O
Story
Vintage Large Language Models
Someone has sort of done this:

https://www.reddit.com/r/LocalLLaMA/comments/1mvnmjo/my_llm_...

I doubt a better one would cost $200,000,000.

We need a library of Alexandria for primary sources. If we had source transparency then referencing back to original sources would be more clear. We could do cool things like these vintage models to reduce bias from current events. Also books in every language and books for teaching each language would help with multimodality. Copyright makes it difficult to achieve the best results for LLM creation and usage though.
Ironically enough, that would be practical for "vintage LLM" - perhaps (morally) obligatory?
As if the language models currently would give a damn about copyright...
The problem is they have to hide their sources due to copyright. So they train on copyright data but must obscure it in the output. Thus they must hide the sources of truth making it impossible to fact check them directly and the reason that hallucinations are so common and unavoidable in the current pattern.
The talk focuses for a bit on having pure data from before the given date. But it doesn't consider that the data available from before that time may be subject to strong selection bias, based on what's interesting to people doing scholarship or archival work after that date. E.g. have we disproportionately digitized the notes/letters/journals of figures whose ideas have gained traction after their death?

The article makes a comparison to financial backtesting. If you form a dataset of historical prices of stocks which are _currently_ in the S&P500, even if you only use price data before time t, models trained against your data will expect that prices go up and companies never die, because they've only seen the price history of successful firms.

  • dboon
  • ·
  • 2 weeks ago
  • ·
  • [ - ]
The talk explicitly addresses this exact issue.
  • alalv
  • ·
  • 2 weeks ago
  • ·
  • [ - ]
It mentions that problem in the first section
Not a financial person by any means, but doesn't the Black Swan Theory basically disproves such methods due to rarity of an event that might have huge impact without something to predict (in the past) that it might happen, or even if it can be predicted - the impact cannot?

For example: Chernobyl, COVID, 2008 financial crisis and even 9/11

All models are wrong, but some are useful.

If you had a financial model that somehow predicted everything but black swan events, that would still be enough to make yourself rich beyond belief.

This would be a good way to verify emergent model capability to synthesize new knowledge.

You give an LLM all the information from right before a topic was discovered or invented, and then you see if it can independently generate the new knowledge or not.

It would be hard to know for sure if a discovery was genuine or accidentally included in the training data though.

I saw Musk repost a boast that Grok created a whole new ("superior") element design for a incandescent bulb using Edison's patent. The implication was that Grok was superior to Edison's team. I was just sat there thinking about the 100+ years of incandescent bulb research that Grok has sucked up from various science papers and random Internet archives. Surely none of that was any help at all /s.
I've been wanting to do this on historical court records - building upon the existing cases, one by one, using llms as the "Judge". It'd be interesting to see which cases branch off from the established precedent, and how that cascades into the present.

Any thoughts how one could get started with this?

Using old models is a good way to received less biased information about an active event. Once a major event occurs information wars happen that try and change narratives and erase old information. But because models were trained before this the bias that the event causes is not yet present.
  • lukev
  • ·
  • 2 weeks ago
  • ·
  • [ - ]
I’m sorry I don’t quite follow… how can a model provide information at all about events it was trained before?
For instance I want information about 2 countries currently at war. By asking about these countries from an older model then we get more factual information about the countries. If we ask about them and the information is seeded from news articles etc AFTER the war started then they will be biasedly influenced and often have disclaimers like "But it should be noted that x y z" showing that there is some MAJOR bias that occurred from the training on the news.

If I want an unbiased reason for what happened before a war started i would want all the information about 2 countries at different points before the war. Because after a military war starts an INFORMATION war also starts. Propaganda will be spread from both sides as wars are just as much about global support as they are about military dominance.

Overspecialization of models is a thing.

>Overspecialization of models, often referred to as overfitting in machine learning, is a condition where a model learns the details and noise in the training data so well that it negatively impacts its performance on new, unseen data. This prevents the model from being able to generalize its knowledge effectively.

You provide the info... and the bias.
Everyone introduces bias. But for instance getting a model trained pre war vs after a war starts is super different. If I want to get raw information about 2 nations then models are in some ways a good source. Because most other parts of the internet can get changed or wiped. A model is "stuck" with the information it had exactly at that point so cannot be directly affected by new information attacks.

It is crucial to have a good framework in how you ask your questions though to avoid bias when using these systems and to try and focus on raw facts. To test ideas I like to make it fight for both opposite extreme sides of an argument then I can make up my own mind.

I like the idea of using vintage LLMs to study explicit and implicit bias. e.g. text before mid-19th century believing in racial superiority, gender discrimination, imperial authority or slavery. Comparing that to text since then. I'm sure there are more ideas when you use temporal constraints on training data.
Cool idea. This might be a interesting literary project along this line ;-)

https://www.gutenberg.org/cache/epub/86/pg86-images.html

Over the long term LLMs are going to become very interesting snapshots of history. Imagine prompting an LLM from 2025 in 2125.
  • lukan
  • ·
  • 2 weeks ago
  • ·
  • [ - ]
I would probably prefer wikipedia snapshots (including debate) as a future historian.
The more options you have, the better IMO.
You're right: I wish OpenAI could find a way to "donate" GPT-2 or GPT-3 to the CHM, or some open archive.

I feel like that generation of models was around the point where we were getting pleasantly surprised by the behaviors of models. (I think people were having fun translating things into sonnets back then?)

Maybe in the sense that a CueCat is interesting to us today.
  • ijk
  • ·
  • 2 weeks ago
  • ·
  • [ - ]
I was hoping that this would be about Llama 1 and comparison with GPT-contaminated models.
I love the ideas about how we might use historical LLMs to inquire into the past!

I imagine that (the author hints at this), to do this rigorously, spelling out assumptions etc, you’d have to build off theoretical frameworks used to inductively synthesize/qualify interviews and texts, currently around in history and the social sciences.

Very cool! I’ve been wanting to do this do a long time!