Falcon LLM, JP Morgan IndexGPT, & TikTok Tako

AI Daily | 5.26.23

Discover the latest in artificial intelligence on AI Daily. In this episode, we cover three major stories. First, we explore Falcon LLM, an impressive open-source language model with 40 billion and 7 billion parameter models. Falcon excels in hugging faces evaluations and offers unique features for chat-based experiences. We also discuss its licensing model, providing free access for personal and research use and a fair royalty structure for commercial applications. Next, we dive into JPMorgan's patent filing for IndexGPT, their foray into AI-driven securities analysis based on customer needs. We highlight the growing trend of companies developing their own GPTs and the ambitious AI goals of JPMorgan's extensive team. Finally, we delve into TikTok's exciting venture into chatbots with TikTok Taco, a feature that analyzes videos and provides answers to user questions. Join us as we explore the cutting-edge advancements of Falcon LLM, JPMorgan IndexGPT, and TikTok Taco, and their impact on AI and society. Don't miss out on this engaging episode!

Links Mentioned

Key Take-Aways

Falcon LLM

  • Falcon LLM, a newly released open-source language model, offers a 40 billion parameter model and a 7 billion parameter model.

  • Falcon is currently topping hugging faces charts for their evaluations, showcasing its impressive performance.

  • Falcon provides both a raw unfiltered model and an instruct model for chat-based experiences.

  • The licensing model for Falcon allows free access for personal and research purposes, with a fair royalty-based structure for commercial use, similar to Unity's successful model.

JPMorgan IndexGPT

  • JP Morgan has filed a patent for IndexGPT, which focuses on analyzing and selecting securities based on customer needs.

  • Financial institutions like JP Morgan are building their own GPTs internally, following the trend of companies developing their own language models.

  • JP Morgan's announcement appears to be a strategic PR move to showcase their AI efforts and drive value, emphasizing their large team of data scientists, machine learning engineers, and AI researchers.

  • The use of GPTs in financial services, despite being banned publicly, is rumored to be prevalent within JP Morgan, leading them to trademark their own IndexGPT.

TikTok Tako

  • TikTok is introducing chatbots similar to Snapchat, aiming to analyze videos and provide answers to user questions based on the content.

  • The introduction of chatbots by TikTok was discovered by an analytics company, indicating that it is still in the testing phase and has not been officially announced.

  • The chatbot feature can understand the videos users are watching and suggest related questions or serve as a search tool for finding more content.

  • TikTok's expertise in computer vision and data collection from videos positions them well to gather valuable insights and improve their search engine capabilities, potentially impacting competitors like Google.

Follow us on Twitter:

Subscribe to our Substack:


Transcript:

Ethan Aldrich:

Good morning and welcome to AI Daily. Today's a fantastic show covering some great technical innovations and also some interesting announcements from big players in the space. So we're going to kick off with Falcon LLM. So Falcon is a new LLM that was just released, trained, open source. There's a 40 billion parameter model and a 7 billion parameter model. The most notable news here I saw was this is topping Hugging Face's charts for their evaluations. Farb, can you tell us a little bit more about what you saw with Falcon?

Farb Nivi:

So this is pretty interesting. So Falcon is the new leader. Who knows how long this will last. They have a, you know, sort of a raw unfiltered model. They have an instruct model as well. If you want something that's more specific for a chat based experience, they also have what they call this little brother, a 7 billion parameter model. And, uh, you know, one of the cool things I found, I think they use four or so different benchmarks to, to rate these different LLMs against. And one of them is called Hellaswag. And it's a model that for, you know, I don't know if anybody knows yet, but humans perform at about 95% or still some of the state of the art AI LLM models are only performing at around 50%. So I just thought that was a cool, cool name for a benchmark. Hella swag.

Ethan Aldrich:

Conner, what'd you see here with Falcon? What was interesting to you?

Conner Aldrich:

Yeah, they released some of their data and stuff, but I think the most interesting part here is the licensing model that they're pursuing. So for personal and research uses, it's completely free and open. But then for commercial uses, it's kind of like Unity, the game engine where if you're making under a million in revenue, it's also free. But once you hit above that, you have to pay a little bit of royalty, which is a pretty fair model. It's worked very well with Unity. So you can see it working well with AI models.

Ethan Aldrich:

Yeah, this commercial licensing is super important. I think it opens it up to small businesses to use and kind of still use commercially. It's not too constricted like something like Llama, but you know, the model providers get their royalty when it starts to get bigger. Unity's had great success with it. And I really like this kind of framework here for open source models. So fantastic. Let's move on to our second story. We're talking about JP Morgan, the financial giant, of course, they have filed a patent for what they call index GPT. So very broad right now, but quote unquote, analyzing and selecting securities based on customer needs. You know, JP Morgan's been famous for the past month or two saying, hey, none of our employees, we don't want you using chat GPT. You know, all these financial institutions are starting to build their own GPTs internally. We saw Bloomberg GPT that we talked about a bit ago, and now it looks like JP Morgan wants to get in the fold. Farb, what's your take on this?

Farb Nivi:

You know, I think we are squarely in the PR era, heavy PR era of AI. This story reads as something that's almost a form of both external and internal PR. I have to imagine that the people that got this story out need their, you know, leaders and executives at JP Morgan to know that this stuff is happening and they need the market to know this. You can tell that in the information they're sharing. They sort of. doubled down on saying we're going to drive $1 billion of value from our AI efforts this year. We're on track to hit that number. Honestly, the number seems kind of low given the scale of the business. They also touted they have, you know, 900 data scientists, 600 machine learning engineers, 1000 people involved in data management and the 200 person AI research team. If it sounds like a long list of stuff for me to rattle off, it's because there was a long list of stuff for them to rattle off, which they were very specific about doing. And they want the market to know that they are all AI, tons of AI, billions of dollars of value in AI. It's actually cool to see everybody, you know, getting behind this new technological trend. I feel like it's been a decade since we've seen something like this.

Ethan Aldrich:

Yeah, no, I'm completely on point with you. I think it's a lot of PR hype and, you know, possibly just patent trolling at some points, releasing all these. But, you know, I talked about last week, I think there's some big funds beginning to use LLM in interesting ways. So Connor, you know, do you think there's gonna be anything here to index GPT? I'm not sure if you got a chance to look at the patent or anything like that, but you know, GPTs and financial services, do you think we'll see benefits to consumers, to the business themselves?

Conner Aldrich:

For sure. I think even internally, as we said, it's banned publicly, but there's a lot of shadowy rumors about how internally everyone from executives to middle to everyone throughout the company is using chat GPT on their personal devices or on their phones at home, et cetera. So I think JP Morgan is pushing hard on GPT. They're just trying to be careful and be safe and reduce liability, which is probably why they're trademarking their own indexed GPT here.

Ethan Aldrich:

Yeah, no, that's a good point. And I think that is the main story here is you have these big enterprises and big companies across law and finances that want to use these models. Their employees want to use these models, but they're not going to be using chat GBT. So whether it's PR or they're actually doing something internally, you have a lot of big companies trying to use this tech. So really cool to see JP Morgan trying to get in the folds. Our last story today is Tik Tok taco. So TikTok of course is going through a lot of interesting things here in the US, but overall they are the future of social media and have been for a while now. And they're starting to inject chat bots similar to something like Snapchat. So Farb, I don't know if you got a chance to look at it, but they look like they're trying to analyze videos and answer questions based on that, less so a personal kind of AI like Snapchat. What's your take?

Farb Nivi:

One thing that I thought was interesting from this story is that this wasn't leaked by a user that saw it

Ethan Aldrich:

Mm-hmm.

Farb Nivi:

or a reporter that saw it. It was from an analytics company that sort of saw it happening because they haven't released it in the United States yet. Of course it would be all over the news if they had done that. They've been testing it around the world. This analytics company discovered it, sort of blew their cover on it. Not that obviously since it was out there, they weren't trying to hide it. And it seems like it. It's able to understand the video that you're watching and sort of suggest other things that you may ask it based on the video that you're watching. But you can also use it as a form of search to find more content on the web. And I'm sure they're not staking their claim on this being their ultimate AI vision, that they're probably gonna continue to do things. This is their first step. And again, it's nice to see big players getting in the mix.

Ethan Aldrich:

Yeah, definitely. I think as a, you know, much different angle than something like Snapchat's, you know, my AI within their user experience. And the most interesting thing to me and Connor would love to hear your thoughts. But, you know, we talk about multimodal models and trying to take videos in and understand videos, ask questions based off videos, what better place to do that and collect data than TikTok themselves. What's your take on this? You know, I think there's going to be a lot of interesting things here to search videos.

Conner Aldrich:

I think it's very early days still. Like of course they were testing just in Philippines. They didn't even tweet about it or talk about it at all until after this analyst company discovered it. Even looking at the UX and the UI of it, it feels very raw. It feels very just like a really, it's just a test bit. Even the model internally, we don't know what they're using. I'm assuming it's something Chinese. It's officially, they said nothing from TikTok or from ByteDance, but you can probably assume it's not OpenAI. They want to keep the data in China.

Farb Nivi:

I have to imagine this was not a ton of work for them, given that they've been doing insane amounts of, you know, computer vision on their videos for a long time now. They know what's going on in your videos. Whether

Conner Aldrich:

They are very good understanding of what you have in the videos. That's why, as I talked about in this article too, like the search engine inside of TikTok is better than Google in a lot of ways. If you search for a video, you can find what you want pretty easily and it can show you a lot of what you want very easily. And now having more of a chatbot experience to find that could hurt Google some. So.

Ethan Aldrich:

Absolutely. Yeah, I think the data they're going to have on what people have questions for based on a video is going to be really insane. You know, you watch a video, what else is looming in your head, that's going to help you know what videos should be next as well. So probably another way for them to test some experience, but also improve their core business.

Farb Nivi:

It would be cool if as an influencer, if you could use it to understand trends that are happening, maybe in TikTok, you know, are there as a way to sort of like get early on a trend before it gets super popular.

Conner Aldrich:

Or even like what people are asking about your videos. Yeah.

Ethan Aldrich:

Well amazing, well as always what else are you guys seeing?

Farb Nivi:

I saw a cool example of a guy being able to walk because of AI. It's tough to know exactly what everyone means, given that we are in this PR frenzy of AI and anything you're doing, if it involves a computer, you're better off just saying there's AI involved and letting people figure out how accurate it is. But they said that they were using AI to connect this guy's brain to his spinal cord or something to the machine that's controlling him. subverting his spinal cord that's injured and was able to start walking again. And who knows if this is, you know, LLM type AI. I don't think so. My guess is this is an example of some machine learning that's happening, which is amazing. And I think we are going to see this stuff accelerate rapidly in a few years. I think, you know, in three to five years, it's probably going to be standard to be like, no, you don't need to be in a wheelchair anymore. We've got the tools and the technology to get you, get you walking again. And, you know,

Conner Aldrich:

Yeah, I saw that yesterday. I wasn't sure what they were using, like the AI for in that scenario,

Farb Nivi:

Yeah.

Conner Aldrich:

but it seemed to work very well. Like it made him be able to walk from a wheelchair and then even when it was turned off, it healed him to a point where he could walk with crutches. So very nice.

Farb Nivi:

Yeah, and I think, you know, in the world of medicine and biology and health care, this stuff is going to change the world rapidly. I feel like we're going to be in a scenario where you look back at photos of people in wheelchairs from polio and people dying from things that nobody today even knows what they are because nobody's died from some disease for a long time. I feel like we're entering another era of that, which is, you know, dynamism at its best.

Ethan Aldrich:

Absolutely. Connor.

Conner Aldrich:

Yeah, someone found a pretty interesting note inside the GitHub privacy policy. So if you're using GitHub Copilot, we're all fans of GitHub Copilot, of course, works very well inside of VS Code. I've been playing around with their new chat feature. Um, but if you use GitHub Copilot and you're accused of copyright infringement, GitHub will actually protect you. GitHub will actually defend you in law. Um, I'm assuming because they want to defend all Copilot users. There's some stipulations, of course, you have to have their filter turned on and you have to not be intentionally goading it to give you copyrighted content. Besides that, it seems like GitHub will defend you if you're in queues of copyright infringement, which is very interesting.

Ethan Aldrich:

That's a good play to try to get more businesses to use it to. Yeah, I saw, you know, similar to far story. I saw that this sup, if you don't know what a super bug is, it's these kind of, you know, bugs that are resistant to antibiotic treatment that are starting to become very widespread killing a lot of people every single year. And I saw they used AI to pretty much simulate finding a new antibiotic to kill that super bug. So similar far story, the effect on bio and healthcare. and these uncontrolled environments, you know, this is where AI, whether it's LLMs or just statistical AI comes into play really well running these simulations you couldn't do before. So super cool to see the impact on real people's lives, you know, on the ground.

Farb Nivi:

The only way to fight a super bug is with a super intelligence. So,

Ethan Aldrich:

I like that.

Farb Nivi:

And of course, as we like to stay here on the show, it's, it's super cool.

Conner Aldrich:

very ultra on fighting Thanos vibes.

Farb Nivi:

Yeah, welcome to the super era.

Ethan Aldrich:

The super era. Well thank you guys as always for tuning in to AI Daily and we will see you again next week.

Conner Aldrich:

Peace guys.

0 Comments
Authors
AI Daily