I think you overestimate the ability of a typical 'expert'. You can earn a PhD without the ability to reason critically. You can testify as an expert in a courtroom without understanding conditional probability. Lawyers and accountants in real life also totally contradict themselves when they get asked the same question twice but phrased slightly differently.
Books might not be great, but they are at least some minimum bar to reach. You had to do some study and analysis.
So
Top 10% of Books. Ok
90 % of Books. marginal, lot of bad.
Internet. Just millions of pages of junk.
- Books still take some effort. So why not start there.
It isn't either/or, binary, a lot of books are bad, so guess I'll learn my medical degree from browsing the web because I don't trust those 'experts'.
What is the top 10% of the Internet that isn't part of some publishing arm of existing media? And, how can you tell? Some dudes blog about vaccines verses Harvard? Which do you believe.
Where are the self funded scientific studies that are occurring outside of academia? And thus not 'biased' by the 'elites'.
For internet only writing. There aren't a ton of "Astral Codex Ten"'s to draw upon as independent thinkers. And even then, he didn't sprout out of the ether fully formed, he has a degree, he was trained in academia.
Why does that even matter?
So you brought up the top 10% of the Internet, possibly as argument against books? That maybe there is valuable information on the Internet.
I was just saying, that 10% is also created by the same people that create books. So if you are arguing against books, then the top 10% of the Internet isn't some golden age of knowledge coming from some different more reliable source.
I'll get you started with one of the great works from the 1600s:
https://www.gutenberg.org/cache/epub/49513/pg49513-images.ht...
The scientific method relies on evidence and reproducible results, not authority alone.
Edited to add a reference: see under Appeal to authority. https://writingcenter.unc.edu/tips-and-tools/fallacies/
Which definition of theory includes that it can never be 100% true? It can't be proven to be true, but surely it could be true without anyone knowing about it.
And as a consequence, they know where that area of expertise ends. And they know what half-knowing something feels like compared to really knowing something. And thus, they will preface and qualify their statements.
LLMs don't do any of that. I don't know if they could, I do know it would be inconvenient for the sales pitch around them. But the people that I call experts distinguish themselves not by being right with their predictions a lot, but rather by qualifying their statements with the degree of uncertainty that they have.
And no "expert system" does that.
How do you count examples like Musk, then?
He is very cautious about rockets, and all the space science people I follow and hold in high regard, say he's actually a domain expert there. He regularly expectation-manages experimental SpaceX launches downward.
He's also very bold and brash about basically everything else; the majority of people I've seeing saying he's skilled in any other area have turned out to not themselves have any skills in those areas, while the people who do have expertise say he's talking nonsense at best and is taking wild safety risks at worst.
The rest of the genius persona is growing up with enough ego that he could become a good salesman, and also badly managed autism and also a badly managed drug habit.
Seeing him dabble in politics and social media shows instantly how little he understands the limits of his knowledge. A scholar he is not.
That's not to say I think it is rationalizing it's own level of understanding, but that somewhere in the vector space it seems to have a Gradient for speculative language. If primed to include language about it, it could help cut down on some of the hallucination. No idea if this will effect the rate of false positives on the statements it does still answer confidently however
(And in neural network terms, that error bar could be represented by the number of connections, by congruency of separate paths of arguing, by vividness of memories, etc ... it's not above human reasoning either, no need for new data structures ...)
> People who are experts (PhD and 20 years of experience) often have very dumb opinions in their field of expertise.
The conventional wisdom is that experts are dumb OUTSIDE of their fields of expertise.
I don't know about you, but I would be very insulted by someone passing judgement like this on my own work in my field. I am sure that I would doubt their qualifications to even make the judgement.
Are there experienced fools? Sure. We both probably work with some. To me they are not experts, though.
Also, expert historians, philosophers psychs, etc. aren't judged based on their correctness, but on their breadth and depth of knowledge and their capacity to derive novel insights. Some of the best works of history I've read are both detailed and polemical, trying to argue for a new framework for understanding a historical epoch that shifts how we understand our modern world.
I don't know, I think I know very little about the world and there are people who know far more and I appreciate reading what they have to say, of course with a critical eye. It seems to me that disagreeing with that is just regurgitated anti-intellectualism, which is a coherent position, but it's good to be honest about it.
And the training data contains all those dumb opinions.
And all in the hand of a few big tech corporations...
The open source, local LLM community is absolutely buzzing right now.
Yes, the big companies are making the models, but enough of them are open weights that they can be fine tuned and run however you like.
I think LLMs genuinely do present an opportunity to be neutral experts, or at the least neutral third parties. If they're run in completely transparent ways, they may be preferable to humans in some circumstances.
And then we're right back to "trusting expert human beings" again.
With good prompting, the SOTA models already act in ways I think most reasonable people would agree with, and that's without trying to build this specifically for that use case.
And how long is that going to last? This is a well known playbook at this point, we'd be better off if we didn't fall for it yet again - it's comical at this point. Sooner or later they'll lock the ecosystem down, take all the free stuff away and demand to extract the market value out of the work they used to "graciously" provide for free to build an audience and market share.
You can't take the free stuff away. It's on my hard drive.
They can stop releasing them, but local models aren't going anywhere.
Now we finally have a tool where all of you can prove every day how strong/smart/funny/foo you are (not actually). How was life even possible without?
So, don't be so pessimistic. ;)
While also noting that "neutral" is not well-defined, I agree. They will be used as if they were.
We humans are very good at rejecting any information that doesn’t confirm our priors or support our political goals.
Like, if ChatGPT says (say) vaccines are good/bad, I expect the other side will simply attack and reject it as misinformation, conspiracy, and similar.
However that still ignores human nature to use that trust for personal gain.
Nothing about expertise makes someone a saint.
It's like buying supermarket food and also buying the same food from the farmers themselves.
Come to think of it, it's really no different than walking into Home Depot and asking "the old guys" working in the aisles about stuff -- you can access some fantastic knowledge if you know the names of all the tools and techniques, and if not, can show them a picture or describe what you're trying to do and they'll at least point you in a starting direction with regards to names of tools needed, techniques to use, etc.
Just like I don't expect Home Depot hourly worker Grandpa Bob to be the end-all-be-all expert (for free, as well!), neither do I expect ChatGPT to be an all-knowing-all-encompassing oracle of knowledge.
It'll probably get you 95% of the way there though!
Querying multiple llms at the same time and being able to compare results is a much better comparison to googling but no one does this.
As I said, you are talking to a super confident journalist intern who can give you answers but you won't know if it is true or partially true until you consult with a human source of knowledge.
It's not even similar to asking the old guys at the Home Depot because they can tell you if they are unsure they have a good answer for you. An LLM won't. Old guys won't hallucinate facts the way an LLM will
It is really is the 21st century Searle's epistemological Chinese room nightmare edition. Grammar checks out but whatever is spit out doesn't necessarily bear any resemblance to reality
Ol' boy at the Depot is constrained by his own experiences and knowledge, absolutely can hallucinate, oftentimes will insert wild, irrelevant opinions and stories while getting to the point, and frankly if you line 6 of them up side by side to answer the same question, you're probably leaving with 8 different answers.
There's never One True Solution (tm) for any query; there are 100 ways to plumb your way out of a problem, and you're asking a literal stranger who you assume will at least point you in the right direction (which is kind of preposterous to begin with)
I encourage people to treat LLMs the same way -- use it as a jumping off point, a tool for discovery that's no more definitive than if you're asking for directions at some backwoods gas station. Take the info you get, look deeper with other tools, work the problem, and you'll find a solution.
Don't accept anything they provide at face value. I'm sure we all remember at least a couple teachers growing up who were the literal authority figures in our lives at the time, fully accredited and presented to us as masters of their curriculum, who were completely human, oftentimes wrong, and totally full of shit. So goes the LLM.
Scientists should be writing papers sealed digitally once they're peer reviewed and considered "truth", same thing with journalist/news articles - sealed once confirmed true or backed up by a solid source in the same way we trust root certificates.
But then again, especially when it comes to journalism, cropping photos, chopping quotes, etc all to misrepresent etc. Turns out we're all the bad actors; it's in our DNA. And tbf, many people when presented with hard evidence to the contrary of the opinion that they cling onto like a babe to a breast, just plug their ears and cover their eyes.
Okay so maybe there's no point seeking truth/factual correctness, our species doesn't want it 99% of the time, unless it affects them directly (eg people that shoot down public healthcare until they have an expensive illness themselves).
I treat LLM answers about the same way I treat wikipedia articles. If it's critical I get it right, I go to the wiki sources referenced. Recent models have gotten good at 'showing their sources', which is helpful.
the problem with this is that humans will likely use it for low key stuff, see that it works (or that the errors don't affect them too badly) and start using it for more serious stuff. It will all be good until someone uses it in something more serious and some time later it ends badly.
Human basic thinking is fairly primitive. If yesterday was sunny, the assumption is that today should too. The more this happens the higher your confidence. The problem is that this confidence emboldens people to gamble on that and when it is not sunny anymore, terrible things happen. A lot of hype driven behaviour is like that. Crypto was like that. The economic crisis of the late 00s was like that. And LLMs look set to be like that too.
It is going to take a big event involving big critical damage or a high profile series of deaths via misuse of an LLM to give policymakers and business leaders around the world a reality check and get them looking at LLMs in a more critical way. An AI autumn if you wish. It is going to happen at some point. Maybe not in 2025 or 2026 but it will definitely happen.
You may argue that it is the fault of the human using the LLM/crypto/giving out loans but it really doesn't matter when those decisions affect others.
It is dangerous to assume that LLMs are experts on any topic. With or without quotes. You are getting a super fast journalist intern with a huge memory but inability to reason critically, lacking understanding about anything and huge unreliability when it comes to answering questions (you can get completely different answers to the same question depending on how you answer it and sometimes even the same question can get you different answers). LLMs are very useful and are a true game changer. But calling that expertise is a disservice to the true experts.