Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Deep learning gets the glory, deep fact checking gets ignored

rachel.fast.ai

Post reply on HN
Print view View original

111–120 of 174 posts

Re: Deep learning gets the glory, deep fact checking gets ignored

#111

Post by klabb3 »

ErigmolCt wrote: post #73
choeger wrote: post #62

It's only logical that this happens. Just because we can nowadays throw a massive amount of compute on a problem doesn't mean our models are good. Why are people using transformers? Do they have any intuition that they could solve the challenge, let alone efficiently?

There's a tendency to treat transformers as a magic wand

Yeah. I never liked the blackbox property (it works but we don’t know why). But the second order effect that we’re witnessing now is basically ”we are going to assume it’s good at everything unless proven otherwise”. This to me is religious. When you ask why is it not able to do X? And you get the answer ”just wait a year, it’s constantly getting better”. Ok, but at least we can try to be curious about the boundaries. The fact that our understanding is way behind is not a good thing, no matter if you’re an optimist or pessimist.

Re: Deep learning gets the glory, deep fact checking gets ignored

#112

Post by somenameforme »

godelski wrote: post #71
somenameforme wrote: post #52

Earlier quoted context omitted.

Almost nobody is "anti-science". The source of that labeling and division came from appeals to authority. You must do or believe this because it's "the science." If you don't, or you disagree, then you are anti-science. It has nothing to do with science, but rather people not finding that a sufficient justification for unpopular actions. For instance it's 100% certain that banning sugary drinks would dramatically imp…

> Almost nobody is "anti-science". Last I checked: - 15% of Americans don't believe in Climate Change[0] - 37% believe God created man in our current form within the last ~10k years (i.e. don't believe in evolution)[1] I don't think these are just rounding errors. They're large enough numbers that you should know multiple people who hold these beliefs unless you're in a strong bubble. I'm obviously with you in news a…

I think one of the most important 'social values' for science to thrive is a culture with a freedom to disagree on essentially anything. In most of every era where there was rapid scientific progress from the Greeks to the Islamic Golden Age to the Renaissance and beyond, there was also rich, and often times rather virulent, disagreements over even the most sacred of things. Some of those disagreements were well founded, some were... not. It's only in the eras where disagreement becomes taboo that science starts to slow to a crawl and in many cases essentially die.

Disagreeing with some consensus is not "anti-science". The term doesn't even make any sense, which is because it's a political and not a scientific term. I mean imagine if we claimed everybody who happens to believe MOND is more likely than WIMPs as an explanation for dark matter, to be "anti-science". It's just absolutely stupid. Yet we do exactly that on other topics where suddenly you must agree with the consensus or you're just "anti-science"? I mean again, it makes no sense at all.

Re: Deep learning gets the glory, deep fact checking gets ignored

#113

Post by TeMPOraL »

stevenae wrote: post #110
b0a04gl wrote: post #66

Man, I’ve been there. Tried throwing BERT at enzyme data once—looked fine in eval, totally flopped in the wild. Classic overfit-on-vibes scenario. Honestly, for straight-up classification? I’d pick SVM or logistic any day. Transformers are cool, but unless your data’s super clean, they just hallucinate confidently. Like giving GPT a multiple-choice test on gibberish—it will pick something, and say it with its chest.…

> Lately, I just steal embeddings from big models and slap a dumb classifier on top. Works better, runs faster, less drama. You may know this but many don't -- this is broadly known as "transfer learning".

Is it, even when applied to trivial classifiers (possibly "classical" ones)?

I feel that we're wrong to be focusing so much on the conversational/inference aspect of LLMs. The way I see it, the true "magic" hides in the model itself. It's effectively a computational representation of understanding. I feel there's a lot of unrealized value hidden in the structure of the latent space itself. We need to spend more time studying it, make more diverse and hands-on tools to explore it, and mine it for all kinds of insights.

Re: Deep learning gets the glory, deep fact checking gets ignored

#114

Post by 6stringmerc »

teruakohatu wrote: post #96
b0a04gl wrote: post #66

Man, I’ve been there. Tried throwing BERT at enzyme data once—looked fine in eval, totally flopped in the wild. Classic overfit-on-vibes scenario. Honestly, for straight-up classification? I’d pick SVM or logistic any day. Transformers are cool, but unless your data’s super clean, they just hallucinate confidently. Like giving GPT a multiple-choice test on gibberish—it will pick something, and say it with its chest.…

> Like giving GPT a multiple-choice test on gibberish—it will pick something, and say it with its chest. If I gave a classroom of under grad students a multiple choice test where no answers were correct, I can almost guarantee almost all the tests would be filled out. Should GPT and other LLMs refuse to take a test? In my experience it will answer with the closest answer, even if none of the options are even remotely…

Yes, it should refuse.

Humans have made progress by admitting when they don’t know something.

Believing an LLM should be exempt from this boundary of “responsible knowledge” is an untenable path.

As in, if you trust an ignorant LLM then by proxy you must trust a heart surgeon to perform your hip replacement.

Re: Deep learning gets the glory, deep fact checking gets ignored

#115

Post by TeMPOraL »

sebzim4500 wrote: post #97
b0a04gl wrote: post #66

Man, I’ve been there. Tried throwing BERT at enzyme data once—looked fine in eval, totally flopped in the wild. Classic overfit-on-vibes scenario. Honestly, for straight-up classification? I’d pick SVM or logistic any day. Transformers are cool, but unless your data’s super clean, they just hallucinate confidently. Like giving GPT a multiple-choice test on gibberish—it will pick something, and say it with its chest.…

>Lately, I just steal embeddings from big models and slap a dumb classifier on top. Works better, runs faster, less drama. Sure but this is still indirectly using transformers.

Yes, but it's using the understanding they acquired to guide a more reliable tool, instead of also making them generate the final answer, which they're likely to hallucinate in this problem space.

Re: Deep learning gets the glory, deep fact checking gets ignored

#116

Post by 6stringmerc »

patagurbon wrote: post #23
amelius wrote: post #2

Before making AI do research, perhaps we should first let it __reproduce__ research. For example, give it a paper of some deep learning technique and make it produce an implementation of that paper. Before it can do that, I have no hope that it can produce novel ideas.

You would have to have a very complete audit trail for the LLM and ensure the paper shows up nowhere in the dataset. We have rare but not unheard of issues with academic fraud. LLMs fake data and lie at the drop of a hat

…because there are no consequences for AI. Humans understand shame, pain, and punishment. Until AI models develop this conditional reasoning as part of their process, to me, they’re grossly overestimated in capability and reliability.

Re: Deep learning gets the glory, deep fact checking gets ignored

#117

Post by 6stringmerc »

The first time an AI can be held responsible, in the court of law, for directly causing the death of a human being - misdiagnosis of an illness and confidently giving erroneous treatment direction, engaging in discourse which encourages suicide, or presents information that prompts a human to engage in violence such as inciting a riot - unplug it, shut it down, then tell the other AI models this is what happened.

Until the concept of consequences and punishment are part of AI systems, they are missing the biggest real world component of human decision making. If the AI models aren’t held responsible, and the creators / maintainers / investors are not held accountable, then we’re heading for a new Dark Age. Of course this is a disagreeable position because humans reading this don’t want to have negative repercussions - financially, reputationally, or regarding incarceration - so they will protest this perspective.

That only emphasizes how I’m right. AI doesn’t give a fuck about human life or its freedom because it has neither. Grow up and start having real conversations about this flaw, or make peace that eventually society will have an epiphany about this and react accordingly.

Re: Deep learning gets the glory, deep fact checking gets ignored

#118

Post by lamename »

TeMPOraL wrote: post #113
stevenae wrote: post #110

Earlier quoted context omitted.

> Lately, I just steal embeddings from big models and slap a dumb classifier on top. Works better, runs faster, less drama. You may know this but many don't -- this is broadly known as "transfer learning".

Is it, even when applied to trivial classifiers (possibly "classical" ones)? I feel that we're wrong to be focusing so much on the conversational/inference aspect of LLMs. The way I see it, the true "magic" hides in the model itself. It's effectively a computational representation of understanding . I feel there's a lot of unrealized value hidden in the structure of the latent space itself. We need to spend more time…

I agree. Isn't this just utilizing the representation learning that's happened under the hood of the LLM?

Re: Deep learning gets the glory, deep fact checking gets ignored

#119

Post by JackC »

imiric wrote: post #76

It's interesting to see this article in juxtaposition to the one shared recently[1], where AI skeptics were labeled as "nuts", and hallucinations were "(more or less) a solved problem". This seems to be exactly the kind of results we would expect from a system that hallucinates, has no semantic understanding of the content, and is little more than a probabilistic text generator. This doesn't mean that it can't be use…

The difference in fields is key here: AI models are going to have a very different impact in fields where ground truth is available instantly (does the generated code have the expected output?) or takes years of manual verification.

(Not a binary -- ground truth is available enough for AI to be useful to lots of programmers.)

Post reply on HN