The Six Five team discusses GenAI Weekly: NotebookLM, NVIDIA LLM, OpenAI, Microsoft
If you are interested in watching the full episode you can check it out here.
Disclaimer: The Six Five Webcast is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors and we ask that you do not treat us as such.
Transcript:
Daniel Newman: … a lot going on in Gen AI. It’s like every time you turn around there’s something cool. You want to play with RAG, there’s NotebookLM. You want another massive mega LLM out and video was here to make news. OpenAI and what I think is a somewhat ridiculous evaluation, but you know what, the money’s piling in, so what the hell do I know? Copilot. Take your pick man. Go ahead and I’ll fill in the blanks. But what caught your eye in the Gen AI news of the week?
Patrick Moorhead: Yeah, we didn’t catch this last week, but Google brought out a service called NotebookLM, and essentially what it is an almost infinite RAG model where you can do up to 2 million tokens, up to 50 different data types, and even have cool buttons you can press. “Hey, you want a summary, you want a table of contents? Do you want a blog style?” And this is Wild, Daniel, is it will actually do a podcast between two people, a man and a woman. And you can just imagine where this can go in the future. First of all, when it comes to personal RAG for a lot of use cases that even people like us do, you want to pile in all this content from disparate sources and be able to query it, summarize it, do all that stuff. And it’s been a criticism that I’ve had of some of the big makers here, because quite frankly between Google and Microsoft, their personal RAG capability just was non-existent.
You had to go to OpenAI, you had to go to perplexity to do that right. Perplexity has a cool feature that says, “Don’t get anything from the internet, just take whatever is in the content that you’ve thrown me.” The context window isn’t high, but this context window is insane. Like 2 million from Google. Of course, we can improve, does not support video content, does not support Word, does not support XLS, any of the Microsoft formats. You can put in a link to YouTube, but only if YouTube has had time to go in and do a transcript. But I’m going to be playing around with it a lot more in the future, but I’m super excited about this. Microsoft did a major drop, two fronts. So first of all, overall Copilot, and again, I’m still piling through it right now, but to me it’s the AI assistant we had always envisioned, right? There’s more personalization, there’s voice, there’s vision.
And by the way, vision not in the way that you might think. And there’s also a much more simplistic interface about what you can do. There’s a daily readout that essentially is the top news stories. Doesn’t look personalized yet, but it’s read as if it’s an NPR episode. So I’ve kicked the tires a little bit on it. And then, I think the biggest one is essentially you just have a conversation with it, right? It’s like ChatGPT advanced voice. And by the way, Gemini cranked theirs out. Google did, they cranked theirs out the same week, and I think Meta even came out with theirs. But I would say, this was the consumer launch for Microsoft and very provocative, better than I thought. By the way, more risky for them but big. And you can get this on your smartphone, Android, iOS, you can get this just on a web browser.
You can get this on Edge, you hit the Copilot key, it automatically comes up with the new capabilities. Super exciting. There’s no RAG capability on that. I think Copilot is the last service to not have RAG capabilities. Here’s the second thing that I’m super excited about, and these were Copilot+ features, okay? And Copilot+ features have the integrated NPU, amazing battery life, amazing performance. But I am so super excited about the features that are coming out. Because there’re features that we all use. Do you use Microsoft Photos? Do you use File Explorer? Everybody who uses Windows uses file Explorers to go in and do searches on that. And do a semantic search. “Hey, show me all my files that I’ve created about a client, an SOW in the last 12 months.” And this sucker pops up. “Show me a photo of me on vacation with my three kids and wife and give me those photos.”
For photos. Upscaling, photos up to 8X. You have a low quality photo, upscale it. You want to do generative filling erase, a lot of the features that we’ve seen through a paid Adobe subscription. It does that too. One of the most popular Android features right now getting a lot of use, and this came out through Samsung, is circle to search. Microsoft has their version called Click To Do. On vision, vision is like this. Vision is, I’m on a web page and I just start querying it, and I literally ask the questions or type in the questions and it’s looking at everything. To me it’s a real time recall, okay, to be able to query, and then it’s using data from the internet to come in and make it even better. Finally, just to make a long story longer, good progress on VPNs for ARM. Looks like going to fill in that gap in the first quarter of ’25. Super excited. And of course, recall. Bring it on Windows Insider, not there yet. I’m excited.
Daniel Newman: Yeah, there’s so much going on, Pat. We’ll talk a little bit more about some of the interesting stuff with Microsoft too when we get to the signal, what we’re doing there in the lab and the testing stuff that we’re doing. But you hit on a lot of them, Pat. I think a couple of the most exciting things for me is one is just the very accessible RAG technology with what they’re doing with Notebook. I’ve seen some reports, I’ve played with it a little bit. I’ll be candid, I got to play with it some more. I’d like to start putting some of our large data sets from our intelligence platform through there and some of the big reports and start querying it, just to see what’s available now. You can basically ingest the data, use language, and get really thoughtful responses with very little effort. Pat, I mean, I think about it a year and a half ago, what it would’ve taken to build these capabilities and how it’s been basically dropped at our doorstep now. We have what’s going on with o1 and these reasoning engines. I mean, you have systems now that basically have the capacity to reason like an executive team. Remember when Benioff talked about having AI in your boardroom? I mean, the problem is these things are way too inefficient power-wise right now to actually be used all day every day.
But if you could use o1 the way we’re using search and ChatGPT, Pat, you may not even need me anymore. You could just literally talk to the reasoning engine when you’re trying to work your way through a modest dilemma. Like, how do I get my biceps over 18 inches? The other thing I think that was really interesting to me this week too is this sort of pile-on thing that’s going on. We talked a little bit about the OpenAI valuation last week, but this NVIDIA large language model got me rethinking now, Pat. We’ve got all these different players contributing here to these frontier models and now you have a company like NVIDIA that’s actually just added one more layer to the stack. It’s always been able to do sort of specialty models. It’s been in this space a long time, and of course, with its own hardware, it can train pretty much anything it wants.
But you’ve got this really, I’d say, critical two lanes running of open source with Meta and NVIDIA that are really pushing and barreling down. And then, of course, you’ve got this black box models that are gaining in popularity. And some would make me wonder why this NVLM thing didn’t get more attention this week. It got some, but why it didn’t get more. Because if I’m an investor, I mean, you heard Apple pulled out. And obviously, there’s been a ton of brain drain in OpenAI. But if I’m an investor and I’m investing 6 plus billion dollars as a group in 157 billion evaluation, I would want to know what makes it sticky in the long run. And when you’ve got companies like Meta with Mark Zuckerberg, I think he’s not the second-wealthiest person on the planet, and you’ve got Jensen saying, “Hey, we can open source this, deliver a product. I think this NVLM one outperformed GPT4o in a number of different benchmarks.”
Not all of them, but some of them. And then you start to think how you could integrate that with NIMS. You could integrate that with CUDA, you could integrate that with hardware. And gosh, then you start to see why Accenture wants to double down. But I don’t know, Pat, I just look at this and I go, “This is going to get crowded. This is going to get noisy.” And unless there’s something in the back room that’s so exciting and so… I’m not sure there’s an advantage. And then on top of that for ChatGPT and OpenAI, I’m not sure there’s an exit other than going public. It’s really hard to go public and evaluation in the hundreds of billions when you haven’t made a dollar in profit yet. So there’s a lot going on, Pat. I think you hit a lot of the tech well. This space is evolving and changing so quickly. It’s really, really exciting times.
