text stringlengths 0 2.68k |
|---|
Also, another thing that happened that I believe is extremely cool is that we are acquiring new people that are like that. What happened is that one of my colleagues, Oran, which is truly a genius of programming, together with Yossi - they are two really superstars that nobody knows. For example, Yossi was the one that... |
\[52:00\] Also, another guy, which is the one that implemented the expires in the hash type elements, single elements - he's very talented, and he optimizes this thing to a so low level. And for example, one thing that I still love about Redis, and I think that we have an edge here compared to, for example, Valkey, is ... |
For example, one of the leaders of the project said, after seeing in my blog post the vector sets thing, "We will never do something like that", because they don't trust in this line of, you know, AI is this terrible hype. For example, this is one way to approach the problem. I am, for example, an AI enthusiast, and th... |
**Jerod Santo:** Sure. |
**Salvatore Sanfilippo:** So I believe they will diverge a lot. |
**Jerod Santo:** That's really interesting. So -- and you're happy about that, because now you have basically... Like, Redis gets pushed forward, but also Valkey gets pushed forward in different directions, and different things work for different people, in different circumstances, so it's kind of spurring some innovat... |
**Salvatore Sanfilippo:** No, no. I heard about it, but I didn't go to check what they are doing. |
**Jerod Santo:** You didn't check it out. This was just a re-implementation with a SQLite backend, trying to just get the Redis API up and running with a different backend. It was kind of a greenfield thing that was inspired by the re-license, but I thought it was pretty cool, the explosion of new projects and ideas th... |
**Salvatore Sanfilippo:** Yeah. It's like, all the teams together optimize to search the space of potential possibilities, basically. |
**Jerod Santo:** Well said, yeah. That's cool. So where do you take it from here? AI, it sounds like, huh? Redis AI. \[laughter\] |
**Salvatore Sanfilippo:** I think that vectors -- I mean, learned embeddings. Because vectors -- okay, the way I am implementing vector sets is one way that allows the user to also use vectors outside AI. For example, I implemented binary quantization, and if you want, you can put binary vectors where just each bit is ... |
\[56:22\] I believe this is going to be useful in the future, even if I am not a fan of rug, because I believe that RAG created a lot of issues. For example, one of the reasons I believe Claude Sonnet is often in the real world so strong for programmers is that when you attach a file there, the model sees all the file.... |
**Jerod Santo:** So do you know what technique Claude Sonnet is using to get that information that's not RAG? |
**Salvatore Sanfilippo:** They just put everything in the prompt. |
**Jerod Santo:** Oh, they put all -- I mean, do they have a huge prompt, I guess? Big context. |
**Salvatore Sanfilippo:** Exactly. That's the reason why Cloud Sonnet, after you fiddle with it a little bit, says "Return in five hours." |
**Adam Stacoviak:** Yeah, I was gonna say... |
**Jerod Santo:** \[laughs\] "Return in five hours..." |
**Adam Stacoviak:** My experience with Claude so far has been just that. They almost -- they start to yell at you, essentially, when you have too many back and forths. And I'm like "That's the whole point, is to have the back and forths", essentially. The prompt, the response, the prompt response, and you iterate and m... |
**Salvatore Sanfilippo:** I agree, but I believe that this is a feature, and the way why so many programmers speak at the Claude. Why I'm saying this - basically, like Redis, they are putting you face to face with the limitation of the system. Instead, what Open AI does is to implement two things. One is memory, which ... |
**Adam Stacoviak:** They do, that's true. I experienced that as well. |
**Salvatore Sanfilippo:** So I am more happy \[unintelligible 00:58:39.18\] but they can understand very well what's happening. And also, DeepSeek V3 and R1 does the same, that when you attach a file, the file is completely inside the prompt. And since attention is quadratic, it burns a lot of GPU compared to using RAG... |
**Adam Stacoviak:** How much are you studying this? Because this is something that I'm still sort of catching up to how the GPU interacts with a model. Let's just say you've got an off the shelf one, 24 gigs of VRAM. How does that translate to the model being in the VRAM to consuming the GPU's ability to compute? How a... |
**Salvatore Sanfilippo:** \[59:44\] Basically, what is happening right now is that all the frontier models are MOE, mixture of experts. They are no longer like, for example, LLaMA 3.3 or the initial models, dense models. Starting with GPT-4, Open AI started with the idea of using a mixture of experts. Basically, the mo... |
So basically, with MOEs you still need the VRAM to contain all the model in your RAM. So for example, if you have -- let's talk about DeepSeek version three, which is more or less as big as Claude, and we have the actual data public for DeepSeek. It's a 600 billion parameters model. |
**Adam Stacoviak:** That's big. |
**Salvatore Sanfilippo:** Yeah, it's very big. And it seems it is 8-bit for a parameter. You need 600 gigabytes just for the model. And then you have the cache for the \[unintelligible 01:01:26.25\] And for a very long context, you need more RAM. So basically, with mixture of experts you need less GPU, and you need mor... |
**Adam Stacoviak:** Let's clarify runs. Tell me tokens per second. |
**Salvatore Sanfilippo:** Four tokens per second in my computer right now. |
**Adam Stacoviak:** Which is so slow. |
**Salvatore Sanfilippo:** It's very slow. |
**Jerod Santo:** \[laughs\] But it runs, Adam. It runs. |
**Adam Stacoviak:** Yes, I know. I want to be clear though, because -- |
**Jerod Santo:** 600 gigabytes, he said. |
**Adam Stacoviak:** Yes. No, I get it. But, I mean, you've got people who are dabbling more and more into this, and runs and is usable is potentially two different things, because... It can run, and I do agree with that. Your hypothesis that more RAM and less GPU makes sense because the Mac, the M1 systems or the M sys... |
**Salvatore Sanfilippo:** \[01:03:52.01\] Yeah... Also, a lot of people don't realize that if you are talking in English with the model, often a single word is one token. However, because of how the tokenization work of large language models, if you talk with them in Italian, for example, since the Italian was very und... |
**Adam Stacoviak:** It really makes you weigh your words that you put into the prompt too, right? Because the weight and the structure -- it almost kind of brings back prompt engineering in a way, because you want to, not so much engineer, but you want to be particular with how you prompt the LLM in this scenario, when... |
**Jerod Santo:** So what you're saying, Adam, is don't waste precious tokens by thanking your models when you're done. |
**Adam Stacoviak:** Precisely. |
**Jerod Santo:** \[laughs\] |
**Adam Stacoviak:** I longer -- that's a deep cut, Salvatore, because I am the person that says "Thank you" or is generous with my words, let's just say... "Can we please", you know... Or "Let's do this", versus "Generate", which is -- |
**Salvatore Sanfilippo:** Me too, but this is nothing. The trick is when you start to have a mid-longer chat, to start a new one... Because since attention is quadratic, once they go -- okay, there is the KV cache, but still, you are accumulating a too much longer chat... And instead, you start anew, fresh. Also, somet... |
**Adam Stacoviak:** Yeah, I like that. So you pull out the latest -- you may have like a 10-minute session, let's just say, a couple of back and forth, but then you take the final artifact you're liking from that session, sort of whole brand new session, and bring that latest version of whatever you're liking to that n... |
**Salvatore Sanfilippo:** Exactly. |
**Adam Stacoviak:** Okay. |
**Jerod Santo:** That's a pro tip right there. |
**Adam Stacoviak:** There you go. Prompt engineering is back. |
**Jerod Santo:** It's back... Until we have sufficient VRAM that we don't need these things. So Salvatore, when do you think this new Redis vector set stuff you're working on will be baked? When will people begin using it? It's out there, it's somewhat open source? What's the status? |
**Salvatore Sanfilippo:** I think it will be released in one or two months, something like that. |
**Jerod Santo:** Cool. And what are some use cases or some people that are currently using Redis that they would be like -- what would this unlock for certain folks? |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.