[0] https://gregoreite.com/drilling-down-details-on-the-ai-train...
[1] https://insightcivic.s3.us-east-1.amazonaws.com/language-mod...
It's not that there aren't normal people on there, it's that organic feel-good posts have a harder time gaining traction vs the sea of flamebait and sponsored ads and astroturfed spam. The signal to noise ratio was very, very low by the time I left. I don't know how it is these days...
It's just such a toxic, manipulated environment... I don't use social media anymore, just text some friends, and am much happier.
And there's extremely strong moderation here that does the opposite of Facebook: It optimizes against controversy and vitriol rather than encouraging it.
We end up with a bunch of nerds mostly talking shop and sometimes complaining about the job market, but that's still far less ragebaitey than most social media.
Much of what's bad about Facebook and other social media pretty much boils down to their need to show as many ads to as many eyeballs as possible.
Something doesn't add up here. Nerds talking shop and controversy and vitriol about the their technical preferences is the same thing.
Perhaps what you're saying is that controversy and vitriol is only apparent when you're an "innocent bystander" who doesn't have a passion for the subject? Which HN avoids by usually remaining focused on a fairly narrow set of subjects, to a user base generally passionate towards those subjects, so you don't notice?
That is an astute observation, if that is what you're trying to say. Indeed, if you showed your non-technical grandmother HN, she would no doubt see it the way Facebook was talked about earlier.
I don't think – I hope, at least – no one is going to start a shooting war over their framework of choice. You can't say the same about much of the content circulating around social media.
(Edit: You added more to your post after I replied. To your point of "Which HN avoids by usually remaining focused on a fairly narrow set of subjects", that's not just my observation, that's the actual guidelines: https://news.ycombinator.com/newsguidelines.html. We self-select into a narrow slice of nerdtalk or we end up getting downvoted or banned from the site. To me that is the big difference between an interest-based forum that generally stays a functional monoculture vs a general social media site that brings diverse strangers together into shouting wars about whatever the controversy du jour is.)
Is there? Perhaps the trouble here is that your examples are too far apart to recognize how they compare?
What if Facebook, instead, said "Fat workers are too bloated, you should hire skinny workers"? Or if HN said "so and so projects are all ____ and they should all be ____, and oh, your vacuum doesn't suck".
I see no practical difference. It seems the only difference is that Facebook tends to talk about people, HN about tech. But that's not a significant distinction – aside from where your interests lie. Certainly tech-minded folks often find people to be uninteresting.
Oh, and my vacuum does suck just fine, thank you very much.
Why do you say that? I share in the tech-minded proclivity towards not having much interest in people, so I admit to being largely out of the loop, but what is there to be worked up about where you wouldn't equally get worked up around some tech-based topic?
It just sounds boring an uninteresting to me. When I have encountered discussions about those topics, all I have is some laughter at how silly the people sound. Just like I'm sure how the "Next.js is awful" conversations sound to the metaphorical grandma.
But out in the greater society, people get all worked up about those other topics, and will go protest in the streets, blockade representatives' houses, send death threats, rage endlessly online, attempt assassinations, go on shooting sprees, etc. Those things are usually triggered by some emotional attachment to some tribal identity and anger towards some out-group.
In a monoculture like this, we're mostly peers. Yeah, we might disagree about some particular things, but there's less of an "us vs them" mindset as you see politics, geopolitics, values-based communities, etc.
Discussions of technical merit, even when heated, thankfully don't reach those emotional heights quite as often. In my experience, at least...
I understand that. Believe me, I can definitely get worked up about Next.js, if you really want to go there...! It is in no way surprising that people interested in people can end up in the same place around topics related to people.
> In my experience, at least...
But the underscore here is that your experience is based on where your interests (perhaps not the perfect word to describe this, but it's what I got) lie. If you have a greater interest in people, you're going to feel more strongly about it. I'm not bothered by it at all.
But, by the same token, what is to say that I don't feel more strongly towards technical controversy than you do?
It doesn't need to, of course. Moderation, outside of blatant spam perhaps, is rather pointless. If someone wants to exclaim "Next.js is the second coming of Jesus", I want to hear it. There is a reason they are saying it.
To not get stuck in my bias: For those interested in people, why wouldn't you want to hear "so and so group of people are all _____ and they should all be ________, and oh, your mom sucks"? It is a ridiculous statement at face value, admittedly, but it is someone's attempt at communicating. What they are actually trying to say might teach you something important. Wouldn't you want to surface what they are truly trying to get at?
I certainly would for its tech-based counterpart. Especially if it seems to contradict what I hold dear.
> But I hope your experience on HN is still a net positive.
Naturally! I'm out of here the second it is in any way not positive. No sense wasting your free time on something you don't enjoy.
The difference is that talking about a framework with passion doesn't generally end up with severe real world consequences for other people.
The difference between tech and people is that people are... people. They have lives, feelings, all those fun human things. Passionately talking about how X group of people should die (or <insert rhetoric of your choice>) is quite different than passionately talking about how Y framework is the worst thing programmers have ever concocted. Generally the latter won't end up with someone getting stabbed. (Generally...)
> What if Facebook, instead, said "Fat workers are too bloated, you should hire skinny workers"? Or if HN said "so and so projects are all ____ and they should all be ____, and oh, your vacuum doesn't suck".
I think this is a terribly unfair switcharoo to make. Let's fill in the blanks for that second quote.
> "so and so group of people are all inhuman and they should all be killed, and oh, your mom sucks"
> "so and so projects are all terrible and they should all be deleted, and oh, your vacuum doesn't suck"
A technology may not be literally stabbed, but I think it is fair to say that it very well may get stabbed metaphorically. It may even lead to the complete demise of the technology.
Passionately calling for the death of Next.js isn't really any different than calling for the death of a person, other than you may perceive it as being different if you hold people more dear than technology. But, in that case, that's just your arbitrary opinion that means nothing to anything other than you. There are no doubt some who would be more concerned about the loss of Next.js (bad example; nobody would miss Next.js, but you get the idea) than the loss of a person, which is equally valid. The true arbiter of truth, the universe, certainly doesn't have greater feelings towards one than the other.
I love these HN takes.
Well, you certainly are.
Admittedly, when you consider people as a technology, one does start to wonder what it would take to fix the glaring bugs. But then that questions if the technology provides enough value to bother with at all.
> And there's extremely strong moderation here that does the opposite of Facebook: It optimizes against controversy and vitriol rather than encouraging it.
I used to have an account for marketplace and our local neighborhood group and that's my experience.
Most users also don't post much.
Most Facebook posts (as opposed to users) are nonsense and memes.
But also, ML model training is a bit like the stock market: the noise/stupidity in individual examples points in a bunch of random directions, and so ends up cancelling out; while the signal all points in the same direction, and so ends up captured in the distilled model. (You might call this the "Anna Karenina principle of Information Distillation": all right answers are the same, while each wrong answer is wrong in a different way.)
The "initial fine-tuning of a 'raw' base model to produce a 'generalized pre-trained' base model" process, is commonly talked about in terms of "alignment" — making the model ethical, making it not swear at you, making it refuse to engage with certain content, etc. And really, that part is all optional, with there existing "non-aligned" or "orthogonalized" models that don't have these steps performed on them or have had them reversed, but which are still useful.
But a large part of this initial fine-tuning process, consists of debiasing the model's default activation, moving it away from making associations with "common misconceptions" and toward making associations with "right answers." And this process is crucial to a model being able to reason intelligently — as these common misconceptions aren't coherent in a chain of reasoning they appear in, and so lead to the chain of reasoning falling apart / being non-productive. This is, in large part, the "secret sauce" that makes a model of a given size "more intelligent" than another model of the same size.
Every base model that anyone actually cares about or uses — "aligned" or not — has had some process of de-biasing like this applied to it; or, at least, has short-cutted this process by training on a training dataset generated by or filtered by a model that has already had this de-biasing applied to it, such that the derived training dataset doesn't contain the "common misconceptions" in the first place.
And when OpenAI and Meta brag about using RLHF, a large part of what they mean, is crowdsourcing recognition of long-tail "common misconceptions" at scale, to allow a much more thorough version of this de-biasing process[2].
---
[1] Of course, if nobody in the training data ever demonstrates the "right answer" knowledge/associations, then the model will never learn that knowledge/associations. But then, given that these training datasets usually represent decent samples of the population, a "right answer" being missing entirely would likely mean that nobody on Earth knows the "right answer" — so we humans wouldn't be able to recognize the model was wrong. The stock market can be wrong too, for the same reason.
[2] Which, perhaps surprisingly, can add up to more than the sum of its parts. The more of this human-labelled "common misconception"-response RLHF data you have, the more you can derive patterns from this bias data. You can distil out negative examples that you can use to prompt a model to filter the training dataset; but more interestingly, you can distil positive examples of the sort of structured chains of reasoning that work best to inherently avoid triggering the bias. Where, if you can overlap the activation-space hyperspheres of many such inversion-of-bias examples, then you get, essentially, the hypersphere within the model's activation space that contains its instrumental rationality. You can then just bias the model toward living in that part of activation space as much as possible — and this shoots its apparent reasoning capacity way up.
"But how can anyone tell?"
I mean hop on chat GPT right now and ask it to come up with any kind of novel joke. I can almost guarantee it will either be some kind of word play or other equally low hanging kind of pun that you would find on Reddit fifty replies deep.
Where's the Mensa member only social media dating platform for a properly erudite and snobbishly arrogant LLM to train on?
- Video games (epic waste of time and moral debasement) gave us GPUs and NVIDIA
- Social media (rage bait-fueled argument machine and cat pic repository) gave us a huge corpus of text on which to train language models. Yes scientific papers and Wikipedia are necessary but probably not sufficient?
- Crypto (fraud, giant waste of resources) gave us a generation of young people who were comfortable building ludicrously oversized GPU clusters, and to some extent funded NVIDIA and TSMC R&D
all of which led to where we are with AI, which is that it is making (in my humble opinion) impressive dents in the problem of solving intelligence -- cue HN commenters telling me that AI is a giant fraud and waste of resources also :)