So "less is more" or what this video should tell ? :) Where is actual comparison of these AI capabilities at actual work, not on paper...
@DonaAI
8 ай бұрын
One of the benchmark used in the ranking is The "ELO score". This benchmark is way more realistic and close to normal usecase. People like you and me have access to a webpage where we can type our prompt and get two responses from two separate LLMs. The name of the LLM is not shown and we then have to vote for what we think is the "best response". As you can see in the ranking, this is where GPT-4 and Mistral 8x7b or Mistral-medium are really close to GPT-4 while using way less resources.
@starblaiz1986
8 ай бұрын
Me a year ago: "Wow, this GPT 3.5 is pretty neat and powerful! Its a shame it will probably be many many years before I can run this locally..." Me today: "HOW DID WE GET HERE SO FAST I JUST BLINKED!" 😅
@BruceWayne15325
8 ай бұрын
I agree, I think in the long run Mistral will overtake ChatGPT because ChatGPT is spending a lot of time working on guard rails, and otherwise neutering their model, and making it less competitive, and less useful. As a writer I prefer using both Mistral and Claude 2. I know that Claude 2 is also moderated, but it's better at prose than any other LLM I've found, so it can give me better suggestions when I'm editing my book.
@webinnovationspartners9293
7 ай бұрын
Great work covering this amazing and rapidly evolving space.
@nedamahmoudi8216
7 ай бұрын
Great content, great presentation 👏🏻👏🏻
@MrMShady
8 ай бұрын
Great metaphors/comparisons, the best I've found so far. I think it's the key to learning. Im subscribing and waiting for more knowledge.
@jaysonp9426
8 ай бұрын
I hope so....open source 👍👍 plus OpenAI is convinced more compute is better. Mistral is crushing the small models. AGI will be stacked small models.
@samatkadyrov5779
8 ай бұрын
I love the microphone upgrade!
@DonaAI
8 ай бұрын
Thanks!
@PaulSchwarzer-ou9sw
8 ай бұрын
🎉
@evelynshi4475
8 ай бұрын
👏👏
@metalmonkey128
8 ай бұрын
Good one!
@notu483
8 ай бұрын
This explanation of MoE is incorrect. The name « expert » is a poor choice of name, but each “expert” is only a feed-forward network that is not specialized in a single field. In fact according to their research their activations are unpredictable and pretty much random.
@DonaAI
8 ай бұрын
I would say it's rather oversimplified than incorrect. I feel like I would have lost a lot of people if I went that deep into the topic (but I could still make a more detailed video in the future). You are right. It's not like every math, physics or music topic will be redirected to a specific expert. It's more like every "experts" is a neural network on its own and the gating network give the work to the "best expert". It can also give it to multiple experts at once. One expert can also be a multititude of other experts 🤯 If you want to learn more about MoE, this is a really good article on huggingface.co: huggingface.co/blog/moe
Пікірлер: 16