267 | China vaults ahead: SeeDance 2.0 leaves Sora & Veo in the dust, and starts a deep-fake tsunami, Harvard: AI usage increases labor intensity, ReantAHuman.AI allows AI agents to rent human employees, and other big AI news for the week of Feb 13, 2026

0:00Hello and welcome to the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage AI to improve efficiency, grow your business, grow your business, and advance your career. Today's episode is going to be a little different as I'm going to deep dive even deeper into the main topics and there's not going to be enough time to go through all the rapid fire news. So I'm going to pick very few of them and even on those I'm going to be very short. So we have three main topics to talk about. One, the first two are somehow interconnected, which is the rise of China in this race.

0:36And in some cases, they are ahead of the US right now, which is the biggest news of this week. Then we're going to talk about deep fakes, and you'll see shortly why these two things are connected. And then we're going to dive into how AI is impacting the workforce in surprising ways. And then we're going to have time for a few rapid fire items, and we will be done for today. So lots of interesting things to talk about, not necessarily a lot of things to talk about. So let's get started.

1:12As I mentioned, our first topic is the rise of China. We've been talking about the race between the U.S. and China for a very long time. But for a very long time, U.S. models were ahead and the Chinese models were able to come very close or in very narrow cases with very narrow margins on a very specific topic. maybe beat the US models on one or two benchmarks, but overall, the US was ahead. The Chinese models' benefits was that they were significantly cheaper to create and significantly cheaper to run, which means they were a great alternative for the larger US models, especially that most or maybe even all the Chinese models are open source, which means you can run them on your own box, more or less safely, despite the fact they're coming from China.

2:00So that gave a great cheap alternative to companies or individuals who wanted to run powerful AI on a budget, but they were still behind. And there were many, many questions and many senior leaders of the US labs that says that they believe that China will break and will be able to lead in the race. They just weren't sure where and when. We heard Dario Amadei speak very bluntly about the fact that we should not be selling advanced and vinyat chips to China because that's going to help them close the gap and potentially pass us. We know that for now, NVIDIA has an approval to sell H200 chips, which are significantly more powerful than the chips that we're able to get before.

2:40The first batch is supposed to come online in the next few weeks. So that's going to be the first time the Chinese will be able to train models on more powerful NVIDIA chips. They're still not their latest chips, but there's still a very serious pushback against selling these chips to China, and there's a lot of discussions in the Capitol Hill as far as whether to cancel that capability for them to get these chips, at least legally. Well, for the Chinese New Year, we're getting more and more new Chinese models. One of the biggest news, which comes more or less exactly one year from the original DeepSeek moment, DeepSeek has just upgraded the context window of their current model from 128 ,000 tokens to 1 million tokens.

3:26That is a very big jump and it makes it the Chinese model with the largest context window there is out there right now, aligned with Opus 4.6 as an example. Very significant jump in the context window and if you compare it to the original DeepSeek version 3, so now it's 3.2, but the original DeepSeek version 3 had only 64 ,000 token context window 3.2 had 128 ,000 tokens and now it is not exactly clear but I think it's still the same model just with a much larger context window but DeepSeek version 4 is supposed to be released sometime in the next few weeks and potentially before this podcast even goes live maybe in the next few days and version 4 has been delayed for a very very long time.

4:12So we had different variants of version 3 for a while now and my guess is that DeepSeek is preparing something big and significant as far as their next model as they have done with version 3. It will be very interesting to see when this model comes out how good it is and does it really break ahead of US models. Another big release this week comes from a company called Zipu, which has released GLM-5, which is a massive large language model for a really small company. So it has 744 million parameters. It's an open source language model, and it is another private agency, so none of the big giant players, that is able to generate a solid large language model.

4:57It does not compete with the frontier US models, but it is aligned with the previous version of models from the US at a fraction of the cost. The other interesting parameter about this is this model was trained 100%, at least as far as what they say, on GPUs generated in China and not generated by NVIDIA. And so it is a capable model, again, not top of the line model, but a capable model that was trained purely on Chinese hardware and software capabilities. Minimax, which has been around for a while as a private smaller company has also released a new model called m2.5 this is the first top to bottom multi-modal model that is achieving really high scores on multiple benchmarks it is again not top of the line model but it is aligned with claude sonnet 3.5 and gpt 4.0 and these kind of models across multiple benchmarks and it is actually really really good at complex multi-step reasoning tasks better than that generation of US models.

6:05And it's released as full open weights and open source model that anybody can use however they wish. And like all the other Chinese models, it is significantly cheaper to run than most of the US models. That means, by the way, that China right now has three independent smaller labs, DeepSeq, Zipu and Minimax that are producing models that are not at the frontier but are pretty freaking close and again there's a very good chance that DeepSeek version 4 will be aligned or in many cases ahead of the top models in the world and that adds to obviously the big players in China like Alibaba and ByteDance that are doing really interesting things which is a great segue now if you're asking yourself how can really small labs generate models that are close to the frontier is twofold.

6:55One is they are training them significantly cheaper than the US models. If you remember the deep seek moment, there was discussions that they trained the entire model at a$5.6 million investment. It was obvious afterwards that that wasn't taking into account some other components that needs to be taken into account. And then they were distilling US models, But it doesn't really matter the fact that they are actually doing it significantly cheaper. And when I'm saying significantly cheaper, we're talking two or three orders of magnitude cheaper than the U.S. companies are training the large leading models right now.

7:28However, the other thing that became known this week is that the Chinese government funds that are managed by the U.S. government has poured over$15 billion to domestic AI companies just in the first quarter of 2026. And the first quarter of 2026 is not even over. The largest amounts went to DeepSeek and Zipu and the rest goes divided between most of the other companies. Zipu, by the way, is planning to go public at a 10 to$15 billion valuation, which is going to make it a very large IPO for a relatively small and young company. Now, by funding the innovation of AI in both hardware and software, where the Beijing government is ensuring that the Chinese lab can continue to advance the Chinese capabilities in the AI arms race.

8:18And they're obviously getting a cut in the eyes and a complete view into what's happening in these companies so they can use the technology for the government's needs. Now, not surprisingly, a significant chunk of that state funding goes to building domestic AI hardware, compute and GPUs that will enable China to run independently of the US. The Chinese government has described AI self-sufficiency as, and I'm quoting, core national security priority, and they understand exactly what the US government understands and any other government in the world right now, that the AI's race is going to play a very big role in future supremacy, and hence they're investing a lot of money and they will invest as much as they need in order to get that We also know that China has built a very significant capacity on the electricity generation side, and they are leading the world right now in that, which gives them extra bandwidth in order to be able to compete in a more aggressive way.

9:18On the US, by the way, there is a push currently in the White House to put out a bill that will force in a collaborative way the main labs to pick up the tab for any new electrical infrastructure in order to reduce to zero the increase in cost or issues to the grid by paying for any new electrical capacity that needs to come online. I think this is a very smart bill. It will be interesting to see how this actually turns out. But the biggest news of this week when it comes to China actually comes from ByteDance. ByteDance is the company behind TikTok. And ByteDance just released a new video model called Seedance 2.0, which is their next generation AI model.

10:07And this AI model is way ahead of the models that were released by the West right now. So it is the first time that a Chinese company is not releasing something that is maybe competing on specific benchmarks, but that is clearly both in means of the performance of the output as well as the underlying technology ahead of everything anybody else in the world has, including the leading companies in the US, such as Sora from OpenAI or VO3 from Google, etc. It is a very unique model because it has two very interesting things that, again, do not exist in any other model right now. The first one, it has a quad modality input system, which means you can provide inputs to the system in text, images, audio, and video all at the same time.

10:57Meaning, when you create a new scene or a new movie or a new segment, you can do that with different inputs to direct the model to do the right thing. The other half is at least as interesting because C-Dance 2.0 is the first model that knows how to process visual and audio streams in parallel at the same time, fusing them into one coherent and amazing output. So all the models that we know right now that include sound in their outputs, such as Vio and Sora and Runway and a few others, the way they do this is the sound gets added in post-processing. In other words, one model generates the video and then it goes back to the beginning and another model generates the audio and try to do the lip sync and the background noise and so on.

11:46And in C-Dance, it happens at once seamlessly, which means it understands exactly what's happening in each frame in order to generate the sound in real time as it's happening. This is a very significant jump forward. In addition, it comes with a very solid control over multiple parameters, which really takes you from a creator that needs to experiment a lot to a Hollywood director that controls everything that's happening in every single scene in a very detailed way. Bytance is claiming a 90 plus percent usable rate for generated clips, meaning that 9 out of 10 outputs require minimal or no editing for professional use.

12:30to give you an understanding what that means. The industry benchmark right now is between 20 to 40%, which means you have to generate a video between two to five times in order to get the output you want. I can tell you that in many cases, that is more than that because I've generated multiple videos. So going to a 90 % success rate and getting a clip to do exactly what you want, including sound is incredible. Now, when I tell you it generates sounds, it generates any sound that is required for the scene. That means synchronized lips for dialogues. It means ambient sound and music. All of it gets created when the video gets created.

13:09And the output is absolutely stunning. There's already thousands of videos online showing really amazing scenes generated by this model. And to make it even more attractive, it is also significantly cheaper than the US competition, such as Runway and VO and Sora. Now, this model is going to catch like wildfire for two different reasons. One, everything that I mentioned so far, it's an incredible model. But as importantly, ByteDance, as I mentioned, is the owner of TikTok that has potentially the largest creator community in the world. And these creators are eager to try these kind of tools. And that means that immediately across a massive creator base, we're going to get a flood of AI-generated videos that are indistinguishable from real life.

13:58Now, the combination of its ability to keep consistency of video angles, scene, characters, et cetera, together with low price and with full sound will mean that we're going to get most likely more AI-generated videos in the relatively near future and most likely this year on TikTok. And then I'm sure afterwards with all the other video platforms, then we're going to get real videos that are created and uploaded to these platforms. So why is this the number one topic for this week? There are two reasons. Reason number one is that it is a significant moment in the history of the AI race, because it is the first time that a Chinese company comes out with model that is much better very clearly both in the outputs as well as in the underlying technology than the leading US models.

14:54The other reason this is critical and it may be the most important part of this is the fact that now any person in the world with extremely low budget can generate any video of anything that is completely undistinguishable from real life. Now, this is becoming scarier and scarier. And to be fair, I felt like this three years ago. If you go back to episode 13 of this podcast, the episode was called The Truth is Dead. And what is the basic premise? The basic premise is that our society, any society, is built on trust, trust in communication. You trust that what people tell you, that what governments tell you, that what organizations tell you, that what the news and media tells you, that what books tell you is real.

15:43And the reality is right now, this trust is broken. You cannot know what is real and what is not on digital media, which means you cannot trust anything in communication. And that trust is the fabric of any society. So what's going to happen next? I'm not sure if you want to make it even more interesting, we're going into a mid-election year in the US, which means it will be very, very easy to convince people or sway people's minds in one direction or the other. Now, this is not new. Governments have done this for generations, and large companies did lots of large media campaigns to convince us in one thing or another.

16:26But now any individual and definitely small groups can generate massive amount of fake news and content to sway people in their direction. And that is really scary to me. Now, a little bit of the backlash or what's happening with C-Dance, with what's happening in that arena right now. Well, the first thing that happened is that the Motion Picture Association, the MPA, and SAG-AFTRA, which is the Screen Actors Guild of American Federation of Television and Radio Artists, and major studios are issuing formal legal action and very strong statements against the use of Seedance 2 of known characters, of known leading actors inside videos that are generated by Seedance 2.0.

17:18Maybe the most viral version is an incredible scene of Tom Cruise fighting Brad Pitt on a roof of a semi-collapsed building overlooking a town in the background. It is an incredible scene. It is practically impossible to know that it is AI generated and it is all generated by Sea Dance. So obviously there's a very strong protest from the leading Hollywood studios and the actors' guilds, et cetera, in what they are characterizing as an existential threat to creative professionals. In addition to their immediate cry, they are now working very aggressively and they're calling for immediate regulatory action to block models to illegally use the IP and the likelihood of leading actors and or cartoon generated in the case that is being generated by the models right now and especially by a Chinese model that has no rights to do so.

18:16The MPA issued a formal statement warning that AI video generation models pose a fundamental threat to intellectual property rights and the likelihood of creative professionals. That's the exact language. Fran Dretcher, the SAG-AFTRA president, called for an emergency negotiation with the technology companies, stating that members' likeness are being exploited without compensation or consent. Screenwriter Rhett Reese commented, and I'm quoting, we've crossed a line where anyone's face can be used to say or do anything and that should terrify every creative professional i would say as i mentioned before that should terrify anybody period in an interesting move disney announced it would pursue legal action against platforms that are hosting the ai generated content basically they know it will be very hard for them to go after the labs that are generating them especially if they're in china but if you're in the u.s and you want to watch this video and it's hosted somewhere, that would be illegal as well, or at least that's what Disney is trying to do.

19:21And I actually think that's not a bad idea. Now, this situation is fueling a no fakes act in Congress that is being now pushed forward from multiple angles. But this has definitely been an accelerator to that, which is a requirement from the federal government to protect the digital likeness of individuals and the IP of groups. and companies. Multiple states have already introduced similar kind of laws and yet limited to either the state or a specific thing, with California proposed a bill that specifically targets AI video generation tools that can create realistic depictions of real individuals without their consent.

20:01Now, going back to the movie and television industry, industry analysts now estimate with these new capabilities that these tools can eliminate 100 ,000 to 200 ,000 creating jobs in the industry within three years if it's not heavily regulated. Now, early last year, I predicted that by the end of last year, we're going to have miniseries that is created 100 % by AI, and that already happened as I predicted. I do think that we are not going to see a full featured long film this year, but I'm guessing we will see them in the next two to three years. And again, if somebody really aggressive is going to go at it, it might happen this year, but I think we're going to start seeing real videos that are going to make it to cinemas, or at least to streaming channels that are going to be 100 % generated by AI.

20:58That means that a a small group of creatives can take a budget of a few tens of thousands of dollars and create videos that will be competing, and in some cases, better than Hollywood productions at a fraction of the cost that Hollywood can do this. Why am I saying they could be better? Because the AI generation gives you a lot more freedom, creative freedom, to do whatever you want. That is really hard to do with real videos, and that is including the capability to create special effects, and digital manipulation and so on. Now, in the beginning, I think in the short term, people still want to see Tom Cruise and Scarlett Johansson and Gal Gadot and all the top leading actors in real videos.

21:42But I think over time, what people want to see, they want to see a movie they enjoy or a TV series they enjoy, something that will get them excited or sad and will create whatever emotion they want to have at that point. And I don't think over time they really care if the actors are humans or AI, as long as the movie or the TV series is great. So in the short term, people might be willing to pay premium for movies and or TV series that are made by and with humans. But sadly, and I'm saying that with deep concerns, I have a feeling that this is not going to stay the norm. And I think that AI generated videos, movies, TV series, etc.

22:22will provide much more flexibility and hence much more interesting plots and environments that can be created and with a fraction of the budget. And so if you remember last year when the actors and the screenwriters signed an agreement with the main Hollywood studios that is preventing them or limiting the amount of AI they can use, I said that is a worthless agreement because once people, other organizations, not the current Hollywood studios, will be able to create full featured films with basically a budget of tens of thousands of dollars. Hollywood is done. If the main studios will not follow suit, they will go out of business.

23:05And then it doesn't matter whether the actors had an agreement or not. They would still not be able to have a job and get a job. Now, do I think this is a good thing? No. Do I feel sad for the actors? 100%. Do I want to see the capability to generate movies and TV series with actual humans continue? I would love to see that. However, from a pure financial perspective, I don't see this as a realistic option for the longer run. Now, you want to get another example of this. Novelist Coral Hart has done an interview with The New York Times. She is currently writing under 21 different pen names, and she has produced more than 200 romance novels last year and self-published them on Amazon.

23:54Now, none of these books has been a hit on its own, but she has sold over 50 ,000 copies, raking six figures as an income from writing these books. Now, while being interviewed by Zoom, by the New York Times, she has finished producing a book in just 45 minutes. An average human writer doesn't stand a chance. That is what she said, and I agree with her. I had the opportunity to actually sit in a AI community discussion with somebody who does the same thing, not at the same scale. She didn't write 200 novels. She wrote a lot less. But she showed us step by step the exact framework that needs to be put in place.

24:39And there's a lot of work to create that environment. But once you have that environment, once you have the machine ready to go, you can create new books, maybe not in four or five minutes, maybe in two hours. maybe in two days. An average person takes months to write. And the same kind of story is going to happen for the movie and TV industry. And again, I'm not happy about it. I'm not saying it's a good thing. I'm just stating where we are right now and what is the trajectory and the outcome of the current trajectory is very clear. But my bigger problem is not even the actors guild or the writers and so on, but just the ability to commit fraud and generate fake stuff at scale with zero ability to distinguish from reality.

25:28And if you think about how we know what is real and what is not, it is because A, initially we assumed everything is real. I talked about the trust that we talked about before. But B, now even if you question that, what you're going to look for are clues that are going to tell you this is the real thing. As an example, if you're watching a piece of news, is it reported from multiple channels? and also you're going to check are there multiple videos of the event so as an example if you're going to see a report of something significant happening you will look and you will see multiple videos of that thing from a helicopter from the news media from people with their cell phones filming it on the street etc now with the new models even with the existing models if they don't get any better you can do this tools like cdance 2.0 allows you to create multiple videos of the same scene from multiple angles shot with different tools.

26:20So one of them can be stable and perfect like professional media. And then you can get four or five or 10 different angles of people shooting the same exact scene with their phones. And it is going to look completely realistic. And if people will put this across multiple channels, then the main media is going to pick it up as well. And then it is going to be reported as news. And it will be a while before somebody catches it is not news. And at that time, the harm is already done because even if they roll back and say, oh, sorry, this is not real, we picked up a fake story, not all the people that saw the original story will see that this is not the real story.

26:56And now you understand where this is going, not to mention people that can do actual real deepfake fraud that is happening at scale right now. So global losses from AI-enabled fraud reached$442 billion in 2025 based on a Deloitte's Global Finance Crime Report. Their projections are suggesting we're going to exceed$600 billion by 2027. Maybe the most known case which we reported on when it happened is the Arup incidents where criminals used real-time deepfakes. So you change the face and the voice of a person that is talking in real time on a Zoom call to impersonate a company's CFO. And that confirmed to a controller in a different country to release$25.5 million bank transfer to their bank account.

27:48Because the CFO that they know and that they work with, the one that runs the financial team that they belong to, that they've been on meetings on before, told them to do so as a payment to a specific client. Similar attacks have targeted financial institutions with one bank reporting 47 separate deepfake fraud attempts in a single quarter. Now, to tell you how bad the situation is, there's now this phenomena of DAS platform. If you're asking yourself what the hell DAS stands for, so think about SaaS as software as a service, DAS is deepfake as a service. These are platforms that are operating in the dark web, which are offering real-time face swapping, voice cloning, and document forgery tools for as little as$20 a month.

28:32This means that anyone with$20 a month can have access to really sophisticated fraud capabilities that previously belonged to very unique, to very sophisticated organizations, not individuals who had similar capacity and now anyone can get access to this. One security researcher described this ecosystem as, and I'm quoting, Amazon for identity fraud, complete with customer ratings and technical support. Microsoft Digital Safety Team has reported that humans' ability to identify deepfakes has dropped to approximately 25%. A survey of Fortune 500 companies found that 73 % have no formal deepfake detection protocols, and 89 % of employees could not reliably distinguish between AI-generated videos from unauthentic footage.

29:25Now, governments around the world are very slow to respond, and I'm not sure exactly how they can respond to this. The EU's AI Act now includes a specific provision for synthetic media labeling, while the US Federal Trade Commission has proposed rules that will require AI-generated content disclosure. However, there are several different problems with that. Problem number one is enforcement. You need to be able to catch the people who actually do this, which is not very easy, especially if it is running in the dark web. Problem number two, the people who are using it to commit fraud are already criminals.

30:00They're not interested in using the main platforms that may follow the rules that the US government or other governments or the EU puts in place that's saying they have to label such content as AI generated. People can take an open source model and run it on their own and remove whatever markings the original model has generated because it's open source and you can change it. And now you can still commit fraud despite the rules that exist today. You're breaking the law anyway. You're just going to break another one. So why did I want to dive into this topic? I wanted to dive into this topic because we're in a very fragile place in history right now because of this.

30:38I urge you, A, to go and listen to episode 13 and to see in more detail why I think it is so critical that we figure this out. But I truly do not know exactly how we can figure this out. I am very much for a very aggressive legislation and a collaboration between the labs to give us detection tools that actually work to identify what is AI generated and what is not. And then hopefully the combination of regulation and digital capabilities will dramatically reduce the impact and will dramatically reduce the risk of AI generated deep fakes and beyond. Now, one company is already leading the charge, and that is Microsoft.