Meta Won? Alibaba's "Seedance Killer" & AI Audio Levels Up!

Author: Theoretically Media · Watch on YouTube

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: Meta Won? Alibaba’s “Seedance Killer” & AI Audio Levels Up! · Channel: Theoretically Media · 16:37 · Watch on YouTube

Topics: AI agents, AI content generation, AI models

Summary

The video surveys a packed week in AI, from competing model releases by Anthropic and OpenAI to Meta’s AI assistant and hardware announcements. It highlights upcoming video models from Alibaba and Kling, new real-time world simulations, and advances in AI audio. A sponsored segment demonstrates Artlist’s reusable creative workflows, while the closing feature looks at Jordan Rudess’s AI improvisation project, JamBot. The host frames these developments as part of a broader race to make AI a useful assistant and a more interactive creative partner.

Key points

The host says Anthropic’s Opus 5.5 feels nicer to use and is more efficient than Opus 5, while OpenAI’s Soul 6 is cheaper but can feel overly concise.
OpenAI connected voice interactions to apps including Mail, Calendar, and Slack, moving closer to an everyday AI assistant.
Meta’s Connect announcements included new VR and Ray-Ban glasses, plus Muse Charm, a voice-enabled AI device with 5G connectivity.
Alibaba previewed a video model expected in November, described as a successor to the WAN model, with longer generations and improved consistency and control.
Screenshots reportedly indicate an updated Kling Omni model with video generations of up to 120 seconds at 720p and 1080p.
Odyssey’s Agora 2 lets AI agents and human players share a Diablo II simulation, while PixVerse R2 adds active prompting within generated worlds.
Artlist Flows offers reusable node-based templates for creative workflows, and the platform also supports MCP connections.
Seed Audio 1.5 is expected to debut on Pippit; Fish Audio Drama 3 is available in preview for text-to-speech generation.
Jordan Rudess and the MIT Media Lab developed JamBot, an AI improvisation tool trained to generate real-time musical responses based on Rudess’s playing.

Chapters

Start	Chapter	Summary
00:00	Model rivalry and the AI assistant race	The host compares Anthropic’s Opus 5.5 with OpenAI’s Soul 6 and discusses OpenAI’s voice integrations.
02:34	Meta Connect and AI hardware	Meta’s assistant and new devices take the spotlight as the host considers the race to become users’ AI companion.
04:38	Upcoming video models and world simulations	The video covers Alibaba and Kling previews, Google’s Nano Banana update, Odyssey’s Agora 2, and PixVerse R2.
10:17	Sponsor segment: Artlist Flows	The host demonstrates how Artlist’s reusable node-based workflows can support recurring creative tasks.
12:53	AI audio updates	The host previews Seed Audio 1.5 and tries Fish Audio Drama 3’s text-to-speech model.
15:03	Jordan Rudess’s JamBot	The closing segment introduces JamBot, an AI improvisation research project developed with the MIT Media Lab.

Quotes

I’ve been playing around a lot with 5.5. It’s just a lot nicer to work with. — 01:02

the next battleground for all of these companies is to become your new AI best friend. — 04:38

this is sort of a sandbox environment — 07:41

it can be a massive timesaver. — 12:23

not running from the technology, but actually figuring out a way to incorporate it into performance. — 16:37

Concepts and entities

Meta, Anthropic, Opus 5.5, OpenAI, Alibaba, WAN, Odyssey, Agora 2, PixVerse R2, Artlist Flows, Fish Audio Drama 3, Seed Audio 1.5, Jordan Rudess, MIT Media Lab, JamBot

Full transcript (original language)

00:00 Well, it’s been nothing short of a blockbuster week for AI with major releases and announcements from just about everyone, plus a slew of interesting stuff on the creative side, including a potential sea dance killer coming up in like the next month or so, more realtime video world models, uh some stuff on the image side, audio updates, and a look at what’s just around the corner. Kicking off with the Battle of the Titans. Uh this week Meta came storming back or I mean were they really here to begin with? Regardless, Zuck and Company

00:32 held their big Meta Connect event hyping up obviously Meta Muse. But before even that happened, the week kicked off with a pretty memorable day that really illustrated the Titanic slugfest between OpenAI and Anthropic with the release on the Anthropic side of Opus 5.5 and on the Open AI side with Soul and Luna 6 uh both dropping on the same day. Setting benchmark charts aside because I I mean honestly who believes benchmarks anymore, the overall vibe does seem to

01:02 be that Anthropic is so back with Opus 5.5 with it being able to compete on an intelligence level with Fable at a fraction of the cost and is 40% cheaper or at least more efficient than Opus 5. Also, as a pretty big win, uh 5.5 does seem to have removed that annoying personality that has plagued the Opus models since the demise of 4.6. Overall, I do tend to agree. I’ve been playing around a lot with 5.5. It’s just a lot nicer to work with. Meanwhile, on the

01:32 OpenAI side, Soul 6 does come in cheaper, actually significantly so, but the overall sentiment is that Soul kind of like taps out of projects a little bit quicker. Um, but it is cheap and it does ship. That said, uh, a lot of the complaints seem to be that it feels almost too concise, like almost to the point of it being a bit bland. That said, I don’t necessarily agree with the flat personality aspect of it. And that obviously that is something that you can tune up via the settings, but that also might have something to do with the fact that I’ve been leaning a lot more on

02:04 voice as an interface modality with chat and codecs. Uh, something that Anthropic is kind of sorely behind on. Interestingly, in the middle of all of this, OpenAI also finally connected voice to like mail, calendar, Slack, basically all of your apps, which always felt like a bit of an oversight to me to begin with, but does finally bring us closer to, you know, it being that AI assistant that is actually helpful in your day-to-day. I mean, at least more so than the new Siri. Do you guys remember when uh Alexa got that big

02:34 update and all of the Alexexas were supposed to be smarter and then we all tried it and we were like, “Yeah, so still pretty much Alexa, though.” there. That’s kind of my mini review of the new Siri. So, that persistent AI assistant is definitely something that OpenAI is aiming towards which does bring us to our other big story of the week because this week Meta held Meta Connect and you know brought their AI assistant Muse really storming into the hardware game. There is no question that Muse has been a bit of a blockbuster for Meta actually

03:06 overtaking the chatbt app in the Apple store and generally just kind of having a wider adoption amongst the normie crowd. But the bigger news coming out of Meta Connect seems to be focused on the hardware front with Meta unveiling well not only the new Meta VR glasses and from a form factor. I mean what this to me does this look like what the Apple Vision Pro should have been all along. And to top it off, this is coming in at the non-inssane price point of $1,299.

03:37 There were, of course, some new Rayban Meta glasses as well. But the thing that really seems to have caught the news cycle was the Muse Charm, which is kind of like a I don’t know, like an AI Tamagotchi like device with real-time voice. Uh, basically like Muse, but without a mobile device or a computer. This one apparently will have 5G connectivity as well. So like it’s a it’s a system that doesn’t even need to be tethered to your phone. Now while Meta does seem to have won the week with the gwow of the uh AI Tamagotchi. Um we

04:07 are still not done considering that OpenAI has dev day next week as it has been highly indicated that we will be seeing OpenAI’s AON which is uh you know basically their answer to Meta’s Muse and well Gro’s Grockbot. There are definitely a lot of rumors going around as to what we’ll see next week at dev day. Mostly I’m just curious to see how Open AAI decides to differentiate itself from the other systems. I’ll definitely have coverage of DevDay, but yeah, I mean overall any way that you look at it, it’s very clear that the next

04:38 battleground for all of these companies is to become your new AI best friend. Moving on with some quick hits from our side of the creative AI fence. Big news from Alibaba. uh at their Absera conference uh they showcased a new video model that will be dropping in November. And I mean taking a look at it, it does definitely look like it’s it’s in around like Cance quality. Um not a lot of details about this one thus far. It does appear to be like the new successor to

05:08 the WAN model uh and is likely tied to the launch of Quen 4. Although all we have to go on currently is some footage captured at at the event on a phone. Um it’s, you know, we’re basically looking at bootleg footage, but it does look very good. In terms of what we do know about this upcoming model, uh to be honest, it’s kind of like the standard new video model bingo card. Alibaba stated that it will have longer duration while keeping scene, character, and visual consistency, better control, and

05:38 more complete narratives, and smarter reasoning. No word on pricing as of yet and no word obviously on weights being open sourced. Interestingly, they did end the presentation with a bit of a uh prediction, I guess, uh that within three years, a native unified omnimodal model will emerge and experience will no longer be constrained by modal boundaries. Meaning that, you know, uh there won’t be a seance, there won’t be a minax H3, there won’t be a uh there’s just going to be sort of like one model, Quen, I’m presuming that they’re

06:09 obviously alluding to that will just do it all. Moving on. And well, speaking of new video models, looks like Clling is back on the menu. While I cannot confirm nor deny any of this, uh, Mars for Tech over at Dev Mode had some exclusive screenshots, which will include an updated Omni model and generations allowing up to 120 seconds, like 2 minutes, uh, of video generation at 720 and 1080. Now, I don’t think that I can say much more about this right now other than um well, I mean, make sure you hit

06:40 that subscribe button and I’ll see you soon. In other upcoming or actually released quick hits uh once again via our friends at testing catalog, uh it does look like Google is staging out Nano Banana 2.5 Flash. It is appearing for some people uh over in Google Flow. I don’t have access to it yet, but you know, we’ll take a look at it when it releases. But importantly, it is not NanoBanana 2 Pro, nor is it Nano Banana 3. I mean, I know, I know. Uh, like Google really seems to be stuck in flash

07:10 land right now. At least that seems to be the feeling that I’ve been getting since Google IO. So, uh, I I am hoping that we see Gemini 4 drop fairly soon and then hopefully a flood of big updates after that. Moving on with a fun one here. Odyssey has released Agora 2, Advancing Multi- Aentgent World Simulation, and you can try this one right now. So, what this basically is is a group of AI agents that can play a game along with you and other human players uh for a total of up to 20. And

07:41 yes, everyone is playing Diablo II. Now, the obvious immediate read here, of course, is like, well, great, it’s an AI MPC. But you know what Odyssey is really aiming here for is uh this is sort of a sandbox environment uh where the Agora 2 model understands the world of Diablo II and uh we get to learn how AI systems interact with one another and interact with us in a controlled simulation. Although, I’m not sure that teaching our behavior to an AI agent via Diablo is

08:14 the best idea since the lessons they learn will be uh when humans are threatened, they stab things, and when they’re feeling sad or low, they chug colorful drinks. Uh that said, if you’ve had a hankering recently to play Diablo II, and you know, you want to do so alongside some AI agents, the Odyssey link to Agora 2 is down below. Moving on to the real-time video and world model side of things. Uh, Pixverse has released R2. I haven’t gotten too much time to play around with this one as of yet, but I mean, I’ve been a fan of what

08:44 Pixver has been doing with their realtime models, and this one does look like a pretty big step up in terms of visual quality. But the bigger update here is that, you know, aside from the kind of like Genie 3esque control layout, is that you can like actively prompt within this generated world simulation, which does open up the door to some pretty interesting experiences in real-time world models, uh, such as the ability to chat with a character in that world. Alex Prompter captured a cool example of this. Let’s take a quick

09:15 look. No, I can’t prove that. I can’t prove it to myself and I’ve had 14 hours of nothing else to do. >> Now, I’ll note that it does seem like Alex sped that footage up a little bit. Nothing on him or anything. Uh, when I was playing around with R2, it did feel a bit on like the the sluggish, loady side. Uh, but again, we’re still pretty early in with this technology, but I do think that we’re starting to get to a

09:45 breaking point and something that starting to genuinely feel like it resembles a hollow deck, you know, minus the holograms, but it’s also a massive step forward from the insane wonk that we got in real time uh back in January of this year. Um, I mean, this this one still brings me so much joy. Um, yeah, but I mean this is where things tended to like randomly catch on fire and coherence was uh like let’s call it a suggestion, a very surreal suggestion. Moving on to the audio side of things, some really interesting stuff going on there. But first, a quick word from

10:17 today’s sponsor, Art List. So lately, while I know that everything has been going a little more the MCP route and we’re all spending more time generating in codecs or in claude, there are still like platform based workflows that are more efficient. So today we’re going to take a look at one of them from our friends at Art List who were kind enough to sponsor today’s video. They have introduced Art List flows which yes is a note-based murder board, but the angle here is that this is a reusable murder

10:47 board. The easiest way to get your head around this is to actually look at one of their templates. Uh because these are actually handy templates. We’ve all seen those templates like what am I going to use this for? Uh but these are actually really good. Uh generate a 3D title. Character sheet of course is uh something that is we run a lot. So obviously we have two nodes here. One a character image and the other a provided prompt. Uh both of which feeding into our output here in which you know we get this multiframe generation. Uh now all of this is of course editable and um you

11:20 can swap out any of the models that you want to use. A number of them here are offered with uh art lists unlimited plan as well. Now where things get interesting and customizable is that from here we can build upon this template. If you just rightclick you can add a prompt in uh and a video generator for example. Um, we can just connect our character sheet to an input here and then a text prompt because one of the things that I’ve been doing with character sheets lately is creating 360 video rotations of these character

11:51 sheets uh in this case in Cance 2.5 and then either providing you know this video reference as a reference or uh just basically taking screenshots of it. Uh, and actually this multiframe approach to character sheets, uh, this might be really interesting to try out. But ultimately, the important part is that this now becomes your template. Well, it’s my template, but you can steal it and and then it’ll become yours. Um, and you can, you know, rerun it with different characters. Just simply upload a different character here, hit run, and the same results or the same workflow will then trigger. Now

12:23 again, I know that not everyone is a fan of nodebased workflows. And I do want to point out Art List does have MCP connections, so you can connect Claude or Chat GPT to it. But I do think it’s worth it to take a little time and create a flow for something that you constantly come back to. I mean, it can be a massive timesaver. I mean, at least it saves time in arguing with Claude about what you did last time. So, if you want to try out Art List flows, it is available right now over on Art List.

12:53 The link is down below. And hey, my thanks to Art List for sponsoring today’s video. Moving on to some interesting updates in the world of AI audio. Uh first off, it does look like we’re going to be getting an update to Seed Audio by Dance’s audio generator, which I looked at uh a few videos back. Uh found it to be very good. I mean, uh again, think seance but audio. Uh here, let’s take a quick listen to 1.0. They’ll tell you the old wards never fail. They told me that, too. Until the night the stars went hunting.

13:30 >> You’re late, boy. Fire’s near out. >> Sorry. The steps are ice cleaned through. I >> Steps don’t kill men. The cold does. Feed it. So, Seed Audio 1.5, which apparently was initially announced as a 2.0 and then kicked down uh.5, it looks like it’s going to be officially debuting over on Pippet, which I did not realize was another Bite Dance platform. By dance, why do you have to make things so confusing? I will be keeping an eye on this one. One of the major sticking

14:02 points of like the seed dance video model is well, I mean, besides the cost is the fact that audio is not its strong point. Uh, and 1.0 uh seed 1.0 out does really seem to help on that front. In the meantime, for audio that you actually can use today, uh Fish Audio has released in preview Drama 3. Uh they are calling this the most controllable text to speech model ever. I’ve actually been a pretty big fan of the Fish model as it tends to do emotive readings really well and well, it’s multilingual.

14:32 So, uh let’s take a quick listen here. Just turn off the lights. Stay away from the windows and whatever you hear out there, do not open that door. >> I mean, yeah, overall sounds pretty good. Uh, if you want to give this a run, uh, this is available actually right now. Uh, even on the free plan. So, I I don’t know how many generations you get per free, but um, yeah, I mean, I I just logged in and it was here. So, uh, under model, just go to, uh, the preview for Fish Audio Drama 3 and, uh,

15:03 prompt away. Rounding out with a musical project that I’ve been tracking for I guess the last two years. Man, has it been that long? My pal Jordan Rutis. Yes, that Jordan Rutus, the Virtuoso Dream Theater keyboardist who can play guitar and keyboard at the same time. [music] I mean, that’s not even fair. I can’t even wrap my head around that. Anyhow, he has spent the last few years working with the MIT Media Lab creating JamBot, or like I like to call it, Jordanbot, which is an AI improvisation tool that

15:33 is trained off of his playing. Now, what’s interesting about this approach is that this isn’t like a diffusion model trained off of recordings of [music] Jordan’s work. That’s something that we would see with like a product like SNO, but rather this is a model that has been trained to think like a Jordan, generating out uh real-time music and responses based off of, I guess, like you know, the notes and phrasings that he would most likely choose. [music]

16:06 So, while this isn’t a product, it’s kind of more of a research thing. Although I would love to be able to download Jordan bot and make it suffer through me playing Dance of Eternity at 70 BPM. They are doing a concert over at MIT. So if you’re in the area on October 3rd, uh this is this will feature Jordan uh and Jamot Jordanbot playing together along with live musicians and a full choir. Now, I know I’m preaching to the choir on this one, but what I really

16:37 appreciate about what Jordan is doing here is, uh, you know, obviously not [music] running from the technology, but actually figuring out a way to incorporate it into performance. So, uh, if you want to check out the concert, again, it is at MIT on October 3rd. If you do go, just don’t be the person screaming freeird, although I’m pretty sure Jordan Bot can play it. So, that’s it for today. Uh, you know, AI fall is really heating up and smells like pumpkin spice. [music] As always, I thank you for watching. My name is