News for hackers. Stuff that matters.

A comment on Cerebras-GPT vs. LLaMA AI Model Performance Comparison. Parent comment

Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by thequadehunter on Wednesday March 29, 2023 @07:39PM (#35365532)
I've been wondering...does it even matter at this point when you can just pull out a calculator in langchain?

Also side note...I went to high school with the author and hadn't heard about him for a while lol. That was a neat surprise.

[ Parent | on HN ]
Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by spaethnl on Wednesday March 29, 2023 @09:02PM (#35366298)
I suspect it does matter because the inability to understand arithmetic likely effects higher level learning. For example, If you want it to learn statistics, it likely needs a good understanding of arithmetic.

All tasks that rely on it will perform worse than expected.

[ Parent | on HN ]
Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by thequadehunter on Wednesday March 29, 2023 @09:25PM (#35366508)
Oh, I see. It's about a cap in baseline ability rather than whether it's technically possible to give it the extension?
[ Parent | on HN ]
Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by est on Wednesday March 29, 2023 @09:18PM (#35366451)
> does it even matter at this point when you can just pull out a calculator in langchain?

That's alrithmetics. For math problems:

Each alien chicken has 3 legs, each alien rabbit has 7 legs, for a total of 33 legs, how many alien chicken are there?

Now try to calculator that!

[ Parent | on HN ]
Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by thequadehunter on Wednesday March 29, 2023 @09:26PM (#35366513)
Apparently it can.

https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its...

[ Parent | on HN ]
Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by sbierwagen on Wednesday March 29, 2023 @10:18PM (#35366974)
A couple people have used LLMs as frontends to theorem provers/SAT solvers, which would handle that fairly easily: https://twitter.com/AlbertQJiang/status/1584877475502301184 https://twitter.com/ericzelikman/status/1605385835624378368
[ Parent | on HN ]
Re:Cerebras-GPT vs. LLaMA AI Model Performance Comparison by computerex on Thursday March 30, 2023 @12:02AM (#35367777)
https://github.com/hwchase17/langchain/tree/master/langchain...

Model would probably take that description and write python code to compute the answer.

[ Parent | on HN ]