I've been doing a bit more experimenting with #LargeLanguageModels#LLM and truth, and I've got an interesting one.
my experimental design was that I'd start asking about relationships between European monarchs, and then start introducing fictitious monarchs, but I didn't get that far...
@simon_brooke I hope that attempts to replace thinking with LLM will fail quickly. What worries me more is swathes of people believing their nonsense to be true fact. We have the examples of lemming behaviour, Braveheart, oh good grief if I continue this list it will cause blood pressure to explode its monitor.
It's almost as if Bard is reasonably plausible when it comes to 'one degree of separation' questions, but when there is a chain of facts to connect, it totally bombs out.
Which makes sense, since the longer the chain, the more branching possibilities there are, and the less likely it is that anything in the training data addresses that particular question.
(Says me, with no real expertise in this matter at all.)
"Describe the relationship between Sweyn Forkbeard and Tsar Nicholas II"
Bard answered, in part:
"Sweyn Forkbeard's daughter, Gytha, married Vladimir II Monomakh, the Grand Prince of Kiev. One of their grandsons, Inge I of Norway, was the great-grandfather of Margaret I of Denmark. Margaret I was the grandmother of Christian I of Denmark, who was the great-grandfather of Ivan IV of Russia, the grandfather of Tsar Nicholas II"
I expected an #LLM to run into problems when I introduced Rufus of Rubovia, but the real royal families of Europe were massively intermarried and in particular Scandinavian royalty had a major impact on the formation of Russia; and their geneologies are matters of public record you'd expect an #LLM to have ingested. So if there were any semantic model here at all, this is just trivial data retrieval. But Bard can't do even that, so it just makes shit up.
So if not Sweyn Forkbeard, what about his great rival Olaf Tryggvason? Bard tells me:
"Some people may mistakenly believe that there is a relationship between Olaf Tryggvason and Nicholas II because they both have the same surname, "Tryggvason." However, this is just a coincidence. The surname "Tryggvason" is a common Scandinavian surname, and it is not clear whether Olaf Tryggvason and Nicholas II are actually related."
Wait, WHAT!? There are so many ways this is wrong!
1, Tryggvason is not a surname, it's a patronymic;
2, Tryggvason, as a patronymic, is not now common in Norway or anywhere else;
3, Nicholas II's family name was 'Romanov'; that literally means 'the Roman', but it's clearly not a patronymic;
4, Nicholas II did have a patronymic, but it was 'Alexandrovich';
5, 'Romanov' and 'Tryggvason' are not the same;
6, 'Alexandrovich' and 'Tryggvason' are not the same...
We expect #LLM responses to be plausible; this fails even that test.
Let's be clear: I know that #LLM answers are truthful only by accident. But that is exactly why I started with questions whose answers were simple data retrieval: factual things which, while not necessarily trivial for humans to retrieve, should nevertheless be very simple for a machine. It's tree search, dammit!
@Loukas@simon_brooke That’s the thing though. These LLMs don’t have access to the internet. They don’t look up facts. They just give plausible answers. They’re great if you give them some text and ask them to summarise it. But they will just make stuff up if you ask them general knowledge questions.
@Loukas Someone, in a history department in a university somewhere, has a database with a complete geneology of European monarchs, which could answer these questions in milliseconds. I'd be actually surprised if Google does not have one.
I expected it to give answers like the first one (x begat y, who begat z, who begat...), except accurate, or else say that there was no relationship.
I expected nonsense to emerge only when I introduced fictitious names.
I expected at least some answers to be correct and verifiable. But not only do I have 'hallucinations' in both answers, but I have at least seven actual errors in the answers to two questions.
Anyone who fears that #LargeLanguageModels are coming for their job either has a really shit job, or is really shit at their job.
Anyone who fears that #LargeLanguageModels, or anything which may evolve in the near term from #LargeLanguageModels, might take over the world really doesn't understand what this technology is (in)capable of.
All these delusional credulous nerds out there insisting these things are alive and have internal mental models of the world and meanwhile they’re just spitting out blocks of text that are plausibly correlated to other blocks of text.
Contrasting side note you kind of addressed I think. Iirc when an llm bot is hallucinating, it makes logically inconsistent statements G-- is so-and-so's daughter, G-- & X are married. Is G-- so-and-so's daughter? No.
"Small" LLMs have been applied to constructing proof sketches, in an effort to actually say at least internally consistent things. If you watch articles @Jose_A_Alonso shares for a few days the topic will come up again.
@screwtape@fedops@simon_brooke@Jose_A_Alonso
but of course logical consistency between two statements A and B isn't part of the model, so checking it isn't part of the generative algorithm. Only whether statements like A and B are likely to occur together in the training data.
An automated proof sketching engine that also used a language model to reword maths into natural language might be interesting, but natural language itself is not and can't be what proofs are constructed with.
@petealexharris
Thanks! I believe automatically using proof assistants using a Large Language Model approach on a large language model approach is an established research topic, with the problem that there is not a Large amount of Language examples on proof assistants. IIRC.
In this case, I was suggesting it had heuristic value due to some established properties of hallucination. @fedops@simon_brooke@Jose_A_Alonso
@petealexharris
I would have a look at Jose's timeline of academic articles, including this research skein for context. After which I will feel better except for the comparison to string theorists! Aside, I am an opponent of LLMs for other reasons ( supporting gopher://perma.computer ). @simon_brooke
@simon_brooke@otfrom If all the jobs that just involved the churn of meaningless signs were eliminated from the economy, there would be no economy. Over the span of just a few months, GANs and LLMs have already seen for the livelihood of a few people I know – people who had managed to retain a certain independence from the métro-boulot-dodo world, as well.
@otfrom yes, but backending a natural language parser onto tree search is something I was doing in games programming forty-mumble years ago. It ain't rocket science.
@simon_brooke
Those are some wild 'hallucinations'.
I keep seeing adverts on the work twitter for using these to 'increase productivity ' and 'improve your management writing '. It's like management classes just can't wait to dive in to having these write their emails etc. Maybe they will end up out of a job because they get sacked for really bad emails....
Add comment