AI News: Opus 5.5, GPT-6 Sol, Jev, Muse and More!

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: AI News: Opus 5.5, GPT-6 Sol, Jev, Muse and More! · Channel: Matt Wolfe · 34:23 · Watch on YouTube

Topics: AI agents, AI content generation, AI infrastructure and hardware, AI models, Multimodal models

Summary

Matt Wolfe recaps a busy week in AI, led by Meta Connect announcements about the Muse agent, new glasses, and a lightweight mixed-reality headset. He compares three new models—GPT-6 Soul, Claude Opus 5.5, and Grok 4.7—highlighting their performance, costs, and results on his game-building tests. He also explains Jev, a model designed to produce structured decisions rather than conversational text, and surveys updates from YouTube, Microsoft, Google, and Spotify. The video closes with Google’s Project Sun Catcher, which aims to test whether TPUs can operate in space, and a preview of expected news from OpenAI Dev Day.

Key points

Meta introduced updates to Muse, including voice access, additional connectors, computer use on Mac, and an email address for forwarding tasks.
Meta announced new AI glasses, camera-free Ray-Ban Meta Audio glasses, and a lightweight mixed-reality headset expected in spring 2027 for $1,299.
OpenAI made its GPT Live One API generally available, enabling a voice assistant to converse while an agent researches in the background.
GPT-6 Soul and Luna are presented as lower-cost models; the video reports that their prices were cut relative to their predecessors.
Claude Opus 5.5 is described as a leading model for coding and other benchmarks, though its token use can make its cost per task comparatively high.
Grok 4.7 is cheaper but performs less impressively than Opus 5.5 and GPT-6 Soul in the creator’s game-building tests.
Jev produces structured outputs such as choices, scores, or null values, aiming to make decisions quickly and cheaply instead of generating long text.
YouTube announced features including custom feeds, question-based search, live-stream dubbing, and AI-assisted tools for creators.
Microsoft, Google, and Spotify announced new AI features, while Google’s Project Sun Catcher will test the possibility of operating TPUs in space.

Chapters

Start	Chapter	Summary
00:00	Meta Connect: Muse and new AI glasses	Wolfe covers Muse’s new capabilities, camera-free Ray-Ban Meta Audio, and updates to Meta’s display glasses.
04:42	Meta’s mixed-reality glasses and creation tools	He describes Meta’s lightweight headset, its features and planned release, then recaps AI tools for creating games.
08:22	OpenAI’s live voice API and GPT-6 Soul	Wolfe demonstrates voice-driven research with OpenAI’s API and compares GPT-6 Soul and Luna on price and performance.
13:00	Claude Opus 5.5 in coding and animation	He presents Opus 5.5’s benchmark claims and examples of its game-building, animation, and coding capabilities.
18:12	Grok 4.7 and model comparisons	Wolfe tests Grok 4.7 and compares the three new models’ benchmark rankings and costs per task.
21:54	Jev: a model for structured decisions	He explains Jev’s structured outputs, speed, cost, example applications, and the temporary pause on new sign-ups.
26:34	YouTube and Microsoft AI updates	The rapid-fire section covers YouTube’s new viewer and creator features, followed by Microsoft’s Copilot modes.
30:15	Google, Spotify, and Project Sun Catcher	Wolfe reviews Google’s live avatar, text-to-speech, and Vids updates, Spotify’s taste profile, and Google’s plans to test TPUs in space.
32:50	Wrap-up and OpenAI Dev Day preview	He closes by discussing the week’s pace of AI news and previewing his coverage of the upcoming OpenAI Dev Day.

Quotes

it now has computer use. — 02:07

the conversation and the research can happen at the same time. — 09:24

the result is pretty mind-blowing. — 14:33

Now, Jev, this new model, is calibrated for decisions. — 22:57

Google is actually trying to put their tensor processing units, TPUs, up into space. — 32:20

Concepts and entities

Muse, Meta, Meta Ray-Ban, Horizon Create, Horizon Studio, GPT-6 Soul, GPT-6 Luna, Claude Opus 5.5, Grok 4.7, Jev, TypeSafe AI, Artificial Analysis, Project Sun Catcher, Google

Full transcript (original language)

00:00 This is another one of those weeks that’s been absolutely insane. I just got back from Meta Connect where there was a ton of announcements, but we also got not one, not two, but three new models. The Made on YouTube event was also this week where they announced a whole bunch of new AI features for YouTube and for a couple days the whole AI world went crazy about something called Jev. So yeah, we got a lot to talk about. I don’t want to waste any more of your time. Let’s get straight to it. Let’s start with Meta Connect because well, I literally just got back from this event and it’s the hottest

00:31 thing on my mind right now. They made a bunch of announcements related to AI, their Meta glasses, as well as a new like mixed reality headset. The entire keynote was about 55 minutes and you can find the whole thing online, but here’s like the TL didn’t watch. First off, they made a bunch of announcements around Muse, which is like their AI agent that’s very similar to Open Claw, but without all of the setup and headaches to get it running. And if I’m going to be totally honest, it’s pretty impressive. I’ve been actually using it quite a bit to check my emails and add

01:04 things that it finds in my email to my calendar, as well as to like look at my Instagram and Threads accounts and tell me ideas of things I could post about that are likely to go viral based on other things I’ve recently posted that have done well. So it is actually really cool and impressive and I highly recommend everybody check it out because unlike most of the other agents that are out there, there is no complicated setup and if you’re in one of the regions that currently supports it, it’s free to use right now. But here’s the announcements they made around Muse. First off, you’re

01:35 going to be able to use Muse with your AI glasses. So if you have like Meta Ray-Bans, you’re going to be able to direct your agent to go do things for you on your behalf by just talking to your glasses. That same voice mode is available inside of Muse as well, not just on the glasses, but you can use it inside of the app and inside of the browser as well. They also announced a bunch more connectors. I think one of my biggest complaints when I first tested Muse was that there wasn’t enough integrations with it yet. But, based on their keynote, there’s a lot more integrations rolling out soon, including

02:07 the one that I’m probably most excited about, which is Granola, which is what I use to record most of the meetings I’m on. They also announced that if you use the Muse app on your Mac, it now has computer use. So, it can actually take control of apps on your computer, and you could walk away and it’ll just keep on doing things for you. This isn’t a feature I’ve had the opportunity to test yet. But again, as far as agents go, this is probably the easiest one to onboard yourself onto, and now it has computer use. They also announced that your Muse will be getting its own email address. So, you could add conversations

02:39 to a thread, and your Muse will be able to see those conversations and pick up the information in them inside of your agent. So, you can like forward things to it and ask it to do tasks for you. If somebody says, “Hey, can we schedule a call?” you could forward that email to your Muse and say, “This person wants to schedule a call, find a time on my calendar that we can schedule the call, add it to my calendar, reply to this person, and tell them the time, or give this person three times to pick from my calendar. Once they select one, add it

03:09 to my calendar.” So, you can have it go and just do that stuff on your behalf now. They also teased something coming called the Muse charm, which is, you know, sort of being compared to like a Tamagotchi. It’s a little handheld device that you can put on a keychain, and you can instruct their agent with that. They didn’t give a ton of details on this one. They basically said it’s something that’s still in production, but more details are coming soon, and they expect it to be ready for the holidays. They also announced a ton of new AI glasses, so a whole bunch of different form factors. Like, they had

03:40 these massive displays of just all of the different form factors you’re going to be able to get these glasses in. But, they also announced something that I think a lot of people are pretty excited about, which is the Ray-Ban Meta audio. So, it’s everything you get in the Ray-Ban Meta glasses, they just took the cameras away. These Meta Ray-Bans with the cameras people have been calling them like creeper glasses or pervert glasses or things like that because they’re afraid that people are like recording everyone around them without the consent. And people are starting to get mad when they see other people wearing the glasses with the cameras on

04:11 them. So, Meta said, “Fine, we’ll offer some glasses that don’t have the camera.” So, if you want all of the functionality of the Meta glasses, but you don’t want to be walking around with cameras on your head, well, that’s what these Ray-Ban Meta audios are for. I did get to test these as well and they sound really, really good. They did also touch a little bit on the Meta Ray-Ban displays, which are the glasses they announced last year that has the little like wristband where you can control the glasses with like hand gestures and then you also have like a little display on one frame of the glasses. There’s things coming like cycling and public transit

04:42 directions, navigation that actually sounds more like a person, a more powerful calendar, mirrored notifications from your phone, threads for community feeds, as well of course the ability to control your Muse agent with these glasses as well. But by far the most sort of show-stopping moment that seemed to impress the most amount of people, myself included, was the new Meta VR glasses. Because these are like your Meta Quest but without the big bulky VR headset. They actually just go

05:13 over your ears like regular glasses. And the quality when you’re looking through them are like just as good if not better than the Apple Vision Pro. And the feeling when you’re using them is very similar to an Apple Vision Pro. You get really clear pass-through where you can see everything crystal clear. You use eye tracking to select stuff. You use finger gestures to like click on things. And I think when people get their hands on these, they’re going to be really, really impressed by them. They’re also super light. They’re only 100 g, which is the weight of about like a deck of

05:45 cards. And they do have like this little puck thing. So, you put the glasses on, there is a cord coming off the glasses that goes to this little battery, but pretty much all the compute is happening on this battery as well. So, that’s your power, but that’s also where all of the compute is happening, which is how they get the glasses to be so light is that all the technical stuff is happening in this device, not within the glasses themselves. Again, think of this is almost like an Apple Vision Pro, but without the giant headset, just more like glasses that you put on your face

06:17 with a little like puck coming off of them. But if you’ve ever tested an Apple Vision Pro, like it’s like that kind of quality when you’re looking through it. And you’re going to be able to do everything that you can do on a Meta Quest right now. Pretty much everything that’s available for Meta Quest is going to be available for this as well. So, any games you play on Meta Quest, you’ll be able to play on this device as well. And they also showed off that calling is now going to have a full hologram, you know, with legs and all, not a cartoon character like the old Horizon Worlds, not like what you get out of an Apple

06:48 Vision Pro where it’s just sort of like a boxed person. Like you’re actually going to see the person standing in front of you with legs and all. That feature, by the way, is also coming to the Meta displays. You’re going to be able to see that full hologram with the Meta displays pretty soon as well. They’re coming spring of 2027 for $12.99, so 1,300 bucks. And spring can’t come fast enough cuz like these are really cool. Now, on the second day of Meta Connect, I actually wasn’t there for the second day, but they did make even more announcements for developers, including this new Horizon Create and

07:20 Horizon Studio. This is basically AI-powered platforms to create things for the Meta devices. Two new tools built on agentic creation capabilities of the Meta Horizon engine. Both enable creators to turn any idea into a complete 2D or 3D mobile game complete with progression systems, balanced difficulty, art direction, multiplayer, and more. So, with Horizon Create, you’re going to be able to open an app on your phone, describe the game you want to build in your own words, and then watch it come to life. With Horizon Studio, you can do the same in your browser. And I imagine it’s only a

07:51 matter of time before we’re going to be prompting like VR worlds into existence inside of the VR headsets as well. But, that’s pretty much the quick recap of everything they announced at Meta Connect. I think it was one of the better keynotes they put out in a while. I know there’s a lot of people that anytime I mention anything Meta-related, they’re going to give some pushback because there’s a lot of non-fans of Meta. But, personally, I think they’re putting out some pretty cool stuff right now. So, [laughter] and yes, I am slightly biased cuz I still have like an afterglow from being

08:22 at this really impressive event. So, keep that in mind as well. We also got some big announcements out of OpenAI this week. Not only did we get GPT-6 Soul and GPT-6 Luna, which I talked about in a recent video on my channel, we also got another feature that I did get a chance to play around with while I was at my hotel room. So, check this out. Next up is an update out of OpenAI. They just made the API for GPT Live One generally available. And it lets a voice assistant keep talking to you while an agent does the work behind the scenes.

08:53 I’m going to use it along with GPT Astra and Codex to build a live interactive AI news researcher. So, here’s my prompt. Build a voice-powered AI news app using GPT Live One and the Agents API. Include clickable sources. So, now I can ask Find today’s biggest AI stories. >> I’ll look it up. >> Actually, only include updates about new tools people can actually try. >> Okay, focusing on just those. >> Show me all the original announcements from the primary source, not just the

09:24 news articles covering them. What’s cool is that the conversation and the research can happen at the same time. I can keep talking and change direction as the agent works, instead of waiting for the research to finish before giving it another instruction. So, if you want to build something for yourself, try OpenAI’s APIs in CodeX today. And thanks so much to OpenAI for supporting my channel and sponsoring this portion of today’s video. And like I mentioned, we did also get GPT-6 Soul and Luna. And if you did miss my last video where I

09:54 talked about it, here’s the quick recap. These models are essentially like price cut models. They’re models that are still really, really good models, but are designed to be a lot less expensive than using GPT-6 Astra. So, if we scroll down and we look at the pricing here, GPT-5.6 Soul was $4 input and $20 output. With GPT-6 Soul, they cut that cost in half down to $2 input and $10 per million token output. GPT-5.6 Luna

10:25 to GPT-6 Luna also got cut in half from 20 cents to 10 cents and from a buck 20 down to 50 cents per million tokens. So, this is designed as an inexpensive, but good model. It is not a model that is just burning up all the benchmarks though. We can see GPT-6 Astra on Automation Bench here is still their best model, where the new GPT-6 Soul is still pretty good, right? Using it on extra high is about as good as using Astra on medium in terms of the score,

10:56 but if we look at the cost along the bottom, it’s quite a bit less expensive to use. On Agent’s Last Exam here, you can see here’s Astra, still their best model, but GPT-6 Soul and GPT-6 Luna are still decent models in terms of how they score. They’re just doing it for a lot cheaper. Now, I did give it my normal mega bunk test by giving it this prompt here, and you can see on GPT-6 Soul Ultra, so pretty much the most powerful mode you could put it on, it only worked for 24 minutes and 10 seconds. And

11:28 here’s what it created, and it looks pretty good. I mean, it’s not horrible. I can’t actually change the camera angle with my mouse, but but it’s it’s not bad. Definitely better than we were getting with Fable, but not quite as good as what I think we got out of Astra. So, pretty much what you’d expect from a model that’s almost as good, but it’s cheaper, and again, not quite as good on the benchmarks. I did run it through my Abusy bench here, and surprisingly, it ranked GPT-6 Soul as number one. But like I always say, use your eyes and decide the leaderboard

11:59 here doesn’t really mean a whole lot. This is using an LLM as a judge to decide the ranking, so it’s very subjective. I still think GPT-6 Astra looks a little bit better than GPT-6 Soul, but also interestingly, it put GPT-6 Soul Pro down here at number three. So, it actually ranked the non-pro version above the pro version. And when we compare them side by side, I think I agree. This whole version is actually better than the Soul Pro version. Soul used 30,301 tokens versus 148,000

12:30 tokens. Soul Pro took 8 minutes, while Soul took 4 minutes. Soul took 30 cents, and Soul Pro 92 cents. So, the one on the right cost 92 cents to make, the one on the left cost 30 cents to make. But when you’re using ChatGPT, if you’re just using chat, it appears to be using 5.6 Soul, but if you switch it over to work, you can set it specifically to GPT-6 Soul or Luna now. But again, I said we got three new models, so we’ve got GPT-6 Soul, and we also got Claude

13:00 Opus 5.5. And Claude Opus 5.5, I did go into quite a bit of detail about it in this video here. So, check that out if you haven’t already, but here’s the TLDR on that one. This is officially the new state-of-the-art model. This is pretty much the best model there is, kind of hands down, at the moment. We do have OpenAI Dev Day next week, so who knows, the leaderboards could be shifting again in a few days. But as of today, Opus 5.5, we can see agentic coding here,

13:32 blows everything else out of the water. It doesn’t list GPT-6 Soul, but trust me, it’s beating that as well. When it comes to knowledge work, new state of the art, humanities last exam, new state of the art, computer use, visual chart recognition, new state of the art. It also got a little bit cheaper than the previous version of Opus. So, Opus 5 was $5 per million input, $25 per million output. It went down to four and 20, respectively. When it comes to coding, it’s beating out GPT-6 Astra on both

14:03 cost and score right now. So, yeah, if you’re looking for the absolute best state of the art coding model at a decent price, this is it. This model’s performing better than Claude Fable 5.1 on almost all the benchmarks, but it’s over half the cost. From $10 down to $4, from $50 down to $20 to get even better performance. Once again, I ran my Mega Bonk test on it, and well, it worked for almost 20 hours. Building, testing,

14:33 building, testing, building, testing, and the result is pretty mind-blowing. This is the version it made, and it is pretty close to the actual Mega Bonk game. Like, look at all these characters you can get. I’m going to go ahead and mute the game, but these are all actual characters from the original game. So, let’s go ahead and select our fox here. We have to get to certain levels to unlock other games. There’s multiple tiers. Again, very similar to the real version of this game. Now, when I get into it, my mouse controls my angle. So,

15:05 as I move my mouse around, it’s changing my angle. The characters, look at these level up animations. Like, let’s go ahead and add another weapon. If anything, I think I level up a little too quick, but, you know, the game’s not perfectly balanced. The level ups that we are seeing are pretty much the same level ups that you get in the real game. And yeah, this is just absolutely mind-blowing to me how good this all looks. I’m going to go ahead and unmute it for a second so you can actually hear it as well.

15:36 Because the sound is pretty similar to the original game as well. Even the first boss in this game is similar to the actual real first boss in the other game. This is what 20 hours of Opus 5 running gets you. All right, I need to stop playing now because

16:08 it’s about as addicting as the original if I’m being honest. Look at this, even the pause menu looks excellent. Like there is so much detail in what Opus 5.5 did. I am insanely impressed with what this does in terms of like games. Now the other thing that seems to be blowing a lot of minds around Opus 5.5 is how good it is at animation. Like here’s some examples that I’ve come across. I’m going to mute them, but most of these even have audio that Opus 5.5 helped generate. So here’s an animation from Addy Osmani here which explains how

16:40 browsers work. And I mean, it’s a really solid animation, like really, really impressive. Here’s another one by Chubby here that actually explains how Transformers models work. And all of this is actually being created with JavaScript. So this isn’t it going and like actually generating an animation like something like VO would do. It’s not generating images like ChatGPT image or Midjourney. It’s using JavaScript to draw all of this out and then animate it. Here’s one from Tech Artist. This

17:10 like an architecture schematic animation that looks really cool. This one’s from Alexei Fative. And this one he had it take control of Blender and create a large Hadron Collider animation. So, this was Opus 5.5 taking control of Blender and generating this animation using Blender. So, not JavaScript this time. And here’s one more example from Chris Riley where another animation where it looks like tiles or I’m not quite sure what it’s supposed to be. I guess here’s the prompt if you want to

17:41 read the entire prompt. Looks like a two-parter here. So, there’s the beginning of the prompt. Here’s the rest of the prompt. You might need to pause twice if you want to see all that. But again, really, really impressive doing animations. So, I thought it would probably crush it on BusyBench, but BusyBench actually puts it all the way down here at number eight. Taking a peek at it, it cost $0.49, used 4 minutes, and about 25,000 tokens to generate this version of Busy. Dang, these animations are so good. Check it out. That’s

18:12 another one from Drew. But as mentioned before, we got three new models this week. And the third one that I haven’t mentioned yet is from Space X AI. They released their Grok 4.7. Now, compared to the models we just looked at, and not quite as impressive. If we look at the average cost per task, Grok 4.7 falls over here. So, kind of think of this side as the intelligence and this side as the price. So, it is less expensive, but not quite getting the same results as these newer models. It’s also not

18:43 comparing it to GPT 6 Soul. It’s not comparing it to Astra. It’s not comparing it to Opus 5.5. So, in all of these benchmark comparisons, it’s kind of comparing to like last generation models. But it is cheaper. Of course, you know I’m going to bonk bench it. And well, this is the version of MegaBonk that it came up with. Our character is a cube. Our fox character is like a pill-shaped thing. And our Clanker character is not openable yet. But let’s go ahead and use Sir Oofie here and let’s bonk. I mean, we can control the

19:15 camera angles with our mouse, but we got, you know, squares shooting at little pill-shaped things. So, you know, again, not quite as impressive at game development and game design. So, those were the three models we got this week. Let’s compare them over on Artificial Analysis cuz that’s where we could kind of see like an aggregated ranking. It sort of takes a bunch of different benchmarks, mashes them all together, and then tries to pick the objectively smartest models based on all of that information. And if we come to Artificial Analysis here, we can see that Opus 5.5 Max is the smartest model

19:50 pretty much hands down at the moment, beating out Fable 5.1 and GPT-6 Astra by pretty big margin, honestly. GPT-6 Soul falls down here at about 10 points below Opus 5.5, and then Grok 4.7 on extra high falls down here at 46. So, that’s where those three models we just look at rank on Artificial Analysis. I also like to look at cost per task, and over here Claude Fable 5.1 Max is still the most

20:20 expensive, but Claude Opus 5.5 is, well, the second most expensive to run right now. So, not a cheap model. Yes, it got a lot cheaper when it comes to cost per token or cost per million tokens. However, if we look at output tokens per cost, Opus 5.5 uses more output tokens than any other model. So, they lowered that price on output tokens, but the model is so much more token inefficient, I guess, that you’re using way more

20:51 tokens, so it doesn’t actually drop your cost per task by that much. On the other hand, when you’re talking about GPT-6 Soul, which is still a pretty good model, our cost per task is way down here at a dollar six per task, compared to almost $6 per task for Opus 5.5 and Grok 4.7, which you saw with our mega bunk test, was definitely not up to par with GPT-6 Soul or Opus 5.5 from like a game coding standpoint, is all the way up here at $3.74

21:23 per task. Again, because yes, the cost per million tokens is less expensive, but it uses a lot of tokens. Also, I just realized I didn’t show off the busy bench on Grok 4.7. Yeah, that’s what we got. It did only take 1 minute, used 7,000 tokens, and cost like less than 3 cents. But uh yeah, I mean, it’s not even in like the top 20 on the leaderboard right now. So, those were the three big models that launched this week. Opus 5.5, new state-of-the-art,

21:54 still not super cheap. GPT-6 Soul, almost state-of-the-art, pretty damn good model, super cheap to use. And then Grok 4.7’s also there. But there was something else causing a lot of buzz this week. It was actually a model that came out last week, but it seemed like it really kind of had a lot of people talking this week. And that’s a model out of a company called Type Safe AI called Jev. Now, I’m not going to go super deep into Jev in this video, because I feel like it’s a pretty deep topic, and I do want to cover it well,

22:26 and I don’t think I can cover it well in just a little teeny segment of this video. But essentially, when you think of large language models like the GPT models and the Claude models, those create new text. Even when you ask it to make decisions for you, it’s still creating text and writing code that then goes and makes those decisions. Jev, on the other hand, just skips straight to the decision-making. So, we can see right here, existing LLMs are optimized for human preference, write-ups and chat

22:57 responses that human readers prefer. Now, Jev, this new model, is calibrated for decisions. The output from existing LLMs are, you know, strings and generated text. The outputs from Jev are structured values. So, structured values basically means you’re defining what it could potentially output. And there’s really three types of structured outputs. You can give it a choice, so you give it some options to choose from, a score, or a null, which is basically, you know, true or false. And instead of

23:28 outputting a ton of text, it just outputs within these possibilities, along with a confidence score of that output. And because it’s not outputting just a ton of text, it’s just making these decisions, it is insanely cheap to use. Input tokens at 4 cents per million tokens, or $42 per billion tokens. Output tokens are free, because they’re just too cheap to meter. It’s also insanely fast. So, here’s an example they show on their website with 27

23:58 different questions, one request each. On the left is Type Safe AI, on the right GPT-5.6 Tera. And you can see after they finish typing this, the response is almost instant, while on the right we can see GPT-5.6 Tera still thinking about it. Now, it’s sort of streaming out its response. And that’s really the big difference here. You can see this cost, I mean, fractions of a fraction of a penny, and completed in .11 seconds, while this one cost a little more than a penny, and completed

24:29 in just shy of 9 seconds. You can also see the difference in output here. This outputted the text, this one outputted, you know, nulls, choices, and scores, along with the confidence score. Again, I’d like to do a deeper dive into this exact tool, but here are some cool use cases others have shared with it. Here’s a post from Malay here, where he created a smart drag and drop, and you can see he’s got a few different folders, work, travel, and receipts. And instead of dragging them by one, he’s able to select them all, drag them over, and he

25:01 created an app that automatically filters them into the right folders. Jonathan Yunakovsky here created something where it looks at his email inbox, and then instead of prioritizing by date, looks at how critical it believes it is, and sorts it by how critical the email appears to be. Dev Ed here uses it to moderate comments for him. So, running Jeff to see how fast it can remove negative comments from a chat, and it’s just automatically moderating if it believes it’s a negative comment. Steven Tae here made an app where you can plug in a

25:31 destination URL, and it detects whether or not it believes it’s a malicious URL before you go to the link. It can also apparently play video games. Now, this isn’t a model that’s going to go and develop a video game for you, but because it can make decisions so fast, it can decide whether it should jump, walk forward, walk backwards, attack, do things like that. So, it looks at the screen, decides what it needs to do, and then takes the action automatically. Again, not something that’s going to build the game, but something that’s actually capable of playing games based

26:02 on each frame that it’s looking at. And this is just scratching the surface of what this Jeff can do. Again, we’re going to play with it a little bit later. Now, unfortunately, if you do want to play with Jeff yourself, well, right now they have paused new sign-ups because of the overwhelming demand, so you can’t quite get into it. But, if you do have access, you can literally go to Codex or Claude code, tell it what you want Jeff to do, and let it write the code to run Jeff for you. Jeff also has a playground that you can use, but again, you need to be able to get

26:34 access, and right now access is currently closed off. Hopefully, it opens up again soon. And also, I will be doing a deeper dive into Jeff, so if I didn’t explain it very well, I’ll try to do better next time. All right, those are the biggest stories of the week, so let’s jump into a rapid fire. So, let’s start with Made on YouTube. This is YouTube’s annual event where they announce all of the new features and things they’re going to be rolling out. I put this one in the rapid-fire section because a lot of this is probably only going to be relevant to

27:05 other YouTubers. There are some other things that do apply to people just watching YouTube, but again, it is a lot of updates for YouTubers as well. One of the things they’re rolling out is custom feeds where you can essentially use AI and say, “Create me a custom feed of just AI news videos. Create me a custom feed of just stuff about the Padres.” You know, whatever you want, and it’ll create a custom feed that goes across the top of YouTube. And then you can click on that custom feed anytime you want, and it’s going to show just those videos. I think that’s a very, very welcome change to YouTube. I’m excited

27:36 for that one. There’s going to be Ask YouTube in search where, instead of just typing keywords, you can ask questions and it’s going to try to find videos that help answer those questions. They’re adding auto dubbing to live stream, so while you’re live, it can automatically dub your voice into other languages, so anybody from any other country can watch your live stream and understand what you’re saying without needing to turn on closed captioning. If you use YouTube Music, there’s some new features coming to that as well, including Ask Music, which is a conversational experience across podcasts and the catalog of 300 million

28:09 songs. So you can customize your queue or dive deeper into artist lore. Now, if you’re actually a YouTuber and you put videos on YouTube, you’re getting some new features as well. They’re adding features where you can use AI to help you create your thumbnails. They’re adding new editing tools that also leverage AI and Gemini Omni. They’re adding better likeness detection, so other people can’t like pretend to be you and make ads that seem like you made them. They’re adding some new AB testing features where you’re going to be able to AB test different versions of your video. So if you need like different

28:39 intros in a video, you’ll be able to split test those. And they’re also adding a feature where you can show different thumbnails to different audiences. So you have an audience of people that watch you like every week, you can show one thumbnail to them, and then you can have a completely different thumbnail that shows to people that have never seen your videos before. Those are some of the cool features that are coming. There’s a lot more announced, but I will link it up in the description. I think those are probably the biggest ones that most people care about, but there’s also a lot of, you know, smaller announcements as well. Microsoft is rolling out a new Copilot with home code and autopilot. Those are

29:11 like sort of three different modes you can switch it between. I have not had a chance to try this out cuz this was announced literally while I was recording this video. So, this is fresh in the moment I’m recording this, but you can see they’re sort of simplifying options like a quick response or think deeper. You can choose between GPT and Claude models. You’ve got chat and co-work, very similar to what OpenAI and Anthropic are doing, chat mode, co-work mode. You can see they’ve got a code mode up here, which would be the equivalent of Claude code or OpenAI’s

29:43 Codex. And then you’ve got an autopilot mode here, which would be sort of the equivalent of scheduled tasks and automations. But it can also take control of like all of the different Microsoft apps you use, Excel, Word, PowerPoint, things like that. And then the automation mode does things that you want it to keep on doing over and over again. Their example in this video is stay ahead of supplier delays. And it looks at Excel spreadsheets and things like that to stay ahead of it for you. Again, this is super fresh. I have not actually tested any of this out yet, but

30:15 it seems to be something where they’re trying to unify what ChatGPT and like Claude co-work do, where you have a chat, a code, and an automation sort of section, but then tie it in with the other Microsoft suite of tools like Excel, PowerPoint, Word, etc. Google rolled out Gemini 3.8 live with a live avatar. It basically pairs real-time video generation with your speech. >> Meet Gemini live avatar. >> Bringing real-time visual presence >> to Gemini’s conversational AI.

30:47 >> Watch the interactive demos below to see how native voice dialogue and instant video generation power expressive face-to-face experiences across various use cases. >> We also got a Gemini 3.8 text-to-speech. More like this. >> Introducing da da da, text-to-speech. Transforms simple prompts into consistent, expressive voice output on demand. >> Okay, now let’s add in another voice for this dialogue scene in the script. >> Or choose from over a thousand ready-to-go voices.

31:17 >> Okay, but what if we had that voice a little bit deeper? >> And change it however you want. >> Okay, perfect. >> I think we get the idea, but this is something that is available right now inside of Google AI Studio. Google also integrated Omni inside of their Google Vids platform. Google Vids was more for making like PowerPoint presentation style videos as opposed to like actual animated videos or realistic-looking videos. But now that you can use Omni inside of it, you can get more like dynamic animations, and it also provides

31:49 a free way to use Omni if you use it through Vids here. Spotify also rolled out a new feature this week with their taste profile. So basically, you can click on this button called taste profile. It’ll tell you what it thinks your tastes are already, but then you can sort of explain what you want it to change, and it will use AI to help you dial in your taste profile a little bit. And finally, there’s been a lot of talk about moving compute up into space. Well, this week Google is taking steps

32:20 to do that. This is called Project Sun Catcher, and Google is actually trying to put their tensor processing units, TPUs, up into space. So they’re going to use SpaceX to test and figure out if their hardware can operate in space. Right now, there’s still a lot of questions about like how to dissipate the heat and things like that. And well, the best way to figure it out, according to them, is let’s just start putting stuff up into space and see what happens. They need to figure out if the hardware can survive the actual launch process. They need to figure out if it

32:50 could stay cool in space. They need to figure out how to connect the various satellites to each other. That’s what this project Sun Catcher is and pretty soon we’re going to get some answers to some of these questions, it sounds like, which is pretty exciting. Ooh. All right, that’s what I got for you this week. I know, once again, really, really busy week. I normally record these on Thursdays and then publish them on Fridays. This week I’m recording on Friday and publishing on Friday, but I might have missed some news while in the process of recording and if I did, it’ll be in next week’s video. But every week I spend all week drinking from the

33:21 firehose, reading all the news, watching all the keynotes, going to the events, talking to the people building this stuff, and just trying to make sure I stay fully looped in and tapped in on what’s going on in the world of AI and sort of future technology so that I can make these videos for you on Friday and you don’t have to feel overwhelmed. You can just watch one video, feel totally looped in for the week, and I’m going to keep on doing that. So, if you like this kind of video, well, press that like button and consider subscribing to this channel and I’ll make sure more videos like it show up in your YouTube feed. Again, next week is OpenAI Dev Day, so I

33:53 will be doing even more travel next week going up to San Francisco for that. We can expect some big announcements. I have some ideas, but not a lot to talk about them yet, but there should be a lot of cool stuff coming next week as well. So, again, make sure you’re subscribed so you get those videos when they come out talking about everything that was announced during next week’s festivities. But again, that’s what I got for you today. Thank you so much for hanging out with me and nerding out with me. Sorry this video is a little bit later in the day than I normally publish them. That’s what happens when I’m doing

34:23 a lot of travel, unfortunately, some of these videos get pushed out. But I appreciate you for watching this and hanging out with me and letting me nerd out about all the cool stuff I came across this week. Again, I really, really appreciate you and hopefully I’ll see you in the next one. Bye-bye.