Hacker News Reader

Comment by ris | Hacker News Reader

Still waiting for someone to train an LLM entirely from sources written before a chosen date and be able to discuss concepts with someone apparently lacking any knowledge of the world after that date.

View on HN

Try to get an LLM to admit it doesn't know something, first

They're pretty apologetic about it. Then they tell you a different wrong thing.

That takes you pointing out it's wrong. Ask it something it can't know, and it will answer it.

This is where LLMs are worse than even incompetent people. People at least know when they don't know something, and might even be truthful about it. LLMs don't know they don't know.

https://en.m.wikipedia.org/wiki/Dunning–Kruger_effect

Would be fascinating trying to get an LLM trained with 1900 data to discover Einstein physics

Wouldn't be too difficult. Poincaré/Lorentz/Hilbert were close to developing the same concepts behind special relativity around 1905 as well. If Einstein randomly died in 1904, I think relativity would have been discovered within the decade anyways, just by combining their knowledge.

Lorentz developed the Lorentz contraction independently of Einstein already, he was just hampered by the fact he adhered to the idea of the luminiferous ether as a medium for light propagation. I fully believe Hilbert+knowledge of tensors (Einstein didn't know the concept of tensors in 1905! [1]) would have developed general relativity as well.

[1] Einstein actually had an idea for general relativity much before he actually figured it out, he literally just lacked the mathematical knowledge to formalize it. He had to learn tensors in order to develop the Einstein Field Equations https://www.quora.com/How-did-Einstein-get-the-idea-that-he-...

It wouldn't be difficult for a different (very intelligent) human, it would almost certainly be impossible for LLMs, they have never done anything remotely analogous.

might work for say post the 1800's in literate countries, but for e.g. Rome our sources are so sparse and so far removed from the time they're writing about that it would be worse than nothing.

For a period like the Roman Empire, there's might be too little source material to even train the model to speak Latin, let alone to say anything about its world. IIRC, the entire surviving corpus of ancient Latin would fit into a couple of bookshelves - it's miniscule.

"What would have happened if ChatGPT was invented in the 17th century? MonadGPT is a possible answer. MonadGPT is a finetune of Mistral-Hermes 2 on 11,000 early modern texts in English, French and Latin, mostly coming from EEBO and Gallica. Like the original Mistral-Hermes, MonadGPT can be used in conversation mode. It will not only answer in an historical language and style but will use historical and dated references. This is especially visible for science questions (astronomy, medecine). Obviously, it's not recommended to follow any advice from Monad-GPT." Available to install and run locally -- or you can try it out for free online."

https://www.metafilter.com/201537/O-brave-new-world-that-has...

In the 1950s, most people believed that the Soviets made the biggest contribution to stopping the Nazis. However, today, most people think it was actually the Americans who played the biggest role in defeating the Nazis.

> "In 1945, the French public said the Soviets did the most to defeat Nazi Germany - but in 2024 they're most likely to say it was the Americans"[0]

[0] https://yougov.co.uk/politics/articles/49613-d-day-anniversa...

The Soviets put the men. The Americans put the materiel.

Stalin thought he would’ve lost if it wasn’t for Lend Lease.

The USSR might have lost if not for US supplies, but the Allies would definitely have lost if not for the Red Army. Technology has come a long ways since the 30s, yet even contemporary wars again emphasize that in the end it all just comes down to manpower.

And this is extremely remarkable if you think about it. Germany basically declared war on the world, and very nearly won.

Germany was not in a position to hit the US. They weren’t a great naval power. They needed to cross the Atlantic, but the English Channel was too big a hurdle.

Eventually the fact that the war was happening in their land would grind them down. Plus, the US had nukes and aircraft carriers by the end, which would have presented a challenging situation.

Well you're making a huge shift without accounting for it. The land war picture would have been radically different without the Soviets. In total they deployed more than 34 million soldiers during WW2 [1]. That's substantially larger than the contribution of every other ally, combined. The second largest force was the US with a total of 12.2 million soldiers (by the end of the war). [2]

So what would have happened in this scenario is difficult to even imagine, because Germany would have been under far less pressure. They were already working on the development of 'Projekt Amerika' [3]. It went nowhere, but without the pressures of the Red Army they would have had vastly more resources to expend on such ventures.

[1] - https://en.wikipedia.org/wiki/Red_Army

[2] - https://www.nationalww2museum.org/students-teachers/student-...

[3] - https://en.wikipedia.org/wiki/Amerikabomber

But you said the Allies definitely would have lost without the Soviet Union. Is it hypothetically possible that the Nazis might invent some mystery weapon without less pressure? Maybe… but it isn’t a sure thing.

Getting a plane across the Atlantic with the limited bomb and fuel load that entails might not have accomplished much. If only a couple bombs could have done the job, I guess allies wouldn’t have been building all those bomber fleets, right?

And they didn’t build a serious surface navy before the war, or in the first year-or-so when they were still trading partners with the Soviet Union. It seems there was something beyond manpower pressure holding them back.

A mystery weapon would not have been necessary. You're talking about a Nazi army that would now be facing an enemy force missing the majority of its soldiers, in a war that was already very close. The Nazis clearly would not have fallen, or even really come close to falling, without the Soviets.

Now I think your argument is basically the same, but for America. In that the Nazis were in no position to invade and occupy America - which I fully agree with. But victory in a war doesn't require you occupy every single enemy nation. The most common way wars end is in settlement. And I see no possible path where the Nazis would not have been able to achieve a favorable settlement.

I didn’t say it was either or. It was a symbiotic relationship. The Soviets were willing (and in a position they were forced to) lose millions. It was a tragedy.

Americans were in position to fund and aid it. Both Stalin and Khrushchev acknowledged that American material help was of great importance.

American lend-lease was 11% of the Soviet war effort. The vast majority of that 11% was consumer goods, not weaponry.

Then why did both Stalin and Khrushchev say it was important for the victory ?

Sorry but I’ve read tons of AskHistorians answers about this. I recommend you go there and search this same question. The Soviets needed that 11%. My understanding is that a lot of it helped mechanise the Soviet Army, ie trucks were delivered in large quantities.

From Claude I got this:

“In Khrushchev's memoirs, he recalled Stalin saying in a private conversation that without American aid through Lend-Lease, the Soviet Union "would not have been able to cope because we lost so much of our industry." Khrushchev himself wrote that Lend-Lease was "of utmost importance" and that "we would have been in a difficult position without it." “

Here is some more from Claude (it sounds about right but I don’t have to check this):

“ The significance of Lend-Lease goes far beyond the raw percentage of industrial output. Here's why it was so crucial:

1. Timing and Critical Shortages: - The aid arrived during the most critical period (1941-1942) when Soviet industry was being relocated east of the Urals - During this vulnerable period, American trucks, food, and materials helped keep the Soviet army mobile and fed - Without this bridge of support during the industrial relocation, the USSR would have faced severe shortages at its most vulnerable moment

2. Strategic Materials and Bottlenecks: - The Soviets received specific materials that were severe bottlenecks in their production: - Aviation fuel and high-octane gas - Aluminum for aircraft production - Radio equipment and communications gear - Special grades of steel and industrial equipment - These materials were critical multipliers that enabled Soviet production

3. Transportation and Logistics: - Nearly 450,000 trucks were provided, which revolutionized Soviet logistics - Before Lend-Lease, the Red Army relied heavily on horse transport - American Studebaker trucks allowed for rapid troop movements and superior logistics - This mobility was crucial for later Soviet offensive operations”

And why Soviets put the men? Because they started the war in the first place with Germans.

Too many people forget about that. They were allies at first.

That's very funny. I'd have thought they'd be hard pressed to get details of the Eastern front, while American involvement was right in front of them.

As far as I can tell, the Americans and Brits took too much credit. Then the Soviets and Russians insisted on more credit -- arguably too much. Of late I'm hearing historians say "Yeah, the Germans overextended themselves at the start and likely would have lost even if Hitler hadn't betrayed Stalin". I'm sure that analysis too will change.

I wonder if the “biggest contribution” was parsed differently in different eras. The Soviets clearly sacrificed the most, which probably was very felt by people who were surrounded by all the dying and personally experienced the violence.

The US industrial contribution is easier to understand looking back. It makes a lot of sense to us nowadays, looking at it in a table (not to cheapen it, it was an astonishing amount of stuff that was produced).

It seems entirely possible that the Soviets gave up more for the victory, while the US contributed more to victory.

Are there any successful models that weren't trained with RLHF, or using a system with RLHF. I'm curious if this could be done without a fine tune step that would't meaningfully bias this.

Normally I balk when commenters go “well they you’re the perfect person to go do it!”, but actually… this is the kind of thing that sounds like it could be a fun project if you’re legit interested. The necessary datasets are likely not hard to gather and collate, a lot of it is probably on places like Project Gutenberg or can be gleaned through OCR of images downloaded from various publicly available archives.

Granted, you’d need to spend about a year on this and for a lot of that time your graphics card (and possibly whole computer) would be unusable, but then if the results were compelling you’d get a cool 15 minutes of internet fame when you posted your results.

I got 15 minutes for basically a useless compiler and programming language that I spent 6 months on. Just for the effort-to-result ratio I feel like it's possible to do quite a lot better.

yes! There's this measure of historical expertise that involves "eating the brains", so to speak, of the people living back then such that if you time traveled back to a bar or street in [insert period], you could carry on a conversation about events going on in that time :) I would love something that uses newspaper fragments, books, etc. to simulate this experience!

The only reason LLMs “work” is because they are trained on a vast corpus of (text-based) human interactions online. The main reason LLMs weren’t a thing 25 years ago, was because there just wasn’t enough scrapeable and useful data available online…

Reduce the dataset to “knowledge as of year 1880” - and it’s not certain you’d even be able to “interact” with the LLM in any meaningful way…

The main reason LLMs weren't a thing 25 years ago is because they weren't invented yet, or many of the prior steps. And if they had been, we didn't have the compute create them.

I think I'm slow. Can you explain this again, maybe with more words?

Let's say we choose 1900 as the cutoff date. That means during training the model is only able to access material written before 1900. It would have a good knowledge about everything discovered in the 19th century and before. There's a great deal of mathematics, physics and chemistry available then. What if now we engage a discussion with that LLM on something discovered after 1900? Say transmutation and nuclear weapons, or general relativity, or the ZFC set theory.

Wouldn't it be easier to cutoff pre-2020-ish, and ask it to create the transformer architecture of gpt? 1900 is so long ago I doubt most documents are good quality if they've been digitised at all. Most likely just low quality scanned images of inconsistent, half-illegible typewriter documents. Transcribed with OCR at best.

Also so little training data from that era. Like, exponentially more data was created after, say, <year when most records become digitized = 1970>

The problem I see with any date after the popularity of the internet is that you just can't be sure of the right date. A lot of traditional web forums now have backdated forum posts that are clearly made by LLM with an implausible date: https://hallofdreams.org/posts/physicsforums/

You can use CommonCrawl - which has massive datasets going back to 2008 - and the Internet Archive.

Reply on news.ycombinator.com