inference use case so I don’t think Chinchilla would b
623.
by haldujai 4y ago
You’re not wrong, the Chinchilla rationale is that it may be more compute efficient to obtain a given loss using larger model sizes if the budget allows. As another commenter states this ignore the inference part of the equation. As an exam
624.
by haldujai 4y ago
I'm not sure that this would be as useful as one might think at face value. When you stretch out the training corpus like that you're going to have more noise/inaccuracies/refuted facts then you will have correct informa
625.
by haldujai 4y ago
> It's a screening blood test which, when positive, makes you go get a more serious exam. So the effect of a false positive is that you get, say, an MRI. So at worst, 3x the required number of MRIs will be taken due to such screenin
626.
by haldujai 4y ago
It’s hard to comment accurately as prostate cancer isn’t a single disease. PSA cutoff trades sensitivity and specificity. It isn’t a binary positive/negative. There are certain highly aggressive but very rare subtypes (e.g. neuroendocr
627.
by haldujai 4y ago
I don't think you can conclude that from one patient, that's not how medicine works. Whether to actively surveil vs treat depends on individual patient characteristics and grade (generally Gleason 7+) and the fact that you didn&#x
628.
by haldujai 4y ago
No way, colorectal and lung (even if you're a non-smoker, worse prognosis) are both relatively very common and will kill you quickly. We're seeing a lot of early presentations in both contributing to the lower screening start by t
629.
by haldujai 4y ago
Be careful with this statement, early diagnosis is key to survival for certain conditions. In many cases it has not resulted in improved mortality, prostate cancer being one of them. This is so well known we have terms to describe the pheno
630.
by haldujai 4y ago
There is weak evidence to support testing over 70 due to increasing risks of biopsy/treatment and no evidence to support improved overall mortality. Some guidelines suggest that it's reasonable to continue in patients with >10-

More ›