AI News: Claude Opus 4.8, Insane Omni Use-Case, and A Dog Translator?

Here's the AI news you might have missed this week. Oh, and by the way, it might look like I'm in my normal studio right now. Wow. I'm actually in a hotel room here in Los Angeles. I'm out here for the Colin and Samir Press Publish LA event and definitely not shooting parts of this video on two separate days. Let's start with probably the biggest news of the week, which is the brand new model from Anthropic in their new Claude Opus 4.8 model.

Now if I'm being totally honest, this is probably a fairly minor upgrade. I don't think most people are going to notice a huge difference from Opus 4.7 to 4.8, but here's what Anthropic claims is different about this model. It's a little bit better at coding than 4.7, got a little boost on the benchmarks here, a little bit better at reasoning, very slightly better at computer use, and, you know, just fairly modest boosts on the benchmarks. However, they do say one of the most prominent improvements in Opus 4.8 is the honesty. It's now better at avoiding making claims that it can't support. It's apparently more likely to flag uncertainties about its work and less likely to make unsupported claims. But even on Anthropic's article here, it does say users will find Opus 4.8 to be a modest but tangible improvement on its predecessor. This new model is available in pretty much all of the platforms where you can use Anthropic's models, and the pricing is unchanged from 4.7. So it costs the same as the previous generation model with just a very slight, modest upgrade. Now, along with this announcement of the new model, they also introduced dynamic workflows inside of Claude Code. You probably don't care about this unless you're using it to write code, but essentially it's gonna help you take on more challenging tasks end to end when you're coding. So here's how this new feature works. When a workflow kicks off, Claude plans dynamically based on your prompt, breaks it into subtasks, and fans the work out across subagents running in parallel. Results are then checked before they're folded in, and then you come back to a single coordinated answer. Agents address the problems from independent angles. Agents try to refute what they found, and then the run keeps iterating until the answers converge. Again, this is a Claude Code update, so if you're using Claude Code to go write code for you, it's going to build out that sort of like agentic checklist to work through, spin off a bunch of subagents, get the subagents to all sort of independently work on the project, and then sort of double-check each other's work until they all sort of land on like, yeah, this is the best way to do the thing. Pretty cool improvement. Unfortunately, since I'm recording a chunk of this video from my hotel room, I haven't had a chance to test this myself, but it is something I will be putting through its motions and probably testing in future videos. But that wasn't the only news we got out of Anthropic this week. Anthropic just became the most valuable startup in history. This week, Anthropic raised $65 billion in their Series H round and was valued at $965 billion. Almost a trillion dollar company that's not publicly traded. Again, that makes them the most valuable startup in the world, surpassing OpenAI. But I have a feeling we're gonna sort of see these companies kind of leapfrog each other before the IPO happens. I wouldn't be shocked if we find out that OpenAI raises at a trillion-dollar valuation before their IPO. Who knows? It is just wild. These numbers to me aren't even real. Like, once you get up into like almost a trillion dollars, this just feels like all fake numbers at this point.

Moving along. Microsoft released a new version of their MAI image model MAI-Image 2.5. And according to the Arena.AI leaderboard, this model comes in at number 3 now. So Microsoft leapfrogged a ton of other models and is now the 3rd best model behind GPT-Image 2 and Gemini 3.1 Flash, aka Nano Banana. And here's what they improved about this new model. It follows instructions more closely, renders text more reliably, and shows strong visual reasoning across objects, scene structure, lighting, scale, and spatial relationships. They point out that it's especially good at text rendering and product and branding concepts. Now, it's not actually in the Microsoft Playground to play with yet, but if you go over to the arena.ai that I mentioned earlier, you can come up to the top, click on Direct Chat with One Model at a Time, select Image Model here, type MAI, and you've got the MAI Image 2.5 preview here. Here's a test image I just did with the prompt: Create a flyer for an event called Learn AI with the Wolf. The event is on July 4th, 2023. "Add some relevant copy and images related to this event." And we can see it added some fireworks, obviously referencing the 4th of July, and some wolf images, and it added a bunch of, you know, additional text and things like that. So with a very simple prompt, it built out this whole flyer. So pretty solid model. It understood the task I asked it and made a pretty good image.

Microsoft also introduced a new design for their Microsoft 365 Copilot this week. It seems like it's, uh, you know, a newer, longer prompt box that you get. You can also add things like bullet points directly into your prompt box. You can see it's got inline formatting. It also looks like it can draw directly from your other Microsoft apps, so it can look at your emails, files, chats, meetings, things like that, and pull in that data into the response. We can see here in this example, it'll create charts and graphs and things like that directly inline with the rest of your prompt. And Copilot is available inside of all of the various Microsoft 365 apps, you know, like Excel and PowerPoint and Word and all those kinds of apps as well, sort of bringing their product more in line with what we're getting out of a lot of the other products from OpenAI and Google and Anthropic and things like that. Now, next week is the Microsoft Build event, and I will be there in person, and we're probably going to get a lot more information out of what Microsoft has been working on. This is sort of just scratching the surface of what's new, but expect more bigger announcements out of Microsoft next week because that's when their big event is that they tend to announce all the interesting, exciting Since we were just talking about Microsoft, Perplexity is now available inside of Microsoft products. They announced that Perplexity Computer is now available inside of Microsoft Word, Excel, PowerPoint, and Outlook for Microsoft 365 apps. So if you want more like multi-step things to happen inside of your Microsoft products, you know, you can use the 365 Copilot to do very simple, normal chat-based stuff. But if you want it to be a little bit more complex, like prep the negotiation draft in Word by analyzing the latest redline against our standard template, propose track change edits and generate an issues list with recommended fallback clauses, then it will use Perplexity Computer to do that kind of stuff. Again, with Microsoft Build being next week, I'm going to be talking a lot more about a lot of Microsoft's rollouts. So I'm kind of going through these ones quicker because I know a big chunk of next week's video is going to end up being dedicated to all the cool things that Microsoft rolls out and talks about during Microsoft Build. So we will deep dive more into some of this kind of stuff in next week's video.

If you haven't heard about Hermes yet, maybe because all the AI drama lately has been around OpenClaw, this is your sign to start paying attention. Hermes might actually solve one of the biggest weaknesses other AI agents still have: their memory. You don't want your agent to just forget everything and restart with every new project. So Hermes has a built-in self-improving learning loop where it can actually create new skills from past experiences, refine workflows over time, and build persistent memory that transfers between sessions. And to run an agent like Hermes locally and have it keep working and improving while you sleep, you want to use something like Hostinger VPS to host it. Hostinger has a pre-configured Hermes Docker template, so it's super easy to install. You just pick the Hermes agent VPS, hit deploy, and it launches a production-ready environment in minutes. And because it's running on your own VPS, all your API keys, memory, conversations, and business context stay private on your own infrastructure instead of living on somebody else's servers. Once it's installed, you can connect Hermes across Telegram and Slack and Discord and WhatsApp and even email so it can do things like monitor my competitors every morning and send a summary to my Discord. If you want to run your own Hermes agent without dealing with all the setup headaches, check out Hostinger at hostinger.com/mattwolfhermes and use code Matt Wolfe for a discount at checkout. And thank you to Hostinger for supporting my channel and sponsoring this portion of today's video. Late last week, the company Leonardo AI rolled out a new feature with the ability to turn an image into a 3D model with AI. This seems like it would be pretty handy if you're creating 3D games or any sort of like 3D video and need to create NPCs or props or enemy mobs or things like that. It can also be used to create 3D objects of e-commerce products. So if you're selling clothing or jewelry, you can make a rotatable 3D version of it. And there's all sorts of other examples in their blog post here. But let's log into Leonardo and try it out real quick. The first thing we notice when we log in is a new 3D button right below our main prompt box. So I'll go ahead and click here. Now, it doesn't appear we can do any sort of text to 3D, but we can start with an image. If I click up on this image box and click Image to 3D. I could choose to upload an image or start from something I've already generated. Out of curiosity, I want to know what it would do if I used one of these wolves that I previously generated. So let's go ahead and select this one. I'll confirm it, and I'm just gonna leave everything on the default and click Generate, see what happens. Now it did take about 5 minutes to generate, so it's not the fastest thing in the world, but let's see what it looks like. I can open it up and I can drag it around, and as expected, it looks pretty good from the angle of the original image and decent from the other side. There's some weird kind of stuff going on with the face here, but not too bad. But they also have this 3D reference view creator here. So if I go up and click this Use This Blueprint button, and let's select this version of the wolf here, and then click Generate, it's going to generate multiple angles. And then I can pull in all of those different angles as references for the 3D object. And we can see it generated a top-down angle, a straight-on angle, a from-behind angle. Now if I close out of this, add our images here, I could select all of these images, confirm this, and now going to use all 5 of these as a reference and we can generate again. And let's see if this looks any better here. So when I open this one up, definitely, definitely a lot more detail when you run it through that process. Let's see what happens when I give it an image of myself here. I'm going to have it generate all the reference frames. And here's what it looks like when I bring myself in. I mean, it's a little bit cursed, especially in the face. But hey, we're getting there. The company ElevenLabs released a new music model called Music V2, which improves upon their previous music model. Here's an example song they shared.

Lab open late, every light on the block. Got the whole industry pissing round the clock. They've been promising a wave. I've been making the flood. Pen pressed to the paper while they sketching in mud.

And the model is trained on licensed data and cleared for commercial use. So, you know, they didn't just scrape a whole bunch of artists' music, train on it, and now we're just getting sort of ripped-off versions of other artists' songs. It's all music that they actually had the license to use. So if we go to 11labs.io/music, let's give it our own prompt. We can see we're on V2 here. High-energy upbeat pop punk song with the driving synth bassline, catchy vocal melody, and a powerful anthemic chorus. The song is about eating tacos in San Diego and watching Padres baseball. And it actually starts to give something that we can play pretty quickly. Let's listen. San Diego Knights! Interesting, because it's actually got some world knowledge that it's baking into this song as well. It definitely knew about Tatis and Petco Park and elements about San Diego. So not bad. Gives you another alternative to some of the other music generator platforms out there where you know that supposedly the music was trained ethically.

ElevenLabs also rolled out this really cool new Dubbing V2 feature. It allows you to take videos that you've already created and then dub them, but it's still gonna use your voice, your emotion, your facial expressions. And well, to be quite honest, it's gonna probably do a lot better than like the built-in YouTube dubbing. So I wanted to test this out. So I went to one of these recent videos that I made, which was like an AMA Q&A kind of video, and I plugged this one into ElevenLabs and basically asked it to dub the first 10 minutes for me. Right now they let you dub up to 30 minutes for free. Since this video was 47 minutes long, I didn't get to dub the whole thing, but I said let's just dub the first 10 minutes and see how it comes out. I had it dub it into Hindi, so this is what that looks like.

Now I don't speak Hindi, so I have no idea how well that did, but if you do speak Hindi, you could let me know if if that is better than what we would normally get from like YouTube straight up dubbing. Last week Google released their Gemini Omni model, and since then all sorts of cool use cases have been popping up. But I wanted to share the one that I've found the most fascinating that you can actually do. So I first saw this one from Chris First here, where he uploaded a screenshot of Google Maps to Gemini Omni and he drew a route on the map. Then he prompted it to create a first-person view of someone driving a taxi cab along the route in the reference image. So check this out, he uploaded this map, here Here's the route that he drew on the map, and then it actually generated this video of a car driving down the street in that exact route that it showed. Now it's only gonna generate 8 seconds here. I think it can generate upwards of 10 seconds, but it did a pretty decent job. And then my buddy Belovl here, he took that to another level. He gave Google Omni a sketched camera path and asked it to generate drone POV footage. Now I'll play it with audio to start because you can actually hear the drone. But if we look at this image here, he took this sort of like angled view and drew a map of where he wanted the drone to fly. And we could see the drone follows that exact path. We can see it flying by this tall building up in the top right towards the end of the video. We could see that it flew under the bridge like he asked it to. Here's that big building. Pretty cool use case. If you make short films and things like that and you need an establishing shot for something but don't have a drone to get the exact establishing shot you want, well, there's options to do that now. There's been a lot more updates this week, but most of them don't need like full deep dive breakdowns. So let me run through them real quick in a rapid-fire session. Starting with some new AI updates out of YouTube. They're actually changing how AI videos are disclosed. They're going to move that disclosure to a more prominent position. For long-form videos, the label will now appear directly below the video player and above the description. And for Shorts, the label will appear as an overlay on the video itself. But probably even more importantly is is they're going to introduce automatic AI detection. Right now, in the backend of YouTube, the creator themselves actually has to check a box to say that this was made with AI. And while most people just don't check that box, they say, while we still require creators to manually disclose when they use realistic AI, we want to make the process more seamless and reliable. Starting in May, we're rolling out new internal signals to help identify AI-generated content. If a creator doesn't specify whether or not they used AI, but our systems detect significant photorealistic AI use, We will now automatically apply a label. This is interesting to me because, well, a lot of times my intros are AI, but then the whole rest of the video isn't. So I'll be curious to see if my videos get flagged or not, but I'm gonna keep having fun regardless. This week the Pope actually gave a presentation on AI, and while giving the presentation, he was actually accompanied by none other than one of the co-founders of Anthropic. Now I wanted to share this post from Ollie Lehmann here because he does a great job of breaking down why this is actually important. Pope's only put out a handful of these huge official letters in their entire time as Pope. The fact that one of them is about AI tells you how seriously the church is taking what's coming. The Pope's main line: AI needs to be disarmed. He literally compared AI to nuclear weapons. We have more nuclear weapons than anybody else, but we don't want to send enriched uranium anywhere. We're not enriched uranium. It's a chip, and it's a chip that they can make themselves. The co-founder of Anthropic actually publicly admitted that every AI lab, including his own, faces pressure that can conflict with doing the right thing. Commercial pressure to keep shipping, competitive pressures from other labs, plus the older pressures of pride and ambition. His solution: we desperately need outside critics with no skin in the game who will tell the labs when they're failing. He says there's 3 giant questions the AI labs need to figure out. How do we make sure poor countries actually benefit from AI? What does human flourishing even look like in this new world? And what are these things we're actually building? Artificial intelligence needs to be disarmed. Sam Altman is starting to walk back some of his talks about the AI apocalypse and AI taking everybody's jobs. He said the rapid development and adoption of AI would not lead to a global jobs apocalypse and the technology had not claimed as many white-collar jobs as he had feared. He said, I'm delighted to be wrong about this. I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened. Even admits, people are like, oh, you could have saved the world a lot of fearmongering and a lot of doom and gloom. But at the time, I was like, I see this is a real risk we should probably talk about. And it still may be. Now, if we're being totally honest about this, yes, it is true that it hasn't taken as many jobs as people thought it would take, but it's still getting better and better, so it still could take a lot of jobs. Now, I don't know if Sam Altman is saying this because he genuinely believes it, or if he's saying this because, well, they're trying to IPO later this year and walking back some of this talk of this being a job-taking technology could be very beneficial for when the IPO comes. Along similar lines, NVIDIA CEO Jensen Huang is telling CEOs to stop hiding behind AI as the reason they're letting people go. He went on international television and called the AI layoff excuse lazy and irresponsible. One week after DeepMind CEO Demis Hassabis Alexibus said the same thing in different words. And again, if we're being totally honest with ourselves, most of the layoffs we've seen, they're using AI as an excuse, but the reality is much different from what I can tell. A lot of these big companies that are laying tons and tons of people off, they were bloated companies. There were companies that did a ton of hiring during 2020, 2021, during COVID They got way, way too big. Their expenses got way too high. Now they're realizing they could be much more profitable if they actually cut back. And instead of just saying, we're cutting back to be more profitable because we they don't want as much overhead. They're saying, oh, AI is doing a lot more of these jobs. They also believe this is positively going to impact their stock prices because if they say, hey, we're laying off a bunch of people because we got too bloated, well, that doesn't look good for stockholders. But if they say, hey, we're laying a bunch of people off because AI is making this much more efficient, well, stockholders now think they're geniuses. That's the theory, at least. The reality is actually much different. Many of the companies that are doing these AI layoffs are actually seeing stock declines. Investors just aren't buying the AI excuse. If you want to know if there's a proposed data center being built in your area, well, Erin Brockovich, who yes, famously had a movie made about her, launched a crowdsourced AI data center map that shows where there's operational data centers, where there's data centers under construction, where there's proposed data centers, and where there's community-reported data centers. If you go to brockovichdatacenter.com, you could scroll down and see where the data centers are being built. Interestingly, this site To me, has all of the markers of a vibe-coded website. Nonetheless, it's a good resource to find out if there's a data center proposed, under construction, or reported by the community anywhere near you. It looks like we might see more out of Apple's AI with WWDC coming up. Apple recently added a new subdomain, genai.apple.com, to its domain name servers this weekend. Now, the page doesn't point anywhere as of this recording, but it does sort of indicate that there's probably going to be something related to Gen AI announced at their WWDC event coming up on June 8th. A Chinese company claims that they launched a 95% accurate pet translator. The Chinese startup Meng Jiaoyi— I don't know how to pronounce that— has claimed that their new pet translator device genuinely works. It costs about $118. It's powered by Alibaba's Qwin language model and is worn around the pet's neck and can allegedly recognize their behavior, vocalizations, and emotions and translate them into human speech with an accuracy rate near 95%. I'm hungry.

I need snack.

Speak.

Hi there, my name is Doug Squirrel. And finally, another device out of China designed to use AI to cut your hair. Several cities in China are now rolling out AI-powered robot barber kiosks. The machine scans your head in 3D and cuts your hair with millimeter precision. It costs ¥60, less than a dollar per session. I don't know, I may just be old-fashioned, but I'm someone who I think I'm going to let a human still cut my hair for a little while. And that's what I got for you this week. Every week there is so much happening in the world of AI, and there's channels out there that will make a new video for every single update and make you think it is the biggest thing that's ever happened in the world of AI. I want to be sort of the signal through the noise, cut through the hype, and let you know just what I think is, well, interesting or important for the most amount of people to know. I'll drink from the fire hose all week. I'll keep up to date with everything. And then once a week in these Friday videos, I'll break down just the most important things that happened in the AI world. So hopefully it feels a little less overwhelming for people. If that's something that interests you, maybe consider liking this video and subscribing to this channel, and I'll make sure more videos like this show up in your feed. As a quick reminder, I record these on Thursdays, publish them on Fridays, and well, this week I was even more backwards. I recorded some of this on Tuesday, some of it on Wednesday from a hotel room, some of it on Thursday, and then published on Friday. But anything I missed in this video, I'll make sure I get into next week's video so you don't miss any important details or information. Again, I'm going to do my best to keep you looped in and help cut through all the hype. Thank you so much for tuning into this video and watching it this far. I really, really appreciate you, and hopefully I'll see you in the next one. Bye-bye. Van that he used to come— he said great for San— why is it great for B? Pure days.