Anticipating GPT-5: Release Date, Rumors, and Expected Features

0:00What can 160 years of experience teach you about the future? When it comes to protecting what matters, Pacific Life provides life insurance, retirement income, and employee benefits for people and businesses building a more confident tomorrow. Strategies rooted in strength and backed by experience. Ask a financial professional how Pacific Life can help you today. Pacific Life Insurance Company, Omaha, Nebraska, and in New York. Pacific Life and Annuity, Phoenix, Arizona. after chat gpt was launched it was only a matter of months later before gpt4 came out and we saw some really impressive improvements in ai and in this chatbot in general and this has begged the question a lot of people have asked when is gpt5 coming out so today on the podcast we're going to dive into trying to answer that question based off of some recent comments by the ceo of open ai Sam Altman and also we're going to talk about what we think we will see in GPT-5 what its improvements will be over GPT-4.

1:01So the first thing to know is that it has been over half of a year since chat GPT was initially launched and since that time we've seen some substantial upgrades. GPT-4 specifically brought a bunch of really powerful upgrades that GPT-3.5 did not have and specifically that was logic and reasoning that was a really big one that we saw in addition to gpt4 being able to take a much larger amount of text as an input in order to generate its responses now the one thing i do want to say with gpt4 is that while a lot of people are sitting here you know wondering when gpt5 is going to be coming out i think what a lot of people don't realize is the fact that we haven't even seen a lot of the really headline GPT-4 announced features yet.

1:50So when GPT-4 first came out, you know, they announced a whole bunch of really interesting things that it would be able to do and that it has the capability of doing, but they haven't actually rolled that out. So I think something to think about is the fact that before, you know, GPT-5 even comes out, we are going to be seeing, we're going to see a lot of really big improvements still coming for GPT-4 in the near future. Now, as far as the timeline on GPT-5 goes, I think OpenAI has began really focusing on slow rolling this entire thing, GPT-5. You know, there was a recent article or a recent letter that got passed around to a bunch of, you know, major tech and AI leaders, and they all signed this letter saying that there needed to be a six-month pause on training, you know, a superior AI version to GPT-4.

2:40And of course, you know, OpenAI didn't specifically say, okay, we're going to listen to your pods. But I think that they could see that the sentiment in the field and also in Congress, you know, regulatory-wise and in government, was having the same feeling. And so they really did back off when Sam Altman went to Congress. He said, we're not training GPT-5 right now. We're not planning on it. And this actually contradicted a report that went out a little bit earlier in that year, which came out from Morgan Stanley, which said they believed back in April that ChatGPT had been training GPT-5, and they believed that because they saw that there was 25 ,000 GPUs being trained, which cost about $225 million worth of NVIDIA's hardware.

3:34And so, you know, after seeing that, they saw some receipts. And so they put out this thing saying they believed GPT-5 was going to be coming out, and they were kind of talking about that. And very quickly after that, we saw this letter calling for a pause on, you know, training, and OpenAI all of a sudden starts saying, no, we're not training anything. So what was that that they saw? Perhaps that was, you know, GPT-4 getting better. This is something I talked about way back when this letter first came out. You know, I said this is, it's like, it's essentially not very useful because OpenAI doesn't have to come out and say, look, we're training GPT-5.

4:06They just say, oh no, this is just GPT-4.1, 4.2, 4.3. You know, they can make as many upgrades to GPT-4 as they want without people really complaining that much. And I think that they actually have done that because they have put out updates to GPT-4 since that's launched, that made it better at reasoning, that made it better at a handful of other tasks at math and other areas. And they've integrated plugins, which have given it a lot of those increased capabilities. So despite that, I think it still is already going to be getting quite a bit better, which means that, you know, it like, did they just pivot and stop saying they were training GPT-5 when they realized people were scared instead of hyped about their technology?

4:47I don't know. But in any case, just recently, we had Sam Altman, he was talking, he was an India talking about AI just this month and he actually confirmed that OpenAI wasn't planning to train a new model for the foreseeable future and he said we're working on the new ideas that we think we need for it but we are certainly not close to the start. I think again he's slow rolling this and something that is really interesting that I've talked about with a lot of really smart people in the AI space, he said we're working on the new ideas that we think we may need for it And this is because what a lot of people are saying is that the transformer model, which is, you know, the model that is used to train these AIs, is perhaps has maxed out at its capabilities, what it's really capable of doing with a large language kind of all encompassing model like a chat GPT.

5:39and we may need to start looking at other ways of training AI models because as we know, the transformer model is not the only framework for training AI models. But what's interesting is, you know, there's been a handful of different frameworks looked at by Google and by Facebook and others. But the thing is right now, OpenAI has raised over$10 billion from Microsoft and they've plugged a lot of those resources into the transformer model framework. And so that is the main one that we're focusing on. That's the main one that we see innovation within. And when you have a framework that has so much money plugged into it, it's really hard to move over to other frameworks or other ways of training AI models because we already have so many resources, so much development down this one specific, I guess, pathway.

6:24But there are a lot of different pathways. So it'll be interesting to see if OpenAI continues. I would assume because of the financial investment they've already put there, I would assume they're going to continue down the transformer model framework for training AIs. but there are other options. And so it's interesting, you know, they're talking about the new ideas that they think they need in order to actually build a GPT-5 and what that might look like. So the next thing I wanted to talk about was, you know, what are the future potential features that we'll see in GPT-5? Like, you know, we talk about it getting better, but what does better really mean?

6:58Obviously, perhaps more accurate, perhaps you ask it less and it gives you more. So there's those kind of like conceptual things. But what does that really look like? So here's some things that I believe a GPT-5 could have that would make it a lot more powerful and what GPT-4 does not have. So number one, of course, I think improved context understanding. One limitation of GPT-4 and any AI model right now is just that they really do have like a very finite window of context, which I think limits them. And the number of, you know, it essentially like if you've ever had a really long conversation with chat gpt it starts at some point to forget some of your earlier messages and responses and i believe that's because the way they have this thing set up is it's just so expensive to use you know really far off previous responses because all of a sudden they would have to send your entire chat as a prompt to the uh to the system every time and that gets incredibly expensive and so eventually i think they just start saying okay They've just send the last X amount of tokens or X amount of words and the newest prompt and then respond based off of that context.

8:08So it sort of starts forgetting things that talked about earlier. So I think GPT-5, one of the biggest things it could do is really to improve that context window. So it could essentially understand an infinite or take an infinite amount of content in. And now, of course, it's like, well, gee, that seems like a lot of computational power and a lot of, you know, perhaps waste compared to what, you know, GPT-4 is doing now. But I think the possibility here, I don't think this is very, I don't think this is too far off because we're seeing a lot of models that are using much less compute. They're becoming much more refined.

8:42They're becoming much more efficient. and so I think that we actually can get to models where we're using less compute but still able to give them more. Another thing that I think we're going to see is kind of a multi-model capability. So GPT-4 right now is just text-based but in the future I think we're going to start seeing the ability for us to upload images, videos, and sounds in addition to the text that we currently can do. And you know earlier I was talking about the fact that there are a bunch of announced GPT-4 features that are not even currently integrated. One of those is images. Like, you know, when they launched GPT-4, they said, like, we have the capability, you can upload an image, it can look at the image and understand that, and you can use that as part of your prompt.

9:25But we still don't have that in GPT-4. And so that will be a powerful way. And Google recently started experimenting in their most recent Palm model with having kind of this multi-modal, these multi-modal capabilities where you can upload images, videos, sounds, and other things. I think the other thing that we will see is a more efficient training. So as these models become larger and larger, they haven't even confirmed the size of GPT-4. Some people say it's a trillion. But essentially, as these models grow, the resources required to train them is also going to grow a lot. So I believe that we're going to see a lot of new techniques that are going to be developed to make training these a lot more efficient.

10:04And this is going to enable larger, more powerful models, like I was saying before. Another thing I believe that we're going to see is enhanced fine tuning capabilities. So I believe that future versions are going to offer a lot more sophisticated ways to customize the AI's behavior and allow developers to really fine tune the model's responses more precisely. Another thing I believe they will do is just they'll be better at interacting with external databases. So GPT-4 already essentially has a knowledge cutoff, right? And this is something that Google Bar does not have because it has access to the full internet.

10:38so it knows things that are happening in real time. And of course, you do have the ability to use web browsing with ChatGPT, GPT-4 if you have premium. And actually, I believe it's with the free version as well. And you're able to see a little bit more stuff in real time if you enable that plugin. But I think that'll be something more baked into the system similar to what Bard has, where you can see real-time content live as you're searching and chatting on there. And I think the other thing I expect to see and a GPT-5 would be definitely advanced reasoning and logic. I think that future iterations are going to focus a lot more on improving the reasoning and logic capabilities of the AI.

11:18We've seen a number of updates to ChatGPT that address this, but I still have seen a few flaws and loopholes that I believe they can tighten up and make quite significantly better. Another area I think that we're going to see is just dynamic and interactive learning. So instead of relying solely on pre-training and fine-tuning on static data sets. I think future models might incorporate some form of online learning where they can actually learn and adapt from their ongoing interactions. And I think this is going to be something really, really interesting. You know, we talk about auto GPT that you can, you know, give it a task, it can go out on the internet, it can learn things and do its own research to help you accomplish that task.

11:59I think we may at some point see some of those same features and possibilities in something like GPT-5. And I think that'll be really, really interesting and really powerful. So it's up to be seen what exactly is incorporated in GPT-5 when this comes out. It may be another year or another two before we see that. But definitely in the meantime, we're going to see a lot of updates to GPT-4 where we have the plugin store. And we're going to see a lot of improvements on the platform overall. So it's not like we're at a standstill with the advancement of this technology till GPT-5 comes out But I believe they have a high bar to set they really want to blow people away and impress them with GPT-5 And so I think that they're going to be taking a completely different approach They're really going to try to revolutionize this space with that And that might be quite a powerful tool But in the meantime, we'll have to follow up on all of the new updates that GPT-4 is making Because there's still a lot of stuff they've announced that isn't even out yet