DeepSeek's free 685B-parameter AI model runs at 20 tokens/second on Apple's Mac Studio, outperforming Claude Sonnet while using just 200 watts, challenging OpenAI's cloud-dependent business model.
A bit like a syllable when you are talking about text based responses. 20 tokens a second is faster than most people could read the output so that’s sufficient for a real time feeling “chat”.
Huh, yeah that actually is above my reading speed assuming 1 token = 1 word. Although, I found that anything above 100 words per minute, while slow to read, feels real time to me since that’s about the absolute top end of what most people type.
It’s the generation speed. Internally LLMs use tokens which represent either words or parts of words and map them to integer values. The model then does it’s prediction on which integer is most likely to come after the input. How the words are split up is an implementation detail that can vary from model to model
Not somebody who knows a lot about this stuff, as I’m a bit of an AI Luddite, but I know just enough to answer this!
“Tokens” are essentially just a unit of work – instead of interacting directly with the user’s input, the model first “tokenizes” the user’s input, simplifying it down into a unit which the actual ML model can process more efficiently. The model then spits out a token or series of tokens as a response, which are then expanded back into text or whatever the output of the model is.
I think tokens are used because most models use them, and use them in a similar way, so they’re the lowest-level common unit of work where you can compare across devices and models.
Okay, can somebody who knows about this stuff please explain what the hell a “token per second” means?
A bit like a syllable when you are talking about text based responses. 20 tokens a second is faster than most people could read the output so that’s sufficient for a real time feeling “chat”.
Huh, yeah that actually is above my reading speed assuming 1 token = 1 word. Although, I found that anything above 100 words per minute, while slow to read, feels real time to me since that’s about the absolute top end of what most people type.
It’s the generation speed. Internally LLMs use tokens which represent either words or parts of words and map them to integer values. The model then does it’s prediction on which integer is most likely to come after the input. How the words are split up is an implementation detail that can vary from model to model
Not somebody who knows a lot about this stuff, as I’m a bit of an AI Luddite, but I know just enough to answer this!
“Tokens” are essentially just a unit of work – instead of interacting directly with the user’s input, the model first “tokenizes” the user’s input, simplifying it down into a unit which the actual ML model can process more efficiently. The model then spits out a token or series of tokens as a response, which are then expanded back into text or whatever the output of the model is.
I think tokens are used because most models use them, and use them in a similar way, so they’re the lowest-level common unit of work where you can compare across devices and models.
it’s a little similar to words per second
Not an answer to your question, but I thought this was a nice article for getting some basic grounding on the new AI stuff: https://arstechnica.com/science/2023/07/a-jargon-free-explanation-of-how-ai-large-language-models-work/