AI News: AI Takes Control in January!

Here's everything that happened in the world of AI this week, so let's just get right into it, starting with the fact that Sam Altman is predicting AGI in 2025. In fact, here's the exact snippet that everybody is talking about when they claim that Sam is talking about AGI in 2025: "What are you excited about in 2025? What's to come? AGI." Yeah, excited for that.

So, Gary Tan, the CEO of Y Combinator, asked Sam Altman what he was excited about for 2025, and Sam's response was AGI, sort of implying that he thinks we'll have AGI in 2025. Dario Amodei, the CEO of Anthropic, recently went on Lex Fridman's podcast, where they chatted for 5 hours. In that conversation, he was asked when he thought we would see AGI, and here was his response:

"What do you use the timeline to where we achieve AGI, AKA powerful AI, AKA super useful AI, smarter than a Nobel Prize winner in every relevant discipline? And all the things we've said every time. If I say 2026 or 2027, there will be like a zillion people on Twitter who will be like, 'AI Co said 2026,' and it'll be repeated for like the next two years that this is definitely when I think it's going to happen. Whoever's exerting these clips will crop out the thing I just said and only say the thing I'm about to say. If you extrapolate the curves that we've had so far, right? If you say, 'Well, I don't know, we're starting to get to like PhD level,' and last year we were at undergraduate level, and the year before we were at like the level of a high school student. Again, you can quibble with at what tasks and for what. We're still missing modalities, but those are being added, like computer use was added, like image generation has been added. This is totally unscientific, but if you just kind of eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there by 2026 or 2027. Again, lots of things could derail it. We could run out of data, might not be able to scale clusters as much as we want, maybe Taiwan gets blown up or something, and you know, then we can't produce as many GPUs as we want. There are all kinds of things that could derail the whole process, so I don't fully believe the straight-line extrapolation. But if you believe the straight-line extrapolation, we'll get there in 2026 or 2027."

So, obviously, I'm not going to crop out the portion that he mentioned that people are going to crop out, so I left his disclaimer in. I did edit it for brevity, but Dario's prediction is seemingly 2026 or 2027.

Now, the big problem that I see with these predictions, both from Sam and from Dario, is that the definition of AGI is a struggle. I think there's going to be a lot of debate about when we actually hit AGI. I think we might see OpenAI sometime in 2025 saying we've achieved AGI, but then, you know, Google and Anthropic might come out and say, "Well, that's their definition of AGI; we don't see that as AGI yet." Right? Like, there isn't any sort of clearly defined definition of AGI that all of these companies and all of these people align on. We may blow past AGI and not even realize it, or we might get to a point where companies claim they've achieved AGI, but other companies sort of disagree with their definition of AGI.

So, I think this whole "when do we get AGI" discussion is a bit of a tricky one for those reasons because there's no agreed-upon definition of exactly what AGI is, and it's going to be hard to define it once we've actually achieved it. It's going to be harder to recognize, and I think most people realize that.

On the flip side of that discussion, there's been a lot of news circulating around how the rate of AI improvements has been slowing down. There was an article on The Information; Bloomberg also put out an article with a similar tone to it that AI progress has really been slowing down. What they're actually referring to in these articles is that they're running into sort of diminishing returns when it comes to actually training the AI models. Up front, they've kind of gotten to a point where just shoving more data, going and scraping more websites that they hadn't scraped yet, and shoving that data into the new models isn't enough to really improve the output. That's sort of the issue they've run into.

But companies like OpenAI are actually finding new ways to make these models smarter, and they're not on that training side. They're not on this side where they're stuffing more data into it and training it on all of this information; it's more on the inference side. So, inference is essentially the point when you ask it your prompt, and then it does its thinking and then sort of gives you a response. That's the inference phase. They're figuring out ways to make that portion of the AI even smarter.

We saw it with OpenAI's 01, and the version of 01 that we've seen is 01 preview. It was like a model checkpoint; it wasn't even the full 01 yet. But they're figuring out how to make it smarter when you ask it the question. They're figuring out how to make it think through the response a little bit better before giving you an answer. A lot of the progress we're seeing now is on that inference side and less on the training side.

So, articles like this from The Information and from Bloomberg and some of these companies saying that there's a big slowdown in the rate of improvement of these AI models—well, maybe there's a slowdown on that training portion, but they are finding new and better ways to improve these AI models during other phases of using these models, like during the inference phase.

Since I brought up OpenAI, let's talk about this because this was big news that was circulating this week as well: that OpenAI is about to launch an AI agent that can actually use your computer. Now, we already saw that from Anthropic a couple of weeks ago. We got computer use out of Anthropic. It's still a little buggy; it still runs into rate limit issues. It's not quite what everybody hopes it would be yet, but it is a beta to kind of show where they're at currently.

Well, it sounds like OpenAI has sort of had this technology for a while. They've been able to do this already, but they're getting ready to release it. The word that's sort of circulating right now is that Sam Altman had an all-hands meeting, and during this all-hands meeting, he claimed that they're going to release this tool use feature, this agentic AI for OpenAI and ChatGPT, likely in January. So, we're only a couple of months out from getting this tool use functionality from OpenAI.

As well, a couple of weeks ago, OpenAI rolled out their desktop app. The Mac version came out a few months ago; the Windows version came out a couple of weeks ago. Those versions were for Plus and Enterprise and the paid users of ChatGPT. Well, this week, ChatGPT made these desktop apps available to everybody. So, even if you're on the free plan, you should be able to use the new desktop versions of ChatGPT. If you want, just go to openai.com/chatgpt/desktop, and you can download the macOS version or the Windows version. It gives you a nice little handy shortcut where you can quickly type in a prompt and get a response.

Along with this new rollout of this new version of the app for everybody, if you're on a Mac, this new app can actually read some of the apps on your computer—not all of them. There are some specific apps we can see in this example screenshot here: Xcode, TextEdit, iTerm2, Terminal, and Code. You can enable those, and the ChatGPT app can actually watch that on your screen and give you feedback.

Now, it's not going to interact with it; it's not going to type in new code into Xcode on your behalf, but it can see what is in Xcode and then tell you what to do next. Here's a little demo video that's been circulating, showing off this new feature. I'm not going to show the whole video, but let's take a peek at what it looks like when he's working with Xcode.

"Well, now with this new integration that I enabled, I can click the Xcode button here, and ChatGPT can immediately see the Swift code I'm working on. Check out this example. So, OpenAI first created this entire app from scratch to track the ISS and astronauts in space in real time. But say I want to add a new feature to this app. I can simply write, 'Add a new screen in the center with the live stream.' Now ChatGPT has the context about my existing Swift code and starts suggesting the changes. That's done, so I can just jump back in and update the code. Let's build the app. Let's go in. Great! It's pretty cool; the live stream is right there with even a fancy icon for it."

So, you can see it saw Xcode, it saw his terminal, and it suggested the code changes. He still had to copy and paste them, so it's not going to take control of those apps for you yet, but the version we might see in January that I just mentioned might be able to handle that stuff on your behalf without you even having to copy and paste.

In the final bit of OpenAI news here, Greg Brockman, the president of OpenAI and one of the co-founders, returned. If you remember, a few months ago, a whole bunch of people were sort of leaving OpenAI. Greg Brockman said, "I'm taking an extended leave; I'm taking a sabbatical," and everybody thought, "Oh, Greg Brockman is out too. All of these other people are leaving; Greg's leaving; he must be gone as well." Well, it turns out it was really just a break; it was just a sabbatical, a vacation, and now he's back and ready to get to work right as things seem to be getting exciting again.

But at the moment, as of the recording of this video, OpenAI does not have the top large language model of the moment. In fact, that accolade currently goes to Google. There was a new model that was released on November 14th, which happens to be the day that I'm recording this video, called Gemini EXP 1114. According to the Chatbot Arena leaderboard, it is currently winning the chatbot race according to users.

Now, as a quick refresher, if you're not familiar with the Chatbot Arena leaderboard, there's a little arena here. You could give it a prompt like, "Which company makes the smartest LLMs?" and you can see here that it blindly gives me a response here with two models. We've got Model A and Model B. I don't know which model is which. At the end, I can say A is better, B is better, it's a tie, or both are bad. You can see that this model on the left gave a lot more detailed reply with a lot more formatting here. Model A gave less of a reply, and based on a quick review, I can fairly confidently say that I think A is better. So, I click A is better, and then down here we can see that A is actually labeled as Anonymous Chatbot, while Model B is labeled as C4 AI EXP 32B.

Now, if we go back to our leaderboard here, this is how those rankings are found. A whole bunch of people are going to this site, giving it a prompt, picking either A or B. This happens so many times every single day, and that's what affects the leaderboard. This is as voted on blindly by users, and at the moment, the current model of Gemini, this Gemini EXP 1114, is the model that people are voting has given the best response to them.

While we're on the topic of Google, they just released a Gemini app for the iPhone. We can see a little bit of a peek here. You can actually take pictures with your phone and then ask it to quiz you on that image. Now, if you have an iPhone but you don't have one of the newer iPhones, you don't have the Siri with ChatGPT in it, well, maybe Gemini is a good option. You can download the Gemini app and use this as an alternative. I mean, we also have ChatGPT and Anthropic and Perplexity and all these other tools that have apps as well, but now Gemini is another one that you could add into the mix of your iPhone apps that talk to AI for you.

Since we're talking about Google, let's talk about YouTube real quick. YouTube is testing music remixes made by AI. It says if you're a creator in the experiment group, you can select an eligible song, describe how you want to restyle it, then generate a unique 30-second soundtrack to use in your short. These restyled soundtracks will have clear attribution to the original song, to the short itself, and the short's audio pivot page, and will also clearly indicate that the track was restyled with AI.

Now, I don't have access to this feature yet, so I can't show it off, but essentially it sounds like you could take somebody else's song, and it will make a new version of it for your short to use in the background, and it'll properly give them attribution, and you don't have to worry about copyright and things like that.

As I've mentioned before, I'm a big believer that everybody should have a website. Everyone should have some online real estate of their own to show off their writing or their images or their videos, and that's why for today's video, I partnered with Hostinger. They make it so easy to set up a website because you can literally let AI do it for you. Right now, they're running a special Black Friday deal with 85% off. You can literally get your website for under two bucks a month. Go to the URL in the description, click on "Claim Deal." You're going to want to get the Business Website Builder because this is the one that has all of the really cool AI features. Go ahead and choose that plan, choose the length of time that you want to sign up for the website for. Personally, I always like to just get the longest one possible because I don't want to have to think about monthly payments.

Click this little link that says "Have a coupon code." Notice the price right now. If I enter the coupon code "Matt Wolf" and click apply, it gets even cheaper. Once inside of your Hostinger account, go ahead and click on "Websites," click on this little "Add Website" button, and then select the Hostinger Website Builder. You'll notice it's an AI-powered drag-and-drop website builder. Now, I simply answer a few questions, and I'll have a website built in minutes.

Let's pretend I'm building a drone review site. I'll call it "Matt's Drone Reviews," and for the description, I'm lazy; I like to hit this microphone button and just describe what I want my website to be. This is a website where I share reviews and news about the latest drone technology for consumer drones. Now, go ahead and click "Create Website" and let the AI do its magic for me. A few seconds later, I have a full rough draft of a website complete with blog posts and drone images already on it. Now, I can customize it by simply clicking this "Edit Site" button in the top right.

Check this out: I can drag and drop this however I want. Maybe I want this top section to be a little bit smaller. Let's just go ahead and shrink it up like that. Let's jump into one of these blog posts. Let's let AI write a drone review for me. Give it a quick description and tell it to create the text. A few seconds later, I have a rough draft of a blog post that I can start to build off of, and this is just scratching the surface of all of the cool AI features they have inside of this platform.

Of course, it's got all the other website building features that you'd expect from a website building platform. You can check it out over at Hostinger.com/SLMTWolf. Get in as soon as you can because they got that Black Friday deal going, and don't forget to use the coupon code "Matt Wolf." Thank you so much to Hostinger for sponsoring this video.

The Anthropic console added a new feature that improves prompts for you. If we head over to console.anthropic.com, we've got this "Improve an Existing Prompt" button. We can click on this; I can give it a simple prompt like "Help me solve this math problem," and then in brackets put "math problem" so that it knows that that's the variable that will change every time I use this problem. Then it says, "What would you like to improve?" I could say something like, "I want to ensure that Claude thinks this through and always comes to the correct conclusion," and then click "Improve Prompt."

It's thinking through right now, analyzing my prompt, figuring out how to improve the prompt. I'm actually watching it in real time go through this process of improving the prompt. Each of these steps is going through by itself, and at the end, we have a much more detailed prompt here: "You're an expert mathematics tutor tasked with helping a student solve a math problem. Your goal is to provide a clear step-by-step solution that not only answers the question but also explains the reasoning behind each step. Here are the math problems you need to solve. Please follow the steps to solve the problem: analyze, plan your approach, execute."

So, it's basically forcing it to do this chain of thought, step-by-step, break down why you did it that way. Now, I can open it in the workbench. Here's my prompt right here, where my variable is; that's where I would actually give it a math problem. So, pretty handy feature. If you don't want to be a prompt engineer, you don't want to learn about prompt engineering, you can let Claude be the prompt engineer for you.

Now, Anthropic probably won't be too happy about this, but any prompt that this generates, you can also go use in ChatGPT. You can go use in Gemini; you can go use in Llama. It doesn't matter; the prompt will probably be a lot more effective than whatever you were going to originally write.

If you've wanted to use Grok by XAI but you haven't wanted to pay for a X premium account, well, it sounds like they're testing out free versions of it in certain countries. I don't believe they're testing a free version in the US, but according to this TechCrunch article, they were able to confirm that X is testing access to Grok for free users in New Zealand. So, if all goes well there, maybe we'll start to see free versions of Grok roll out into more locations.

I mentioned several weeks ago that Perplexity was going to start adding ads into their platform, so when you ask questions, there could be ads within the responses. Well, it sounds like this week they're actually starting to roll out that feature. We can see a little screenshot here of what that might look like. You ask the question, it gives an answer, and then down below the answer, they have like a related question section where you could sort of follow on the discussion, and it looks like there might even be a sponsored related question.

The Washington Post this week rolled out its own AI bot to answer questions, I'm guessing about what's going on in the news. We can see it's called "Ask the Post AI," and it's completely free to use right now. Now, the Washington Post, I think, has a subscription model, but you don't need to be a subscriber to use this at the moment. It has some suggested questions down here. Let's go ahead and see what it says for "How can I stop junk mail?" It looks like it uses articles from The Washington Post to respond to your questions, and it's a pretty short response, but it looks like there are three articles that it pulled from here. Not something I see myself using a ton, but it's free, and if this interests you, you can check it out at washingtonpost.com/askthepostAI.

The AI tool AlphaFold 3 was actually released this week as open source. It was put up on GitHub. The creators of AlphaFold 3 actually won the Nobel Prize this year for this work, which is an AI that helps model protein structures. With this being more open now, theoretically, it could lead to a lot more drug discoveries and a better understanding of biology.

Essentially, 11 Labs rolled out a new feature this week called "Projects," which appears to be a sort of organizational tool inside of 11 Labs for if you're converting large books into audio versions. It seems like it'll help you break it down into much more manageable chunks to create the audio from it. It also looks like you can now select a word or a part of a sentence and regenerate only that section. So, if you don't like the way a single sentence sounds, you can have it redo it until you like it, and a whole bunch of other quality-of-life upgrades, like the ability to customize sentence delivery, apply voice changes across multiple paragraphs, improved audio playback when you're listening back to it, and things like that.

Since we're talking about 11 Labs, Jerry Garcia is now one of the voices that's available to use inside of 11 Labs. This was approved by Jerry Garcia's estate, and it looks like it's in the 11 Labs reader app, where it can read PDFs and articles and things like that to you. If you remember, in a previous video, I talked about how Judy Garland, James Dean, Burt Reynolds, and Sir Lawrence Olivier were all people that lent their voice to 11 Labs, and they can sort of read these articles to you. Well, now Jerry Garcia is one of those as well. It could be a lot cooler if it would sing it to me, but still cool nonetheless.

Nathaniel Ruiz here showed off some new research they've been working on over at Google. Nathaniel is actually the original creator of DreamBooth, a tool that I've made several tutorials about and how I actually trained my own face into AI originally. Well, this new research is a sort of re-angling of videos. The tweet here says, "Recapture is a method to generate new versions of a user-provided video with customized camera trajectories." We can see some examples here, like this original video of a parrot, and then this recapture video with new angles. This person blowing a kiss has new angles; this elephant can show new angles. So, if you have an original video, you can upload this original video, and this research can help you make new angles on that original video.

Now, this isn't something that we have access to; it doesn't look like they've made the code available on a place like GitHub or anything. We do have a research paper here on archive.org that goes into more detail about what they've been building, but if you know me and you've watched my channel for a while, you know that I really get excited about stuff when I have the chance to play with it. I'm excited to see that one roll out and actually get my hands on it at some point.

Here's some more research that I came across this week that I thought was really interesting called AutoV VFX. You can upload a video scene in like 3D space, like this table here they uploaded, and you can see there's a vase on the table. They gave it a prompt of "Fireball breaks vase." You can see down here, and if I scroll across, you can see what this looks like. But there's the fireball hitting the vase, or here's the vase actually melting, or here's an animated bird flying around the vase. It looks kind of like a cross between the pigs and the birds in Angry Birds. Maybe a basketball is dropping on this 3D scene.

So, you upload the 3D scene, and then you can give it a text prompt of what sort of additional visual effects you want in that scene. This seems pretty dang cool. This is something I can't wait to play with. I love workflows like this. We have all sorts of examples like this office here, which is empty, but then here's Iron Man emitting smoke in that office, or here's like a little studio apartment or something. I slide this over, and now there's a campfire in the middle of the studio apartment. Some pretty cool research here. I really like this, like a car driving down the road. Now it's hitting a Mercedes G Wagon.

I'll leave the link to this research in the description below, as I do with all the links of the things I talk about. It looks like, let's see if I go to GitHub here, there is actually code on GitHub for this one. So, if you know what you're doing, you can actually set this one up. Personally, I don't know what I'm doing on this; I might have to refer to an expert and get somebody to walk me through how to actually install this and use it myself if my computer is even capable of it.

Maybe something we can play around with more in a future video. The company InVideo AI, actually one of the longer-term sponsors on this channel, just rolled out their new V3 model of InVideo. Now, if you remember, version two of InVideo went and found stock footage that already existed. It listened to your voiceover or took your text and then put the relevant stock video over it. Version 3.0 generates the video, so it's not going in finding stock video; it's actually generating completely new video based on your script or whatever you uploaded for the video. I do have more videos coming out about InVideo 3.0, so I'll be talking about that in future videos, but that update rolled out this week as well.

All right, I'm going to move into some rapid-fire news. Most of what I'm going to talk about from here on out is sort of interesting stuff, but it doesn't require much more in-depth discussion on it. I'm just going to kind of share it quickly. Apple's next device is an AI wall tablet. They actually have a picture here of like an Amazon Echo Hub. Apparently, Apple is going to be building something very similar to this, and it's going to have AI built into it, but it's going to help you manage your smart home.

Since I'm talking about Apple, Apple rolled out a new version of Final Cut Pro this week, Final Cut Pro 11, and it has a ton of AI features in it. It's got a magnetic mask feature, which will sort of do the rotoscoping for you in video and mask out any object in the video. It's got spatial video editing for Apple Vision Pro, so you can edit videos and turn them into 3D videos designed to be viewed on the Apple Vision, and a handful of other features. Now, I don't use a Mac right now; I use a PC. I use DaVinci Resolve, but their magic mask feature, which DaVinci already has, but this Apple version looks pretty dang good. It looks like the best I've seen so far.

Since we're talking about Apple, I had to show off this video. This guy, Alex Chima, here over on X shared his setup where he hooked up four brand new Mac Minis together, and he's using them as his AI machine. He's using this Exol Labs, which allows you to run AI on any device, and he hooked up the four M4 Pro Mac Minis together, and he's running NVIDIA's Neotron 70B on it and said he can even run Llama 40B on it, which is a huge model. I just thought that was interesting and wanted to show that off.

If you're a fan of the Arc browser, the Arc browser is now available as an Android app. We've actually had the Arc browser as an iPhone app for a little while now, but now that there's an Android version, apparently Yahoo Mail is launching an app update with some AI features, like it'll summarize emails for you and help you improve emails or make them more concise or make them more friendly or whatever—very similar to what we've seen with Gmail and Outlook and some of these other tools that have rolled AI into it. And if you're thinking what I'm thinking, it's that people still use Yahoo email addresses. That's news to me.

NASA and Microsoft are working together on Earth Co-Pilot. We can see a demo here; it's using all of NASA's information and it's using Microsoft's AI, so you could ask questions about complex geospatial satellite imagery and data and unlock new insights and discoveries. But basically, it's got all of the data that NASA collected from its various satellites, and it's making that data available to essentially chat with.

It seems like the Vatican just worked with Microsoft to create a digital twin of St. Peter's Basilica so that people can do virtual visits of it, as well as so it can document some of the damage and log issues with the building and things like that. They used over 400,000 high-resolution digital photos to actually get this, and if you head over to this long complex URL, which don't worry will be in the comments, you can actually take a peek here. We can click on different areas of the map here, and by clicking on one of these areas of the map, it will sort of set our vantage point. You could actually see my character down in the bottom right here moving to this point on the map itself, and then once I get there, I can actually drag around and look around at what's in this imagery, and I can actually get really, really close up and get a detailed look at a lot of what's going on here in the Vatican.

Now, some of the shots just show this sort of grid outline, which in my opinion looks pretty cool anyway, but once we zoom in closer on certain elements, then things come into focus. I just think this is really, really cool, and I'm excited to see more projects use this technology and allow us to explore places that we might not otherwise ever get to explore.

This artwork here was generated by AI. It's a photo of Alan Turing, and it sold for $1 million. Now, this wasn't generated with like MidJourney or Stable Diffusion or something like that; it was actually generated by this AI robot that is a humanoid robot with arms who actually physically did the drawing of the image. We can see here it sold for $1,084,800 after 27 bids. Again, here's a closer look at the image that was created, and here's a slightly closer image of the robot that painted the image.

A few months ago, we talked about how a Beatles song was assisted with AI. They used AI to restore some old audio footage, and apparently this song is now up for a Grammy. It'll be competing with Beyoncé, Taylor Swift, and Billie Eilish. Personally, I'm rooting for the AI.

If you've always wanted to ride in one of these driverless cars, they've been available in San Francisco and, I believe, Phoenix and a few other cities. Well, now if you're in Los Angeles, you'll be able to use them in Los Angeles. As somebody who lives in San Diego, I'm actually up in Los Angeles fairly frequently, so I'm excited to finally try one of these.

Unry Robotics open-sourced their G1 dataset here, their data collection model, their learning algorithms, the dataset, and models—all of it was open-sourced so other robotic companies can leverage the information and the research they've done to develop what we're currently seeing from the Unry robots.

In the last bit of robot news, this is a robot here out of China. This is from Deep Robotics, and we can see this is a four-legged robot with wheels that looks like it can pretty much go on any terrain. Now, I don't know if you remember a couple of months ago where I showed that robot dog that had a flamethrower on it. If you can find that robot dog with a flamethrower and this robot, that's like nightmare fuel right there.

But again, I always like to end on robotics because robots are so much fun to me, and they're so exciting, and I just like robots. So, that's my final story of the week. I hope you enjoyed this episode. Lots and lots of cool AI stories coming out right now; a lot of progress happening. I think, like I mentioned last week, there will be a bit of a slowdown at the end of November. Most of December, usually we don't get a lot of new announcements due to the holidays, but come January in 2025, I think we're in for a wild ride in AI like we haven't seen anything yet.

What I know that's coming in 2025 and what we've sort of gotten sneak previews of is going to be the biggest year in AI yet. I mean, bigger than 2023, bigger than 2024, and I can't wait for it. I can't wait to show off some of the stuff that I know is coming, and stick around on this channel. Subscribe if you haven't already, like videos like this one to make sure more of them show up in your feed, and I will do my best to keep you looped in on the latest AI news, bring you cool AI tutorials, and share all the interesting AI tools that I come across.

So much fun stuff to share, so subscribe if you haven't already. I really, really appreciate you, and if you haven't already, check out Future Tools, where I share all the cool AI tools I come across. This news page is up to date on a daily basis—way more news on this page than what I'm even able to share in these videos. There's a free newsletter that will deliver the cool AI tools and the most interesting AI news directly to your inbox completely for free. When you sign up, you get free access to the AI income database, a little database I made of interesting ways to make money leveraging these AI tools. You get that for free as well.

Just head over to Future Tools and join the newsletter. Once again, thank you so much for tuning in. Thank you to Hostinger for sponsoring this one. I really appreciate all of you guys, and I will see you in the next video. Bye-bye!