I have no real quibble with the blog post itself, but I take issue with the title that calls it a "vintage model".
The blog post defines a "vintage model" as one that is trained only on data before a particular cutoff point:
> Vintage LMs are contamination-free by construction, enabling unique generalization experiments [...] The most important objective when training vintage language models is that no data leaks into the training corpus from after the intended knowledge cutoff
But as they acknowledge later, there are multiple major data leakage issues in their training pipeline, and their model does in fact have quite a bit of anachronistic knowledge. So it fails at what they call the most important objective. It's fair to say that they are working toward something that meets their definition of "vintage", but they're not there yet.
I was reading Nate Silver's book "On The Edge" and there is an interesting part where he takes predictions on the usage of nuclear weapons taken from just after World War 2 and compares them to what the Bayesian prediction would be given what actually happened.
Post World War 2, some people had the odds per year at 10%. Some of that is probably a mix of recency bias + not understanding how to use new weapons etc etc but as Silver points out, the odds were much lower.
I mention this only b/c the "could something trained on LLMs of the time predict the future" always makes me think of it.
>Have you ever daydreamed about talking to someone from the past?
It's going to be more like corresponding with someone from the past. We don't have much in the way of recorded speech from that area, so this will be built from written records. Much more than now, the written records are going to be formal and edited, reflecting a different pattern than casual speech or writing.
Having said that, this is cool. I recently had to OCR a two-hundred year old book with the usual garish fonts from that era. It was remarkably easy to do, and accurate.
> A language model trained from scratch exclusively on data from certain places and time periods to reduce modern bias and emulate the voice, vocabulary, and worldview of the era.
>Have you ever daydreamed about talking to someone from the past?
Fun facts, LLM was once envisioned by Steve Jobs in one of his interviews [1].
Essentially one of his main wish in life is to meet and interract with Aristotle, in which according to him at the time, computer in the future can make it possible.
[1] In 1985 Steve Jobs described a machine that would help people get answers from Aristotle–modern LLM [video]:
The idea of talking to a machine that has all of humanities knowledge and gives answers is older than electronic computing. It certainly wasn't a novel idea when Jobs gave that speech. At that time, the field of artificial intelligence was old enough to become US president.
Except... not at all? The vast majority of the training data required to create an artificial Aristotle has been lost forever. Smash your coffee cup on the ground. Now reassemble it and put the coffee back in. Once you can repeatably do that I'll begin to believe you can train an artificial Aristotle.
It is cool. I find the idea of trying to understand whether these types of models can come up with things like General relativity, or maybe some results really interesting.
20GB isn't enough for a 13B parameter model? I thought the 29-31B models could run on a 24GB GTX x090 card?
I'm currently shopping for a local LLM setup and between something like the Framework Desktop with 64-128GB of shared RAM or just adding a 3090 or 4090 to my homelab so I'm very curious what hardware is working well for others.
> 20GB isn't enough for a 13B parameter model? I thought the 29-31B models could run on a 24GB GTX x090 card?
Parameters are like Hertz - they don't really tell you much until you know the rest anyways. In this case, a parameter is a bfloat16 (2 bytes). I'm sure someone will bother to makes quants at some point.
> I'm currently shopping for a local LLM setup and between something like the Framework Desktop with 64-128GB of shared RAM or just adding a 3090 or 4090 to my homelab so I'm very curious what hardware is working well for others.
I grabbed a 395 laptop w/ 128 GB to be a personal travel workstation. Great for that purpose. Not exactly a speed demon with LLMs but it can load large ones (which run even slower as a result) and that wasn't really my intent. I've found GPUs make more usable local LLMs, particularly in the speed department, but I suppose that depends more on how you really use them and how much you're willing to pay to have enough total VRAM.
It's next to impossible to make your money back on local (regardless what you buy) so I'd just say "go for whatever amount of best you're willing to put money down for" and enjoy it.
I think that one could also take a much larger model (35B or 122B sized) and give it a thorough system prompt to only speak in the manner of a well educated Victorian/Edwardian era gentleman, if you want an "old timey" LLM.
As we learn how to train smarter models on less data, it’ll become more and more interesting to see whether models like this can invent post-1930 math, science, etc. and make predictions.
[Edit: serves me right for not reading tfa. My points are well-covered]
I have no real quibble with the blog post itself, but I take issue with the title that calls it a "vintage model".
The blog post defines a "vintage model" as one that is trained only on data before a particular cutoff point:
> Vintage LMs are contamination-free by construction, enabling unique generalization experiments [...] The most important objective when training vintage language models is that no data leaks into the training corpus from after the intended knowledge cutoff
But as they acknowledge later, there are multiple major data leakage issues in their training pipeline, and their model does in fact have quite a bit of anachronistic knowledge. So it fails at what they call the most important objective. It's fair to say that they are working toward something that meets their definition of "vintage", but they're not there yet.
I was reading Nate Silver's book "On The Edge" and there is an interesting part where he takes predictions on the usage of nuclear weapons taken from just after World War 2 and compares them to what the Bayesian prediction would be given what actually happened.
Post World War 2, some people had the odds per year at 10%. Some of that is probably a mix of recency bias + not understanding how to use new weapons etc etc but as Silver points out, the odds were much lower.
I mention this only b/c the "could something trained on LLMs of the time predict the future" always makes me think of it.
>Have you ever daydreamed about talking to someone from the past?
It's going to be more like corresponding with someone from the past. We don't have much in the way of recorded speech from that area, so this will be built from written records. Much more than now, the written records are going to be formal and edited, reflecting a different pattern than casual speech or writing.
Having said that, this is cool. I recently had to OCR a two-hundred year old book with the usual garish fonts from that era. It was remarkably easy to do, and accurate.
Related: https://github.com/haykgrigo3/TimeCapsuleLLM
> A language model trained from scratch exclusively on data from certain places and time periods to reduce modern bias and emulate the voice, vocabulary, and worldview of the era.
Discussed here: https://news.ycombinator.com/item?id=46590280
>Have you ever daydreamed about talking to someone from the past?
Fun facts, LLM was once envisioned by Steve Jobs in one of his interviews [1].
Essentially one of his main wish in life is to meet and interract with Aristotle, in which according to him at the time, computer in the future can make it possible.
[1] In 1985 Steve Jobs described a machine that would help people get answers from Aristotle–modern LLM [video]:
https://youtu.be/yolkEfuUaGs
The idea of talking to a machine that has all of humanities knowledge and gives answers is older than electronic computing. It certainly wasn't a novel idea when Jobs gave that speech. At that time, the field of artificial intelligence was old enough to become US president.
Imagine aiming for Aristotle and landing on Siri…
Except... not at all? The vast majority of the training data required to create an artificial Aristotle has been lost forever. Smash your coffee cup on the ground. Now reassemble it and put the coffee back in. Once you can repeatably do that I'll begin to believe you can train an artificial Aristotle.
Whoa, Alec Radford is on the list of authors! He was instrumental in building the original GPT models at OpenAI.
It is cool. I find the idea of trying to understand whether these types of models can come up with things like General relativity, or maybe some results really interesting.
Darn I've only got ~20 GB of VRAM. I really need to get a stronger machine for this sort of stuff.
20GB isn't enough for a 13B parameter model? I thought the 29-31B models could run on a 24GB GTX x090 card?
I'm currently shopping for a local LLM setup and between something like the Framework Desktop with 64-128GB of shared RAM or just adding a 3090 or 4090 to my homelab so I'm very curious what hardware is working well for others.
> 20GB isn't enough for a 13B parameter model? I thought the 29-31B models could run on a 24GB GTX x090 card?
Parameters are like Hertz - they don't really tell you much until you know the rest anyways. In this case, a parameter is a bfloat16 (2 bytes). I'm sure someone will bother to makes quants at some point.
> I'm currently shopping for a local LLM setup and between something like the Framework Desktop with 64-128GB of shared RAM or just adding a 3090 or 4090 to my homelab so I'm very curious what hardware is working well for others.
I grabbed a 395 laptop w/ 128 GB to be a personal travel workstation. Great for that purpose. Not exactly a speed demon with LLMs but it can load large ones (which run even slower as a result) and that wasn't really my intent. I've found GPUs make more usable local LLMs, particularly in the speed department, but I suppose that depends more on how you really use them and how much you're willing to pay to have enough total VRAM.
It's next to impossible to make your money back on local (regardless what you buy) so I'd just say "go for whatever amount of best you're willing to put money down for" and enjoy it.
How much system memory do you have? Llama.cpp can split layers across cpu and gpu. Speeds will be slower of course but it's not unusable at all.
This is cool. Is it possible to easily install with ollama?
The Python example is fascinating, and a good rejoinder to anyone still dismissing LLM’s as stochastic parrots.
Vintage is a funny thing to call this. Is it running on vacuum tube hardware?
I think that one could also take a much larger model (35B or 122B sized) and give it a thorough system prompt to only speak in the manner of a well educated Victorian/Edwardian era gentleman, if you want an "old timey" LLM.
As we learn how to train smarter models on less data, it’ll become more and more interesting to see whether models like this can invent post-1930 math, science, etc. and make predictions.
[Edit: serves me right for not reading tfa. My points are well-covered]