63 | ChatGPT introduces a long term memory and insane video creation capabilities, Gemini 1.5 has 1M tokens content window, $349 stylish AI glasses, and many more AI news from the week ending on Feb 16, 2024

0:00Hello, and welcome to Leveraging AI, the podcast that shares practical, ethical ways to improve efficiency, grow your business, and advance your career with AI. This is Isar Maitis, your host, and this is a short weekend news episode of the podcast. This episode is brought to you by the AI Business Transformation course. We've been teaching these courses twice a month. Some of them are open to the public and some of them are booked by specific organizations. So if you're an organization that wants to train your people on AI, or if you're an individual that wants to take a course that will dramatically change your understanding of AI and how to use it within a business context, you can check out the information in the link in the show notes.

0:38The courses have been sold out all the way through the end of March. So the next open course starts in the beginning of April. And we would love to see you join us in those courses. Like every week, a lot of stuff happened. But this week has really been explosive with really incredible news.

1:00so the first big news comes from open ai and chat gpt open ai just announced that chat gpt will have a quote-unquote long-term memory what this means is that you will remember things in between chats so far the way chat gpt works every chat was a standalone universe where it remembered what happened through the chat information you gave it in documents you've uploaded etc but it did not know anything across chats other than the stuff you put in the custom instructions, which you have uploaded and created yourself. If you haven't, then you can go and explore that on how to add custom instructions, then carry across all the different chats.

1:37But this new feature will basically learn you. It will start saving different components of things that you write and different responses, such as your tone and information about your personal life and your company and the style you want to write in and the type of outputs you like and don't like and so on and build a profile of you in order to provide you with better personalized customized results to everything that you're doing. Now, on one hand, this is amazing, right? It means that every one of us will have its own chat GPT that is tailored to our needs. The concern is obviously privacy and what data does it store and how long does it store it and so on.

2:18To resolve that, OpenAI introduced two different things. One is you can see everything that it is saving to its memories and you can control that. You can actually actively ask it to save stuff and you can go to the list and then delete anything you don't want it to remember. So that's number one. Number two, they're also introducing a feature that's called private chat that is going to be like an incognito window in a browser, meaning it will not remember anything from that chat. So if you want to have a completely private conversation where open AI is not going to save any of your information, you'll be able to do that as well.

2:51On the bigger picture, what this tells us is that these AI systems are going to be more and more personalized and customized to our needs, whether it's company needs, tone, style, brand, target persona, audience we're writing to, type of content we like to create, and so on and so forth. This is obviously great news. And that comes with obviously some baggage of giving up on personal information, just like any other platform, you give up information to Google, to Facebook, and so on, and it gives you more personalized content. The same thing is going to happen on probably any AI platform. And this is just the first step in that direction.

3:27This feature is currently not available to everybody. It's currently rolled out to some free users and some paid users, and only then it's going to roll out to everybody else and then businesses. But I assume in the near future, we're going to have access to that. Another huge piece of news from OpenAI this week came with the introduction of Sora, which is their text-to-video AI platform. And it's absolutely mind-blowing. If you are on any social media platform and you haven't seen 20 different videos of examples of this, then you're not following any AI titles in that platform because this has exploded everywhere.

4:01It is absolutely incredible. It literally looks the most realistic video. It runs up to 60 seconds and with very good consistency and very high resolution. It's really like nothing I've ever seen before. The most impressive video creation platform I've seen so far, which is not even available to the public, is Google's Lumiere, but that is a research paper and it showed good improvement over what we know today from platforms like Runway and Pika, but it's not available. On the other hand, Sora from OpenAI has started to being available to beta users, which again means probably within the next couple of months, we're all going to get access to it.

4:44And it's even better than Google's platform. If you haven't seen any of their demos, I will put a link in the show notes to the actual announcement page and you can go and check out their demo videos. It's absolutely amazing. I said that several times on this podcast. 2024 is going to be the year of video and audio on AI. It's already pretty good, but by the end of this year, there's zero doubt in my mind. We'll be able to create videos that will be super realistic, high resolution in probably any length we want. Definitely, we stitch them together in a video editing software, and that has amazing benefits for anybody who is a content creator.

5:24and it has a lot of question marks when it comes to safety, knowing what the truth is, or even people in the video creation industry, like people who are photographers, videographers, camera people, light people, actors, et cetera, because you'll be able to create videos at a very high resolution without any of that equipment and any of these people. Now, I don't know if those two huge pieces of news from Chachi PT about the memory and Sora is what broke the internet, but it definitely broke ChatGPT. I'm just kidding. I don't know if it's related, but ChatGPT experienced massive outages on February 14th, and they lasted for a few hours.

6:07There were tens of thousands of people who shared outages across multiple platforms, whether it's on X or on down detector platform in those 24 hours. OpenAI did not comment on this to explain exactly what happened. But what it shows, it's showing that there's a huge demand for that platform or similar platforms. And then that is obviously not bulletproof from having issues. So the problem that I see with this is that we will become more and more dependent on these AI solutions. And if they are down, it's going to be basically like the internet is down because many of our core processes will depend on that.

6:41I think companies that start building business solutions around AI needs to take this thing into consideration and have some kind of a backup plan that will allow you to continue working if these are down. As an example, don't use just one model. So know how to run on ChatUPT and Claude and Gemini and potentially an open source model like Mistral, and then you are covered if your main path is not working, you can fall back to your second or third different options. Last piece of news about OpenAI. On this coming Tuesday, we are releasing an episode about convergence of technologies in which one of the main topics that we're going to talk about is AI agents.

7:20And one of the most interesting people in the world when it comes to talk about AI agent is Andrei Karpathy. He is a researcher in that field. He's one of the dominant researchers in that field in the world. He's one of the founding members of OpenAI. He then left OpenAI to join Tesla as their director of AI. And he came back to OpenAI in early 2023 to lead exactly that field of agent development. And he just announced that he's departing the company immediately. And he said, there's no hard feelings. It's all a good departure and that he's going to be pursuing personal projects and see what that develops.

7:58So it's not great news for the development of agents within OpenAI, even though I'm sure they have a lot of other really talented people that can continue his path. But I'm really curious to see what he comes up with, because I'm sure it's going to be mind-blowing if it's something that is more than what he can do at OpenAI. And now from the insane amount of news that came from OpenAI to the really exciting news that's coming out of Google. Just last week, Google announced their new advanced model called Gemini Advanced, which is running on their model that's called Gemini Ultra. But let's roll back for a second.

8:32At the end of last year, they finally released their new model called Gemini, which is going to be the infrastructure for everything AI within Google. They've released three different levels, Nano for Mobile, Pro, that back then was the backend of Bard. And they said that they will come up with a new model that is going to be Gemini Ultra, that is going to be their most advanced model that, as I mentioned, they released last week. Well, not a week has passed and Google has announced that they're releasing Gemini 1.5. So everything else that I said so far was version one. Now it's version 1.5. And the first platform that's going to get this is Gemini Pro, which is the thing that used to run BARD, but last week they've announced there's no BARD anymore.

9:15So let's try to create some order and everything I just said. They still run their Gemini Nano on their advanced phones, like Pixel 8 Pro. There is no more BARD. It's now everything is called Gemini. And Gemini browser that used to be called BARD is now running Gemini 1.5. And there's still Gemini Ultra, that is the advanced model that you can pay for 20 bucks a month and be able to use, that is their competitor of GPT-4 Turbo, that still runs version one. Now, what is Gemini 1.5 and why is it exciting? First of all, they're saying that Gemini 1.5, which is again the second tier version, it's not their top version, they're saying it's now as good and in some cases better than the Ultra version 1.

10:01So Gemini Pro 1.5 is in par with Gemini Ultra version 1 using a lot less computing power, meaning it's more scalable and it's going to cost less money and create less damage to the environment while getting to the same results. The other interesting part about Gemini Pro 1.5 is that it can run a context window of up to 1 million tokens. Now, those of you who don't know what tokens are, the whole large language model concepts run on tokens. It's basically fractions of words. It's about 0.75 words is one token. So a million tokens is about 700 to 750 ,000 words. To put things in perspective, the largest context windows we have right now is 128 ,000 tokens in GPT-4 Turbo and 200 ,000 tokens in Cloud 2.

10:57So this is 5X, the second best context window. But more importantly, if you tried these large context window in either Cloud or Chachapiti, you know it's not working great. The more data you put into the context window, the less accurate the results become. And what Google are claiming is that they can run these really, really large context windows with very little reduction in quality. So this is obviously really exciting to anybody who wants to run on really large volumes of data. This is still not available to everybody. I mean, the large context window, and it's only going to be available in their AI Studio and Vertex AI.

11:35Vertex AI is their API connection, and AI Studio is their playground sandbox development environment in which you can test stuff and then turn it into things you can connect through the API. But still, if you want to be able to experiment with that, you'll have to wait probably a little longer, and then I assume everybody will have access to it. through AI Studio, which is a really cool tool that you can go and create and test different things. You can even create stuff like GPTs that exist in OpenAI. You can create similar things that they're calling them chat prompts that you can build within Google's AI Studio.

12:07One more thing I want to say about this is if the free version, which is Gemini Pro 1.5, is as good as their paid version, Gemini Ultra version one. There is zero doubt in my mind that Gemini Ultra 1.5 is coming out in the very near future. Otherwise, there would be no point for people to pay for the Ultra version. So this is obviously just another step in this crazy race between these most advanced model companies on what's coming next. I can't wait to see what's going to be coming out of OpenAI with GPT-5 or 4.5. There's rumors about both, but it will be very interesting to see how good that model is.

12:49Speaking of GPT-5, another just interesting geeky fact, one of the Microsoft executives that has been testing GPT-5 is saying that there's a very high likelihood it will be able to do things that we were not able to do before as humanity. One of the cool examples that he gave is that there's a very old written language called Linear A, which has been undecoded since 3 ,000 years ago. We have examples of this language, but nobody was able to identify it. And he's claiming that GPT-5 has a very good shot of actually understanding how this language works, because it's really amazing capability to make sense in sequences and codes and so on.

13:33While this is a very geeky, maybe not mainstream usage, It just shows you the power of these tools to make sense in data that we couldn't make sense in before that is available to literally anybody in the world. So it's the ultimate democratization of amazing power. Another very interesting piece of news this week came from another huge player in the AI world, which is NVIDIA. And NVIDIA, we all know them as a company that makes the GPUs that run basically everything AI right now. That's why their valuation went through the sky in the last two years, because everybody's buying their GPUs. This week, NVIDIA announced that they're releasing what they call RTX Personal AI Chatbot.

14:12It's a chatbot that runs locally on RTX 30 and 40 GPUs. So if you have a computer, a PC, and you can buy an RTX GPU, install it, and then you can run a local chatbot that will have access to all your documents, as well as different links like YouTube videos that you're watching, and you can quickly search and summarize information across all the information you have on your computer. The huge benefit of that is obviously that you're running it locally and you're not relying on cloud services, which means A, it's going to be faster, and B, you're not sharing data with anybody in order to be able to do that.

14:47That being said, it still has a lot of bugs from people who started using it. It takes a huge amount of RAM and it takes even a bigger amount of storage on your hard drive, but if you have enough RAM and you have enough storage and you're running a PC and you can run these GPUs, it's an interesting tool to try out. That being said, it does not have any context, meaning it does not remember anything between questions, and it has still some accuracy problems, but I think the direction is very, very clear. I don't see that as a long-term viable solution. I have very little doubt that the next operating system, both from Microsoft and from Apple will integrate AI capabilities into the actual operating system of our computers, meaning external tools like this are not going to be necessary because the operating system itself will most likely be able to know everything that's on the computer and respond to that through some kind of an AI chat or even beyond that, maybe voice and video and so on.

15:43So why is this important? It's important because it's showing the transition of NVIDIA from just being a hardware provider to being AI software provider, meaning drifting into the worlds of Microsoft and Google and OpenAI and Anthropic and so on. By the way, at the same time, the opposite process is happening. So Microsoft already announced that they're developing AI chips. Google has announced that they're developing AI chips for their phones. Apple announced the same thing, that the next set of iPhones is going to have AI chips built on board. And obviously, there's a huge announcement of Sam Altman that's saying he's trying to raise$7 trillion in order to build AI chips that'll be able to run the AI of the rest of the world.

16:23And as a follow-up to that, by the way, after Sam came up with his announcement of trying to raise$7 trillion to build these chips, NVIDIA CEO Jensen Hung replied by saying that this$7 trillion could buy basically every GPU in the future of the planet. But he's saying that because of advancement in that world and because of economies of scale, he thinks that the world will not need more than$1 trillion to maybe$2 trillion to provide it all the AI computing power it needs. Now, he obviously knows one of two things about developing AI-related hardware because they are the leading manufacturers by a very big spread.

17:02I would assume he's probably more accurate with his number than Sam Altman, but there's definitely a race going on of hardware manufacturers going into the software world and vice versa. So from all of these huge piece of news to one last cool and fun piece of news. There's been a lot of hype in the last couple of weeks from Apple's Vision Pro. Everybody's talking about them. Everybody who has enough money or is a real geek has bought one and is sharing what he's finding and everybody finds it amazing. I assume it's incredible from everything I've seen. It's a game changer. The only problem with it, it's heavy and it's big ski goggles that you're probably not going to wear walking out in the street.

17:42Well, a startup called Brilliant Labs just came out with what they call frame glasses. These are AI driven, really cool looking regular glasses that you can wear anywhere and they can project on a relatively small field of view, but still project different information, AI based on the glasses. It has a microphone that it can listen to getting different commands through OpenAI Whisper and it can identify objects through the camera and it can provide various information and you can communicate with this through your voice and get voice answers and information projected on the screen. These glasses look like pretty cool glasses.

18:19They're round looking, look like John Lennon kind of glasses. And so they're stylish and small and you can wear them anywhere and you can use them to connect with AI capabilities through a normal day-to-day interface, which I find really attractive. They weigh only 40 grams and they cost only$349. So this is not much more expensive than just high-end sunglasses. It's obviously not similar in any way to Apple's Vision Pro, but it shows you where this world is going, meaning definitely in the future, I see small glasses with a form factor of normal sunglasses that will be able to provide us similar capabilities to what Vision Pro can do right now.

19:00And this is definitely something that I see more and more people wearing in the next few years. If you're enjoying this podcast, please subscribe, rate the podcast, give us whatever star rating you feel we deserve in your favorite podcasting platform and share it with other people that you think can benefit from this. This is your way to help additional people learn more about AI and it also helps us and I would really appreciate that. That's it for this week. As I mentioned on Tuesday, we're coming up with a unique episode that is going to talk about the convergence of various technologies, including AI, and where is that leading the world we live in and how much it is going to change.

19:38That's it for now and have an amazing weekend.