Sam Altman Reveals A.I In 2025 - (2025 AI Predictions)

So we are all very eagerly anticipating 2025 in terms of AI. There are many different themes that we could expect, but one company that all eyes are on is of course OpenAI. And recently Sam Altman had a Twitter thread in which he explained exactly what's going to happen. So in this video I'll dive deep into what you can expect from OpenAI in 2025 and the entire AI space. So one of the things that we are actually probably going to get is of course, surprisingly, some voice mode changes, which one user says that when I talk to advanced voice mode, I wish it had better memory of my previous conversations, both verbal and text, and if it were just a retrieval augmented generation system. He explains that he enjoys conversing and brainstorming due to his commute, but he finds himself having to refresh things in great detail. And Sam Altman says that he really wants to figure this one out. Now, this one is really interesting because this brings us to not just voice mode changes, but actually some key memory changes too. Recently in another video I actually mentioned how in 2025 it's quite likely that memory is going to be one of those things that is solved memory and so it just doesn't forget, which is truly transformative. I mean, you talk about inflection points. Memory is clearly an inflection point because it means that it's worth you investing the time. So it's that capability alone, which I expect to come online in 2025 is going to be truly transformative. So if you're unfamiliar with who that was, that was Mustafa Suleiman. He's currently the head of Microsoft AI. And it's well noted that Microsoft AI work closely with OpenAI in order to produce products and services. So it's quite likely that we're going to be getting some form of OpenAI model that basically has infinite memory. And it was today that we got another member of technical staff from OpenAI Roon actually tweeting a reply to this user. Here you can see someone said rune brother us or the user interface is vastly improved, but $200 per month and no increase in user memory is whack. What is the outlook on that? And he said soon infinite memory. And I'm going to show you guys a clip from a recent video in where I actually spoke about how 2025 is quite likely to have infinite memory for these LLMs, which is going to be a remarkable achievement. Of course, in more voice mode we do get some better detection for voice chat, for example, sometimes when you're thinking about what to respond to the voice mode AI. It will just start yapping on about something random, which is of course not what happens in certain conversations. You're allowed the time to think about your response. Now. One of the biggest questions that majority of people are still asking OpenAI is where is the 4.0 replacement? You can see right here that we did get a very important few questions from this user Leo. He said one of the things that I want is a Strong photo replacement GPT Sora, all seamless, integrated into ChatGPT, perhaps a $50 to $70 a month plan, maybe a middle ground. Of course lond context which we already discussed is basically infinite memory and of course knowledge cutoff aggressive updates please. And he says this is a very good list. Hopefully you'll be quite happy with us over the next year. So I'm guessing that mainly from this, what we're probably going to see is, is, you know, a middle ground monthly pricing plan. I do think there is just such a disparity between $20 and $200 a month. I think it would probably be really useful if OpenAI had something that was progressive in terms of their payments. I know that there are many people that wouldn't want to play just a fixed rate, but would rather pay for how much they talk to the AI. I know that many people who use Claude would rather just Pay an additional $10 or $20, however pay as you go kind of way for when they're talking to Claude AI in certain conversations. And I'm pretty sure a lot of people would like to do the same in ChatGPT for rather than just being rate limited. And of course one of the key things here that we're going to be getting is a strong 4.0 replacement. Now 4.0 is of course the model which is a successor to the GPT4 series. And this is essentially an Omni model. So this model is quite important because it's the base model that a lot of people will use in a variety of different applications, considering most people don't have the necessary requirements in terms of the demanding tasks that O1 requires. So you can see when he responds to this other tweet right here, this user Sully says honestly, a good thinking model, GPT4O kind of sucks now and even GPT4O mini isn't that good. And I think having both is important. And he says here that definitely we are going to do this. And we all do know that OpenAI are working on a new model. In fact, there was a recent article that spoke about how this new model actually is a lot better than the current GPT4O. But the current cost that it takes to use the model, which is might not be worth the marginal gains in terms of improvement. And considering OpenAI still don't have enough compute to serve their users in terms of sora, in terms of 01 Pro mode and of course the O3 model that they're currently working on, I think maybe smaller models like GPT4O mini aren't a priority. But I do think that maybe sometime next year we will get a successor model to GPT4O that's going to be a very remarkable replacement. So that is definitely something that you can be on your list. And I do remember last year that GPT5 was something that people were thinking about. Another thing that we are going to have is of course a potentially liberated mode. This one is actually quite hilarious because this person, okay, and most people don't know who this person is, but this is Pliny the Liberator. Okay, And I say most people, I'm just referring to people outside of the AI space. But Pliny the Liberator is a notorious AI liberator in the sense that he will jailbreak every single large language model that comes out. Anytime you think that a large language model has passed all the safety requirements, sometimes this guy manages a will to jailbreak that large language model, basically meaning that, you know, if you want to ask the AI a question that it usually can't respond, such as asking it to do something that is probably illegal. This guy finds ways to jailbreak the AI every single time. I don't think I've ever seen him fail. And he said that, you know, why don't you lose the guardrails, it's cleaner. And Sam Altman said we're definitely going to need some sort of grown up mode. So I think this one is going to be super fascinating because I think there are a lot of people out there that the main reason they use open source models and all of those unrestricted models is because there is a lot of stuff for adults or grownups that people would want to access and explore creatively with large language models. I'm pretty sure you can all know the direction I'm talking about. But maybe due to policy restrictions or yada yada yada, OpenAI have just decided to keep those restrictions on. So it will be interesting to see if in the new year we do get some kind of mode, which is grown up mode where you can simply ask the AI anything and Maybe they just trust you to ask it questions that are responsible. Another feature we're quite likely to get in 2025 is of course a feature that has basically been released by Google. So this user says, do a Deep research feature like Gemini, but better. And Sam Altman said, okay, okay. Now, if you don't know what Deep Research is from Gemini, I'd explain it to you first, as it's quite like Perplexity's search engine for AI. Now, if you don't know what this is, this is basically where you can just search for AI using the current LLMs of your choice. And it's something that is remarkably effective at allowing you to scour the web for various pieces of information. Now, Deep Research, however, is something that is completely changing the game. Now, this is something that I did a tutorial on recently and trust me, by far this is one of the most useful things in AI. I use this almost every day. Now, you can basically search for anything on the Internet and it researches over a hundred different websites and creates a comprehensive report and gives you that in absolutely incredible detail. So this is something that, if you haven't started using that, I would really use this as soon as possible because it allows you to access so much research, so much information at the click of a button and you can click it, you can come back while your AI agent is off researching information, and then you can use that information to create content, to develop marketing strategies, plans. It's just absolutely incredible. Don't forget to check out that tutorial on my channel. I will leave a link to that. But it is something that OpenAI are basically saying that they are going to be having. So of course OpenAI do have their own search GPT model. You can see right here that this is something that they do have. So search GPT is essentially where you can browse the Internet with ChatGPT. Now, I do think that this is something that is good, but of course there are the depths that a lot of people do want in certain use cases, which is why Gemini's Deep Research feature is so good. So OpenAI stating that they're going to be working on this will be surely interesting because we know that their products are usually top tier when it comes to the user interface and of course the positioning. Now, something rather fascinating that OpenAI also spoke about was the fact that they are going to be doing a hardware play. So you can see here that this user managed to ask Sam Altman for a variety of different things, talking about your own retrieval API, talking about a Video input modality would be nice. Of course that would be nice. Of course you're excited for agents and of course a hardware play. Now, video input modality would be nice. This was something that OpenAI already revealed with GPT4O. I just think maybe it is a little bit expensive in terms of the compute side, which is why we haven't seen it yet. But if we're actually starting to talk about a hardware play, this is something that apparently OpenAI are already on. Now essentially. Joanie, I've has confirmed that he's working on a new device with OpenAI. So this is someone who worked at. I've. This is someone who worked at Apple on the iPhone and now he's working with 10 employees, with people who worked on the iPhone in a 32,000 square foot office in San Francisco on a new generative AI device which is going to be super interesting because this year we've seen many devices come out and some of those have failed. You know, notably the Humane pin did actually not do well. We've seen other AI products as well do really well. For example the Friend tab, there's just a variety of different AI products like the Rabbit R1 device, which is really cool. So it's going to be really interesting to see how these products do succeed and if OpenAI does manage to do them. Now, I don't know if this product is going to come out in 2025 or not. I do know that making a hardware product is really, really hard, but I do think that like, I wouldn't say we're at the limits of where LLMs are, but I do think that now we're at that implementation phase where a lot of these companies right now are just figuring out the best ways to implement stuff, which is why we're getting products like Notebook lm, we're getting search products, we're getting deep research products, we're getting agentic products. And this is why I think there's going to be less research on how to make those models better in terms of the GPT series, but a lot of areas of explosion where we're going to be getting products that utilize these models in really unique ways. I don't know how long this product is going to take. Ideally they'd want it to be really good, but it's going to be really interesting. I mean, a lot can happen in a year, so that is going to be something to look out for as well. Now, of course, one of the biggest things that has been going on in the community is of course OpenAI's video generation model Sora. So this user here uncanny. Harry AI says Sam, please, I would like to see more adherence to image and text prompts, also a reasonable content restriction policy. And he said lots of SORA improvements coming now. This comes off the back of the response from the community when regarding OpenAI's if you aren't familiar with the Sora vs OpenAI, so if you aren't familiar with Veo vs Sora, essentially Google released a video model that was just far, far superior to OpenAI Sora and he said lots of SORA improvements coming now. He also responded to this person, say just make SORA really, really good. And he also said that this is definitely going to be coming now. If you aren't familiar with Veo, maybe you've been living under a rock, maybe you're not familiar with the AI space, but VEO is Google's video model. And in this example here from the Venture twins, you can see that between Sora and Veo, VEO just manages to excel in terms of prompt adherence, the physical motion, and all of the things that you really would want out of a video model. So with that being said, it's quite clear that there is a level above that Google have managed to do. But of course this is going to be something that OpenAI are improving on. Another thing that OpenAI are going to be doing in 2025 is of course a proper agent. One of the articles that have consistently reported on what OpenAI are doing has spoken about how OpenAI are going to be releasing some kind of agent early next year. And apparently this is going to be something that's available in January. Now of course timelines are subject to change based on AI developments and world art events, but of course this is something that will be there in 2025 as that will be the year of AI agents. Now this is of course something that I did make a video on which I will include a segment of there that includes a lot more details on to know. So it says here that OpenAI is preparing to launch a new AI agent, codenamed Operator, that can use a computer to take actions on a person's behalf, such as writing code or booking travel, according to two people familiar with the matter. Now interestingly, one of the things that I found quite fascinating was the fact that this happened in a staff meeting on Wednesday. And the fact is, is that there are plans to release this tool in January as a research preview. You can see that it actually states that it's going to be through the company's application programming interface for developers, said one of the people, which is just API and of course someone was speaking on condition of anonymity to discuss internal matters. Now, what's important to note here is that their leadership plans to release the tool in January is going to be a research preview through an API for developers. And what's important is that this phrasing usually suggests a limited release targeted at researchers and developers, rather than a broad public launch. Typically when we look at a research preview, this does imply that access may be restricted to specific groups like those in a developer or research program, than rather being released to the general public. Which kind of makes sense considering the fact that they are likely to test this out, get feedback, then of course release this to the general public. So for those of you who are hoping for an OpenAI release of agents in January, it's quite likely that this won't be January, but would probably be further down on in the year, whilst January should be slated for for a GPT5 Orion release as they iron out the kinks now. But in response to this right here you can see someone says proper agents and he says Happy 2025. Now one of the things that I think most people are completely missing when it comes to 2025 is the fact that AI is probably going to move even faster. You might be thinking what on earth am I talking about? Benchmarks are going to be smashed. And this is because we take a look at a researcher at OpenAI, Jason Wei, and he says that O3 is very performant. More importantly, progress from 01 to 03 was only three months, which shows how fast progress is going to be in the new paradigm of reinforcement learning on chain of thought, to scale, inference, compute, which is way faster than pre training paradigm of every new model, one to two years. Basically stating that. Look, when we take a look at these improvements, we've seen how 01 preview, then to O1 then, then to O3. Between those models it's only three months, okay, which basically means that before when you'd have GPT3 one year, GPT4 another year and you have six months of collecting data, another couple months of training the model for months and months on end. This kind of model doesn't require that much time in order to be improved. So it's going to be really, really intriguing to see just how many AI improvements we do get. If we got 03 recently, that means there are probably going to be four iteration cycles in terms of model improvements, which means by the end of 2025, it's quite likely that we will be on not 03, but we will probably get 04, 05, 06 and 07, which basically means that if we get to the seventh iteration of the model, we can probably not even fathom how smart that model might be. Of course, that is presuming that everything goes smoothly and there are no interruptions. We know that the world is a very volatile place. Maybe OpenAI might even implode. We've seen some crazy things happen, but one thing is for sure, with many other companies working on thinking models as well. And of course since the age of pre training is over, it means that the AI space is about to move even faster. With that being said, let me know what your wishes are for OpenAI in 2025. Leave your comments down below on what you'd want from OpenAI and Sam Altman. I know that what I would want is probably some really cool AI agents that are able to listen to and do exactly what I say, which of course probably going to come next year. But with that being said, I'm excited for the New year and I'll see you guys in the next video.