Live data from Hacker News

How to scale LLMs better with an alternative to transformers

hazyresearch.stanford.edu

31–35 of 35 posts

Re: How to scale LLMs better with an alternative to transformers

#31
post #23
post #2

I wonder how a decentralized, hierarchical LLM would perform. For example: LLM A is trained on all of Wikipedia LLM B is trained on all of Hacker News LLM C is trained on all of Project Gutenberg User asks question Q on webservice W. W sends Q to A and B. Then W sends a question to C "Hey C, I have a user who asked Q. Here is A's reply and B's reply. Given those, how would you answer Q?" Would the answer be as good a…

Isn't that more or less how GPT-4 works? multiple "expert" LLMs giving input depending on the context?[0] [0] https://the-decoder.com/gpt-4-architecture-datasets-costs-an... the biggest issue is if you have too many specialists and spin a lot of them to reply to the same query and after that discard the less optimal answers. Your answer quality might improve, but the computing costs could skyrocket without some smart…

A huge misconception is that MoE is an ensemble of discrete models, when it is in fact multiple FFNN modules that share an attention and embedding module.

Basically the idea is that there's some pars of the model (attention/embedding) that should be trained on everything and used in every inference and other parts (the FFNN) that are fine to specialize on certain types of data (via a routing module that is also trained).

[0] https://arxiv.org/pdf/1701.06538.pdf [1] https://arxiv.org/pdf/2112.06905.pdf

EDIT: Specifically GLaM model architecture. Each MoE layer (the bottom block) is interleaved with a Transformer layer (the upper block). For each input token, e.g., ‘roses’, the Gating module dynamically selects two most relevant experts out of 64, which is represented by the blue grid in the MoE layer. The weighted average of the outputs from these two experts will then be passed to the upper Transformer layer. For the next token in the input sequence, two different experts will be selected.

Re: How to scale LLMs better with an alternative to transformers

#33
post #30

Earlier quoted context omitted.

The idea of decentralized hierarchical LLMs is interesting but your chosen example is not a good illustration as all three of these data sources are small and insufficient, any model trained solely on any of them will not be a good model for anything. Other things being equal, data quality and domain matters a lot, but a hundredfold increase in data quantity makes an even larger difference. Datasets like those can be…

That is not so certain. Microsoft's "Textbooks are all you need" is a case in point. https://news.ycombinator.com/item?id=36413768

That paper kind of does the same thing that my comment above proposed, starting with as large dataset as they can get and then filtering it to extract a much smaller dataset focused on a specific task that still is larger than all of English Wikipedia.

Re: How to scale LLMs better with an alternative to transformers

#34
post #2

I wonder how a decentralized, hierarchical LLM would perform. For example: LLM A is trained on all of Wikipedia LLM B is trained on all of Hacker News LLM C is trained on all of Project Gutenberg User asks question Q on webservice W. W sends Q to A and B. Then W sends a question to C "Hey C, I have a user who asked Q. Here is A's reply and B's reply. Given those, how would you answer Q?" Would the answer be as good a…

I dunno, but humans who are experts in multiple fields are often more useful than humans who are experts in just a single field.

Re: How to scale LLMs better with an alternative to transformers

#35
post #21

...from the same team that brought you FlashAttention, S4, H3, and Hyena. As always, we have to wait until this has been tested at much larger scale.

are those good or bad

FlashAttention is an amazing improvement over the previous state of the art. The others are still highly experimental, but seem like they'll at least contribute significant knowledge to whatever ends up surpassing the Transformer, (assuming something does).
Post reply on HN