Weekly AI Recap: GrokBot Just Did the Unthinkable, ChatGPT Looks Completely Different & More!

0:00The AI news never stops. We got another week packed full of exciting updates in the world of AI. First, Grok bot did something that's kind of insane. It's a bold move, but it puts them way ahead of the competition. Open AI then released GPT-6 to everyone and created a new way of interacting with Chat GPT that will change the way that you use it moving forward. Meanwhile, Anthropic released some very controversial updates to Claude that are upsetting a lot of people and we have a new app from Google that will allow you to create any game you want with zero coding skills. My name is Paul J. Lipsky and every week I

0:39make a video just like this one where I cut through all the noise in the world of AI and just tell you what actually matters. So, to stay on top of all of it, make sure to subscribe and now let's get into it. I want to start by talking about Grok bot because what they just did is honestly shocking for a leading AI company. If you haven't used it, Grok bot is SpaceX AI's always-on AI assistant. From day one, I've been calling it one of the best AI assistants

1:04ever and now it's gotten even better. This week, Elon Musk posted that going forward Grok bot will use the best back-end model for any given task and he specifically named Claude Opus 5.5, Midjourney, and Suno. So, if you ask your Grok bot for an image, it might use Midjourney. If you ask it for a song, it might use Suno. And for a really hard task, it might hand the work off to Claude, which is made by Anthropic, one of Grok's biggest competitors. Think about what that means for a second. Up until now, there's been a division when it comes to AI agents. On the left, we

1:39have the agents like Grok bot and Muse. They just feel really easy to use and they really nail the vibe of what an agent should feel like, but they're powered by weaker models. On the right, we have Chat GPT dots. Since dots is still really new, it's pretty rough around the edges and I don't think it's as useful of a tool as something like Grokbot or even Muse. But it's powered by one of the best models in the world, GPT-6 Astra. That meant that you always had to compromise. Do you want the best app or the best model? You couldn't have both. But now Elon Musk is saying you

2:18can have the best of both worlds. What I think is the best AI agent in the world powered by one of the best models in the world. If this works well, it will be incredible and drive a lot more people to use Grokbot. There are a lot of questions and some concerns around it though. First, it doesn't seem like we as a user will get to choose which model it uses. Generally, I think that's going to be okay, but I can see many times where I would be frustrated if I wasn't able to choose Opus over one of the Grok models. For instance, writing seems like

2:50a very basic task. So Grokbot would probably just use Grok for that. But I much prefer the way that Opus writes. Second, we're not sure how much this is going to affect the pricing of Grokbot, how much faster it will burn through our Grokbot usage once it rolls out. And the third question is, are we allowed to opt out of this? Some people won't want to use it because of the cost or other reasons. But overall, I think this is a great idea and removes one of the biggest reasons people weren't using Grokbot. And already I thought that Grokbot was probably the best AI agent in the world and now that it's backed by

3:24the most powerful models and the best models in the world, I think it's going to be really hard to beat Grokbot moving forward. But that wasn't the only update to Grokbot. Grokbot also got a new bot inside of every account. It's called your primary bot and you'll know it's your primary bot because it has this star icon next to it. And these bots are designed to be proactive. So instead of waiting for you to ask it questions or to assign it work, it looks for work it can take off your plate and offers to handle it for you. You can see example of that right here. Unfortunately,

3:56there's a lot of this I have to blur out, but right here there is this deadline that was approaching where I had an offer in my email inbox that was expiring in a couple of hours, and it let me know that, "Hey, you need to give an answer for this right now, otherwise you're going to lose this opportunity." And it just did that on its own. Every company is headed this way. Dots and Mews both do this, too, where they proactively surface different tasks that they think they can handle for us. But quite honestly, I haven't seen any of these companies really nail this yet. I

4:27would say about 25-30% of the time what they surface for me is actually helpful, and the other 70 or 75% of the time it's just another notification. And I don't know about you, but I don't need more notifications in my life. So, it's getting there, but they all need more work when it comes to proactively helping us, otherwise it's just going to be more noise. And the final bit of Rock bot news is that it can now read, search, and monitor X without any extra setup. So, if you tell it to keep an eye on something that's trending on X, it can do that for you. And it used to be

5:02able to do this, but you had to pay extra through the X API. That is no longer the case, it's just natively built in. And before we move on, let's take a quick moment to hear from the sponsor of today's video, Oku. Oku is an AI teammate that connects to over 3,000 tools like Gmail, Slack, and Notion. It also has built-in data sources for things like social research, people search, and list generation, so it can deliver finished work instead of just suggestions. For example, every week I start my AI news video by digging through a pile of announcements. So, I asked Oku to take that part over. I told

5:37it to check the official blogs from Google, OpenAI, and Anthropic, pull the biggest updates from the past week, and write a brief with a summary and source links for each one. Then save it as a Google Doc, ping me in Slack when it's ready, and do this every Thursday morning. And because it runs in the cloud, I can close my laptop and it keeps going. When I come back, the summary is waiting in Google Docs with a link to the sources. I still check the sources before anything goes into a video, but now I'm starting from a

6:03finished brief instead of a blank page. And since it's scheduled, it'll do this again next Thursday. This example ends in a Google Doc, but Oku can also build reports, slide decks, and web pages. And once the workflow is saved, I can hire anyone onto my team and they can run the same brief from Slack without setting anything up. You also decide exactly what Oku can touch. It can draft an email, but it won't hit send for instance until you approve it. Get started with Oku for free using the link in the description, and now back to the news. All right, let's talk now about what OpenAI has been up to. They

6:38released some pretty cool updates to ChatGPT. So, first of all, GPT-6 is now powering ChatGPT for everyone, including if you're on the free plan. So, coming into ChatGPT to regular chat, if you click on the model selector right here and click through to the more advanced settings, you'll see it is now powered by GPT-6. If you have a plus, pro, business, or enterprise plan, this will be GPT-6 sole. And for free and go users, it'll be powered by GPT-6 Luna. They also rolled out something called intelligent UI. So, instead of giving you a big wall of text, ChatGPT can now answer with an

7:19interactive layout where it makes sense. I've been testing this out and it's actually really cool. Here for example, I asked ChatGPT, "What is a Tabata workout?" And instead of just giving me a text response, it also drew up this little table for me to explain what a Tabata workout is. Then coming down here, it gives an example of one of them. And then I asked it, "Hey, can you build me a timer for this workout that you just suggested?" Then it actually built me a timer for my workout. I can press start here. It begins counting down, and I can go through each of the different exercises here with this app

7:55that it just built for me. I then asked it to help me learn how to do a handstand, and it gave me a What is this? A four-step process, and found images from the web to match each step of the process. And then coming down a little bit further, you can see it's asking me these questions. How to tell when you're ready to progress in your handstands. So, we have these handstand milestones, and as I'm going through this, I can check them off, and then I can come down here to have it start coaching me based on how well I'm currently doing with my handstands. I

8:30mean, look what it's drawing up for me right here. That's pretty cool. All right, let me show you something else. So, here I gave it a link to an IKEA product, and said, "How do I assemble this?" This one it struggled with a little bit. I'm not sure what it was trying to surface right here. I think it was something from the instruction manual from IKEA, but

8:48obviously it didn't show up correctly. So, I prompted it a little bit. I said, "I'm having trouble visually understanding it." So, it not only found an image for me, but also drew up this diagram, which shows very clearly how to actually assemble this unit. There are also some fun things you can do with this. So, here I told it I want to start learning how to meditate, but I told it specifically to give me a response in Microsoft Paint style UI. So, here is an explanation on how to start meditating, but you see kind of looks like Microsoft Paint, and I can come in here and use

9:20this timer. It created this little app for me. I can check off each day as I'm actually meditating day-to-day. And then here is a fun one. I asked it to create for me an interactive scary story. And then coming down here, we can see the story, read it, and then I can make a choice as to what happens next. So, I'm

9:40going to say I want to lock the door. Then the story continues, and I get another choice. So, yeah, this is a fun one. Definitely give the new intelligent UI a shot, and let me know if you find some cool stuff you can do with it. OpenAI has also shared some updates on dots. Like I said before, I think dots is a little bit rough, right? It can use a little bit of work, and luckily OpenAI

10:01has said that they are working on it. Specifically, when it comes to dot voice, when you call your dot, this is what they're working on. They're working on making sure that the connection is faster and more reliable. They are going to keep that phone call UX, but they're going to make it faster. They're going to make sure that you can see the transcript from your calls and summaries of them, and they're going to add more variants. Because right now, when you call your dot, it likes to say checking,

10:27checking, checking over and over again. So, they're going to add some variants in there, and they're going to make sure that it's regularly updated. So, they are working on dots, even though it's not perfect yet. It does seem like it is a priority for them to make it look great. And finally, OpenAI's Thibault, who leads engineering for Codex, made a pretty bold promise this week. For the next 28 days, every single day, they'll either ship a clear improvement for ChatGPT or they'll give everyone a full usage reset. A lot of what I showed you already has been part of that 28 day challenge, but they've also increased

11:02the speed of GPT-6 Astra and GPT-6.1 Soul by about 50%. So, either way, ChatGPT users this month are going to come out ahead every day. You're either going to get something better or you're going to get more usage. Pretty cool. Over at Anthropic, we got a new model, Claude Haiku 5.5. Haiku is Anthropic's smallest, fastest, and cheapest model. So, it goes Haiku, Sonnet, Opus, and then Fable. Generally speaking, the lower you go, the faster the model gets, the cheaper it gets, but it's also the least intelligent. And even though this is an update to their least intelligent model, it's still quite impressive. It's about 75% cheaper

11:43than the last Haiku model, and it also performs better. But for the average person, I still think you're going to be better off using Sonnet or Opus most of the time. This 5.5 family has become very, very good and very efficient in terms of cost. So, if you're on the $100 plan, I would be using Opus 5.5 all the time. That's what I'm doing, and I'm not hitting my limits with it. And if you're on the $20 plan, you can probably get

12:09through an entire week using Sonnet 5.5. Overall, Anthropic is crushing it when it comes to these new models, and it's predicted that Fable 5.5 should be coming out pretty soon, and when it does, I think it is going to blow everyone's mind. So, definitely stay tuned for when that happens. The other big news and controversial news from Anthropic this past week is that Claude Work now runs in the cloud. This was something that took a lot of people by surprise because, quite honestly, Anthropic did a very poor job at warning people about this. So, here's what's going on. As of October 6th, any new Claude Work tasks run on Anthropic

12:50servers in the cloud instead of on your computer, and you can't switch back. And this upset a lot of people, mostly because we had no idea that this was happening. Anthropic claims that they sent an email about it, and you should have received a notification in the app about it. Personally, I didn't receive either of those. I only knew about this because one of you guys emailed me about it, and I saw some people posting about it on Twitter. But the other reason this is upsetting a lot of people is because it kind of ruins one of the selling points of Claude Work, which is that

13:24everything stays on your computer. But Anthropic has said that this is actually an improvement. So, how does this work now? Well, when you come into Claude, what you'll first notice is that there's no way to choose between chat and co-work. Depending on the task that you give it, it will make a choice as to whether to use regular chat or co-work for that query. Either way, what it's

13:45doing will run completely in the cloud. So, let's say for example, you're having it grab something from your Gmail and then based on what's in there, update a page in Notion. Those are both cloud-based apps. So, that use case makes sense. It makes sense that those would run in the cloud. I think that's pretty easy to understand. But, let's say you tell Claude to grab some files from your computer and then to use what's in those files for context when drafting an email. What Claude will do is copy those files to the cloud while it's working. Anthropic said that this is a good thing because now your task

14:19will keep running even when you close your laptop. But, that doesn't make a lot of sense to me. Because in order to upload the files, your computer needs to be on and in order to update the files on your computer, your computer needs to be on. So, if my computer's on anyway, why not just have them run locally? The only real benefit to this I see is that the mobile app is so much better now because now every task that you start inside of Claude, whether it's chat or co-work, you can easily control from your phone. It's a much better experience than it was in the past. And

14:54I will also say that even though I'm a big user of Claude co-work, the experience for me feels exactly the same. Like, I don't even know if I would have known there was a change unless someone told me. But, if you don't like it for some reason, then the best alternative is probably to use Claude code. And even though having the word code in there might make it sound a little bit scary, it's actually really easy to use. And if you were using Claude co-work quite regularly and you really understood how it worked, I think you'll find that Claude Code is very similar in many ways. If you guys want,

15:28I can do a full video on how regular non-technical people can use Claude Code, and you'll see that it's really not that bad. And the final update from Anthropic is that Claude now works inside Google Docs, Sheets, and Slides. So, you can see right here, I have this really dense document here. It's kind of a recap of all the AI news from the past week or so. So, I'll click on

15:52extensions, and I will open up Claude. Now, you'll see it over here on the right, and I can say something like clean this up, make it teleprompter ready for recording a YouTube video. And there we go. Now, it's teleprompter ready. So, like I said, this works also inside of Sheets and Slides. So, if you're a big fan of using Google Drive, but you don't like using Gemini, you now have a choice of using Claude. To add this, you click on the top here where it

16:19says extensions, and hover over add-ons. Click on add-ons here, and then search in here for Claude. Make sure you're downloading the official one. It should say it's from Anthropic in blue here. And then, you'll want to install this one. And this is currently in beta for all paid Claude plans. Speaking of Google Drive, let's talk about what Google has been up to in the world of AI because they've honestly had a bit of a week. But, let's start with some good news. Inside of Drive, you can now preview, edit, and collaborate on markdown files natively across Drive and Docs. That is a big change because in

16:55the past, anytime you would upload a markdown file, you could either preview it, or it would convert it into a Google Doc. But now, you can edit it natively. You can see up here, it is keeping it as a markdown file. That is really important because pretty much every AI tool writes in markdown files, and it's sort of the language of AI. So, the fact that it is now incorporated into Drive is really nice. I think this could unlock a lot of workflows where you keep all your files in Drive and you just

17:27start pointing your AI agents to Drive. I don't have access to it yet, so I haven't really been able to test that, but this is a nice thing to see. But now for some bad news. They are changing the models that you're going to have access to based on the plan that you're on. So, if you don't have a plan, in other words, you're on the free plan, you will only get access to their weakest model, which is Flash Lite. And I think that's pretty rough because inside of Google, I I don't think Flash Lite is all that good. Flash is a great model. Pro is

18:00also really great, although they are pretty behind on updating it. But Flash Lite is pretty weak. And the fact that you don't get access to it at all on the free plan is pretty rough. Now, for Plus subscribers, you will get access to Flash, but you don't get access to Pro. And then for Pro and Ultra subscribers, you get access to all three models. The other thing that's going to start rolling out is that you'll be able to choose the effort level between low, medium, and high. That should be in this drop-down pretty soon. I heard it's rolling out right now. So, my take on this is if you're on a Pro or Ultra

18:35subscription, you're not going to notice any difference. But on a Plus subscription, you probably won't either because honestly, you should be using Flash. It's their best model right now. Once they release a Pro model, then you're not going to have access to that, which will stink. But who's really going to feel affected by this is anyone on a free plan because Flash Lite, like I said, I don't think is a super great model. And for free, you can get GPT-6 Luna on a ChatGPT subscription, which I think is better than Flash Lite. So, this may drive a lot of people away from using Gemini who are using it for free

19:10and towards ChatGPT. But of course, there are probably plenty of people who are on free plans who don't even know this is happening and won't even notice, to be honest. Google also released a new image generation model called Nano Banana 2.1. This is obviously a step up from their last model, Nano Banana 2. And this is supposed to be better across the board. They have actually said that it's supposed to be better than Nano Banana Pro. Now, I released a video right when 2.1 came out comparing it with Nano Banana 2, Nano Banana Pro, and chat GPT images 2.5. Doing a bunch of

19:45prompts and comparing them side by side. And quite honestly, I don't notice like a huge difference with it. You can go watch that video if you want, but the step up, if any, seems very, very slight. And I definitely don't think it beats out Nano Banana Pro or even GPT images 2.5. Sometimes it's better, but a lot of times it's not. So, this wasn't

20:08the huge leap that I was hoping for. Hopefully, they soon come out with a new Nano Banana Pro model because even though that model is quite old in terms of AI time, it still is a fantastic model, and I think it's still one of the best. I still think it beats GPT images 2.5 a lot of the time. So, once they release a new Nano Banana Pro model, I think it will absolutely crush the competition. Google also released a new experimental app called Playground. You can access this at playground.google.com, and the idea behind this is that you can use this to generate games or to play

20:48other people's games. So, it's supposed to be an easy way for non-developers to code or vibe code different games. So, I'm logged in to mine. I click on the left here where it says create, and then you can describe here what game you wanted to create. You can choose between a 2D game or 3D game, single player, multiplayer is coming soon, and you can even upload files as references. So, let me show you some of the games that I created and the prompts that I gave it for them. So, I tried two, it's the same game twice. So, this is the first game, the first time I tried it. I basically

21:22was trying to create RollerCoaster Tycoon. So, I said, "Build a roller coaster park builder game, any roller coasters, I need to track customers, we have upsells like food and merchandise, and other things like that." And so, then it builds the entire app for you. It's very, very simple in that regard. There's no technical language that you need to know, and then you can play it over here on the right. And here is where the problem starts because what I've discovered is that using this, I find it's actually really hard to get a nice-looking game. So, watch. I'll play this game. I can build a roller coaster

21:54here. So, this is like the start of the roller coaster, and then I can build some tracks, and just look how bad that looks, right? It's not connected. You can build different shops and stuff. It just doesn't look like a great game. Now, I know you can get decent-looking games because if you come in here, there are some games that other people have created that look pretty good. But, I think it requires a lot more work. And when I go over to Claude or ChatGPT and tell it to build a game, it can be as simple as build me RollerCoaster Tycoon, and it will build me a great-looking

22:26game. And one of the reasons for that is because it will actually play the game to check for any bugs and to make sure it looks good. Doesn't seem like Playground is doing that. So, if you want to spend the time building a great game, I think it's possible with this, but it's not as easy as it seems the first time you try it out. Let's round out the Google news with some good news, or at least I consider these things to be good news. So, SynthID is getting better. So, SynthID is a tool that allows you to detect if an image was generated with one of Google's image

22:58generation tools. But now, they've expanded their partnership. So the SynthID will work with other images that were generated from other providers like OpenAI and soon Apple. So you can come over to synthid.com, upload an image, and it will identify if it was AI generated. And then finally, Google built an accessibility feature into Gemini Live. This was built alongside the blind and low vision community. It's called Guided Vision, and it should offer conversational real-time visual assistance. So all you have to do is pull out your phone, share your camera, and you'll receive dynamic audio descriptions and natural verbal framing cues as you move through the world. So this should help a lot of

23:43people who are blind or have low vision. I always love when I see that AI is being used to make a positive impact on the world. And to give them a lot of credit, a lot of the best features are coming out from Google. So that was a pretty big week. My big takeaway from this week is that Grok bot started out the week as the best AI agent in the world, and they've just increased that lead by even more. It is really hard to beat what they're doing right now. And I think all the other companies are going to have to find a way to catch up in

24:15that department at least. Do I still use Muse? Do I still use dots? Yes, absolutely, but I think Grok bot is definitely setting the pace right now. The other thing I'll say is that the more I use it, the more I actually like the updates to Claude Co-work. I like that I don't have to switch between chat and co-work, and honestly, the trade-off, which is putting everything in the cloud for us, makes it so much more convenient because I can run all my tasks from my phone now. That is a trade-off, but for me at least right now, it's really worth it. But let me know your thoughts in the comments

24:49section down below. What is your biggest takeaway from this past week that you are most excited for? And while you're down there, make sure to subscribe to the channel. I release a weekly AI recap like this every single week. So, if you want to stay on top of all the AI news, make sure to subscribe, and then I'll

25:05see you next week. Bye for now.