Perplexity’s TikTok Plan, OpenAI & MIT on AI's Emotional Impact, and AI Censorship by Language

Summary

Summary of "AI Deep Dive" Podcast Episode: "Perplexity’s TikTok Plan, OpenAI & MIT on AI's Emotional Impact, and AI Censorship by Language"

Released on March 24, 2025, the "AI Deep Dive" podcast by Daily Deep Dives explores significant developments in the artificial intelligence landscape. This episode delves into Perplexity’s ambitious plan to revamp TikTok, examines the nuanced dynamics of AI censorship across different languages, discusses the competitive shifts in the AI industry with Kai Fu Lee's Zero1AI, and investigates the emotional implications of using AI chatbots like ChatGPT.

---

1. Perplexity’s Bold Vision for TikTok

The episode opens with an in-depth discussion about Perplexity’s revolutionary plan to overhaul TikTok, aiming to transform it into a more transparent and trustworthy social media platform.

Key Points:

• Rebuilding TikTok: Perplexity intends to "totally rebuild TikTok from the ground up" ([00:52] B), focusing on addressing existing concerns related to privacy and algorithmic bias.
• Transparency and Open Sourcing: A cornerstone of their strategy is "building the algorithm completely in the open, maybe even open sourcing that whole For You feed" ([02:00] B), promoting algorithmic transparency.
• AI Integration for Enhanced Performance: Utilizing Nvidia Dynamo, Perplexity’s AI system is poised to "make TikTok's recommendations like a hundred times faster" ([02:00] B), enhancing user experience by making the platform's content suggestions nearly instantaneous.
• Fact-Checking Features: Incorporating Perplexity’s fact-checking capabilities directly into TikTok videos will serve as a "built-in BS detector" ([02:37] A), ensuring information reliability.
• Personalized Cross-Platform Experience: By linking Perplexity and TikTok accounts, users can enjoy a "personalized experience across both platforms" ([03:48] B), seamlessly integrating their search and social media interactions.

Notable Quotes:

• Speaker B [00:17]: "Our mission today is to sort of cut through all the noise and give you the important bits so you can see what's really going on and why it matters."
• Speaker A [02:22]: "So more waiting around for the algorithm to catch up. That would be a pretty noticeable difference for users, right."

Implications: Perplexity’s initiative could redefine social media by merging robust AI-driven recommendations with stringent transparency and fact-checking, potentially fostering a more informed and trustworthy online community.

---

2. AI Censorship: Language Matters

The podcast delves into the complexities of AI censorship, highlighting how language influences the extent and nature of content moderation.

Key Points:

• Differential Censorship Practices: AI models exhibit "different behavior depending on whether you're speaking English or Chinese" ([05:55] A), raising concerns about inconsistent content moderation.
• Case Studies and Experiments: Developer xlr8herder conducted tests revealing that even American AI models like Claude 3.7 Sonnet adjust their responses based on language input ([05:42] B).
• Impact of Training Data: The limited availability of uncensored Chinese text data leads AI to naturally "avoid certain topics when they're talking in Chinese" ([06:54] A), similar to learning from a biased textbook.
• Expert Opinions:
  • Chris Russell ([07:03] B): Highlights that "the ways we build these safeguards don't work the same in every language."
  • Gautam Vagrant ([07:27] B): Emphasizes that AI "just learn patterns from the data they're given."
  • Jeffrey Rockwell ([07:55] B): Points out that AI may "miss the point when it comes to, like, subtle criticism in Chinese."
  • Martin Sapp ([08:19] B): Discusses the dilemma of balancing "universal values versus culturally specific approaches."

Notable Quotes:

• Speaker B [06:00]: "So same AI, but different behavior depending on whether you're speaking English or Chinese."
• Speaker A [07:21]: "So the very methods we use to prevent harmful content might not be equally effective across all languages."

Implications: This segment underscores the challenges in creating universally ethical AI, highlighting the necessity for culturally and linguistically aware safeguards to ensure consistent and fair content moderation across global platforms.

---

3. Kai Fu Lee’s Zero1AI and the Open Source Challenge

The discussion shifts to Kai Fu Lee’s AI startup, Zero1AI, and its strategic pivot towards open-source Deep SEQ models as a competitive alternative to proprietary systems like OpenAI’s ChatGPT.

Key Points:

• Adoption of Deep SEQ Models: Zero1AI is transitioning to "using Deep SEQ models, which are open source" ([09:35] B), positioning itself against mainstream AI providers.
• Industry Impact: Lee describes Deep SEQ as "the ChatGPT moment for China" ([09:29] B), indicating significant traction among Chinese CEOs and the broader tech ecosystem.
• Economic Viability: By leveraging open-source models, Zero1AI boasts much lower operational costs, with Deep SEQ’s expenses being "like 2% of OpenAI’s $7 billion expenditure in 2024" ([11:33] B).
• Strategic Focus on Niche Markets: Zero1AI aims to "customize these Deep SEQ models for different industries like finance, gaming, and legal" ([10:09] B), capitalizing on specialized applications rather than broad, general-purpose AI.

Notable Quotes:

• Speaker B [09:44]: "So it's making waves in the Chinese tech scene big time."
• Speaker B [11:45]: "Deepseek is like infinitely lasting because of its funding and low operating costs."

Implications: Zero1AI’s approach signifies a potential shift towards more affordable and customizable AI solutions, challenging the dominance of high-cost, proprietary AI models and fostering a more competitive and diversified AI industry.

---

4. Exploring AI’s Emotional Impact on Mental Well-being

The episode concludes with an exploration of the psychological effects of interacting with AI chatbots, based on collaborative research by OpenAI and the MIT Media Lab.

Key Points:

• Study Methodology:
  • OpenAI's Approach: Analyzed nearly "40 million ChatGPT conversations" and conducted user surveys ([12:57] B).
  • MIT Media Lab’s Experiment: Conducted a controlled study with "almost a thousand people using ChatGPT for four weeks" to assess impacts on loneliness, social interactions, and dependency ([13:06] B).
• Findings:
  • General Use: "Most people don't really get emotionally involved when they're using ChatGPT" ([13:53] A), using it primarily as an informational or task-oriented tool.
  • Voice Feature Impact: Users engaging with the "voice feature... saw ChatGPT as a friend" ([14:06] A), with mixed effects on well-being based on usage frequency ([14:13] B).
  • Conversation Type:
    • Personal Conversations: Linked to increased feelings of loneliness but decreased dependency on ChatGPT ([14:29] B).
    • Non-Personal Conversations: Associated with greater dependency and potentially unhealthy usage ([14:39] B).
• Expert Commentary:
  • Researchers caution that findings are preliminary, noting limitations such as lack of peer review and reliance on self-reported data ([15:25] B).

Notable Quotes:

• Speaker A [14:29]: "That suggests it's not just about the quality of the interaction, but the quantity, the sheer amount of time spent engaging with the AI that can impact well being."
• Speaker B [15:07]: "And of course, everyone's different. People who tend to get attached easily, who see the AI as a friend and who use it all the time, those are the ones who might have more negative experiences."

Implications: The research highlights the nuanced relationship between humans and AI chatbots, emphasizing the importance of mindful usage and the need for further studies to fully understand the long-term psychological effects of AI interactions.

---

Conclusion: Interconnected Threads Shaping AI’s Future

The episode synthesizes the discussed topics, illustrating how advancements in AI technology are intertwined with societal, cultural, and psychological dimensions. From reimagining social media and addressing global censorship challenges to fostering industry competition and understanding human-AI interactions, the podcast underscores the multifaceted impact of AI on our world.

Final Reflections:

• The rapid evolution of AI requires continuous dialogue and critical examination of its broader implications.
• Ethical considerations, cultural sensitivities, and psychological well-being are paramount as AI becomes increasingly integrated into daily life.
• The future of AI development lies in balancing innovation with responsibility, ensuring that technological progress benefits society as a whole.

Closing Quote:

• Speaker B [16:47]: "The future of AI is something we're all shaping together, whether we realize it or not."

This episode of "AI Deep Dive" provides a comprehensive overview of pivotal AI developments, offering listeners valuable insights into how artificial intelligence is reshaping various facets of our lives.

Transcript

[A]

Wow.

[B]

AI is just moving so fast these days. It's like every time you turn around there's some new thing that's changing how we live online.

[A]

Yeah, it could be hard to even keep up with it all.

[B]

Exactly. And that's why we're doing this deep dive today. We're going to look at all sorts of cool stuff from rethinking like a major social media platform to. To figuring out how AI gets different languages and then even how these AI things are affecting us personally. Our mission today is to sort of cut through all the noise and give you the important bits so you can see what's really going on and why it matters.

[A]

We really want to highlight those key trends, the ones that are going to shape the future.

[B]

Okay, so let's jump right in. Have you heard of Perplexity?

[A]

Yeah, they're the ones with that answer engine, right?

[B]

They're all about giving you good, accurate information. Well, they just came out with this really bold plan for get this TikTok in America. You know, the one with all the short videos.

[A]

I think everyone knows TikTok at this point.

[B]

Uh huh, probably. But Perplexity wants to like totally rebuild it from the ground up.

[A]

That sounds like a huge undertaking. What's their thinking behind that?

[B]

They're basically saying there are, you know, some big concerns about how TikTok works right now and they think they can fix them.

[A]

Okay, so what are they proposing?

[B]

Well, the biggest thing is they want to make everything totally transparent. Like they're talking about building the algorithm completely in the open, maybe even open sourcing that whole for you feed. They want to make TikTok the most neutral and trustworthy platform out there.

[A]

Wow, that's a pretty radical idea. Open sourcing the algorithm. That's something you don't hear every day, especially with all the talk about how biased these algorithms can be.

[B]

Yeah, and that ties into their whole plan to have all the data and servers right here in the US under US oversight. They say that'll make sure everything's on the up and up with our privacy rules and all that.

[A]

Makes sense, but how are they going to actually rebuild TikTok?

[B]

Here's the really interesting part. They want to use AI, their own AI, to make it super powerful. They've got this system based on Nvidia Dynamo, which is made for really, really fast computing, and they're saying it could make TikTok's recommendations like a hundred times faster. The for you page would practically read your mind and know what you want to watch instantly.

[A]

Wow. So more Waiting around for the algorithm to catch up. That would be a pretty noticeable difference for users. Right.

[B]

And it's not just about speed. Perplexity wants to add their fact checking stuff right into TikTok videos so you can see if something's true right then and there.

[A]

That's a cool idea. Sort of like a built in BS detector for TikTok.

[B]

Exactly. Plus they want to use AI and feedback from the community to make sure the most reliable info gets highlighted. Kind of like how they have that at Ask Perplexity account on X. So basically they're saying, hey, we can make social media trustworthy again by being open about the algorithm and giving everyone the tools to check the facts.

[A]

It's a pretty ambitious vision. It could totally change how we think about those short videos. You know, instead of just entertainment, it could be a place to actually learn real verified stuff.

[B]

Totally. And think about this, Perplexity has that answer engine and TikTok has this massive library of videos. They're saying you put those together and you get the best search experience ever. You get answers with citations and everything. Pulling from both Perplexity's knowledge and all those TikTok videos.

[A]

That's a pretty cool idea. Combining like a traditional search engine with all that user generated content. I could see that being really powerful.

[B]

Oh, and it gets even better. Perplexity wants to make TikTok more than just like a time waster. They want people to feel like they're learning something, that their time is well spent.

[A]

So it's not just about keeping people glued to the screen, but actually giving them something valuable in return.

[B]

Yeah, exactly. They're even saying you could link your perplexity and TikTok accounts and it'll personalize your experience across both platforms. Like what you're looking for on Perplexity could change what you see on TikTok and the other way around.

[A]

Huh, that's interesting. So it would really tie those two platforms together in a personalized way.

[B]

And get this, they're talking about using AI to like automatically translate and annotate videos so everyone in the world can understand them.

[A]

That's huge for accessibility, especially for topics that are super niche or have a lot of jargon.

[B]

Imagine watching a video about quantum physics and then bam, you can just ask a question right there and get a deeper explanation without having to go anywhere else. Going from just watching to like really diving into the topic. That's powerful stuff.

[A]

It sounds like they're really trying to blur the lines between passive consumption and active learning. That's a Pretty exciting prospect, especially in a world where information overload is already a problem.

[B]

Okay, so now let's switch gears a bit and talk about something a little more tricky. Have you heard about this whole thing with AI censorship and how it's different depending on what language you use?

[A]

I've seen some headlines, but I haven't had a chance to dig into it. What's the story?

[B]

Well, we know that in China they have AI models that are trained to censor stuff, you know, politically sensitive Topics. Back in 2023, they even had a law that said AI can't say anything that, you know, hurts the country's unity or anything like that.

[A]

Right. So there's a very clear line drawn there in terms of what's acceptable for these models to generate.

[B]

Yeah, but here's the thing. It turns out that how much the AI sensors can actually change depending on whether you're talking to it in English or in chines. There's this developer on X, they go by xlr8herder, and they made a test to see how different AI models respond to questions about the Chinese government. You know, like criticisms and stuff. They tried asking in both English and Chinese.

[A]

So they were basically probing the limits of the censorship built into these models.

[B]

Yeah, exactly. And the results were kind of shocking. Even some American models, like Claude 3.7 Sonnet, would give you different answers depending on the language.

[A]

So same AI, but different behavior depending on whether you're speaking English or Chinese.

[B]

Exactly. And then There's Alibaba's Quinn 2.5 model, the big one with 72 billion parameters. It was super careful when you asked in English, but when you switched to Chinese, it would only answer about half of the tough questions.

[A]

So it was more willing to push the boundaries in English.

[B]

It seems that way. Even Perplexity's model, the one they call uncensored, wouldn't answer a lot of questions in Chinese. So this is kind of scary, right? The stuff we build into AI to keep it safe might not work the same way in different cultures or languages.

[A]

Yeah, that raises some pretty serious questions about how we approach AI safety on a global scale. If the safeguards are culturally and linguistically relative, then what does that mean for the universality of ethical AI development?

[B]

Exactly. The theory is that a lot of the Chinese text that these AIs learn from is already censored, so they just naturally learn to avoid certain topics when they're talking in Chinese.

[A]

That makes sense, right? If the AI has mostly been exposed to censored data, then his responses will likely Reflect that. It's like learning from a biased textbook.

[B]

Right. And some real experts are backing this up. Chris Russell, who works at the Oxford Internet Institute, said that the ways we build these safeguards don't work the same in every language. Like, you could ask a question one way in English, and it's fine, but if you ask the same thing in Chinese, the AI might shut you down.

[A]

So the very methods we use to prevent harmful content might not be equally effective across all languages.

[B]

Exactly. And then there's vagrant Gautam. He's a linguist, and he said that these AIs are just machines that learn patterns from the data they're given. So if there's not a lot of, like, uncensored Chinese text out there, the AI won't know how to talk about those things. He gave this example of, like, to whom it may concern just a normal phrase, but the AI has to learn that specific pattern.

[A]

That's a great point. It really highlights how the AI's capabilities, what it can and can't say, are directly tied to the data it's trained on.

[B]

Oh, and another guy, Jeffrey Rockwell from the University of Alberta, he said that sometimes the AI translations miss the point when it comes to, like, subtle criticism in Chinese. Like, there are ways to say things that are critical without being direct, and the AI doesn't always get it.

[A]

It's almost as if the AI is learning the grammar of a language without fully grasping the cultural context and nuances embedded within it.

[B]

Yeah, like it's missing the subtext. And then Martin Sapp from AI, too, brought up this whole dilemma about, like, do we want AIs that are good at everything or AIs that are really tailored to specific cultures? He even talked about this thing called cultural reasoning, where the AI learns the rules of a language, but not the social rules behind it.

[A]

That's fascinating. So it's not just about translation. It's about understanding the cultural norms and values that shape how language is actually used in different societies.

[B]

This all gets back to that big question of, you know, who gets to decide what these AI things can say and in what language?

[A]

Exactly. It really forces us to rethink some fundamental assumptions about how we design and develop AI, especially when it comes to ethics and responsible use. Do we strive for a universal set of values, or do we acknowledge the need for more culturally specific approaches? And how do we even begin to navigate those complexities?

[B]

Okay, let's move on to something completely different. Now let's talk about Kai Fu Lee. He's a big name in tech used to run Google China and now he's got this AI startup called Zero1AI. Well, they're doing something really interesting. They're switching over to using Deep SEQ models, which are open source.

[A]

Kai Fu Lee, he's definitely someone who knows what's going on in the AI world, so this move is definitely noteworthy.

[B]

Yeah, he even said Deep Seq is like the ChatGPT moment for China, meaning it's a really big deal there. A ton of Chinese CEOs are super excited about these Deep Seq models.

[A]

So it's making waves in the Chinese tech scene big time.

[B]

And Lee thinks these free open source models are going to be a huge problem for OpenAI and their business model. He even said something like Sam Altman's worst nightmare is that his competitor is free. Apparently people are ditching their ChatGPT subscriptions because Deepseek is free.

[A]

That makes sense. If you can get similar functionality for free, why pay for it, right?

[B]

So 01AI is focusing on customizing these Deep SEQ models for different industries like finance, gaming and legal stuff. Lee's basically saying that if you want to make money from AI, you got to focus on specific applications, not just the general stuff. He thinks those big pre trained models only work if you have like millions and millions of users and that's only something a few huge companies can do.

[A]

So he's betting on a more niche approach, leveraging the power of these open source models, but tailoring them to meet the specific needs of different industries.

[B]

Yeah, exactly. But he did say 01 AI isn't making a profit yet. They're predicting big revenue in 2025 though, like way more than they made in all of 20 for.

[A]

Okay, so they're playing the long game, betting on future growth. But there's always that risk with startups, right?

[B]

Totally. And it's not all smooth sailing for deep seek either. OpenAI and Anthropic have been telling the US government to like restrict access to Deep Seek, saying they might be state controlled. But Lee thinks that's just paranoia.

[A]

It's interesting how these competitive pressures, especially when it comes to such cutting edge technology, can quickly become intertwined with geopolitical concerns.

[B]

Yeah, it gets messy fast, but Lee thinks things are going to settle down eventually. He predicts we'll end up with just a few companies making those big pre trained models, but most of the development will be open source.

[A]

So a more collaborative and decentralized approach to AI development in the long run.

[B]

Yeah, he really seems to think that's where Things are headed. And he brings up this crazy stat. OpenAI spent something like $7 billion in 2024. Deep Seek's costs are like 2% of that.

[A]

Wow, that's a massive entrance. It makes you wonder about the long term sustainability of these huge proprietary AI.

[B]

Right, and that's exactly Lee's point. He doesn't think it's just about, like, which model is a little bit better. It's about whether OpenAI can keep up with the costs. He says Deepseek is like infinitely lasting because of its funding and low operating costs. So yeah, he doesn't think Sam Altman is sleeping very well these days.

[A]

Haha. It definitely paints a picture of an industry that's in flux, where the economic realities of developing and maintaining these powerful AI models are becoming as important as the technological advancements themselves. It'll be interesting to see how it all plays out.

[B]

Okay, ready for one last deep dive? This one is about how using ChatGPT affects our mental well being. OpenAI and MIT Media Lab are working together to figure out what happens to us socially and emotionally as we use these AI chatbots more and more.

[A]

That's a really important area of research. We need to understand the potential impact, both good and bad, of these technologies on our psychology and how we interact with each other.

[B]

Right. They're saying that ChatGPT isn't meant to replace real human relationships, but you know, people might start using it that way because it's so conversational.

[A]

It makes sense. If you're interacting with something that feels responsive and engaging, it's natural to develop a sense of connection, even if it's with an AI.

[B]

So they did two studies. OpenAI looked at tons of real ChatGPT conversations, like nearly 40 million of them. And they also did surveys, being careful not to like, invade anyone's privacy. And then MIT Media Lab did this controlled study where they had almost a thousand people use ChatGPT for four weeks. They wanted to see how it affected things like loneliness, how much people talk to others, whether they felt dependent on ChatGPT and if they were using it in, you know, unhealthy ways.

[A]

So two very different approaches to setting the same issue. Analyzing real world usage and conducting a controlled experiment. That's a smart way to get a more comprehensive picture.

[B]

And the results are pretty interesting. They found that most people don't really get emotionally involved when they're using ChatGPT. They're mostly just using it to get info or do tasks.

[A]

So for the majority of users, it's a functional tool rather than an emotional outlet.

[B]

Right. But there was a small group, especially the ones using the voice feature a lot, who did get really attached to ChatGPT, even seeing it as a friend.

[A]

It's fascinating how the mode of interaction, voice versus text, can influence the level of emotional engagement.

[B]

Oh, and speaking of the voice feature, they found that it can be good or bad depending on how much you use it. If you only use it for a little bit, it might make you feel better, but if you use it every day for a long time, it might not be so good for you. And it didn't seem to matter if the voice was like super realistic or not.

[A]

That suggests it's not just about the quality of the interaction, but the quantity, the sheer amount of time spent engaging with the AI that can impact well being.

[B]

They also found that the type of conversation mattered. Personal conversations where people shared more feelings were actually linked to more loneliness. But they were also linked to less dependence on ChatGPT and less unhealthy use, at least if you weren't using it too much. And conversations that weren't personal seemed to make people more dependent on ChatGPT the more they used it.

[A]

It's a complex picture, isn't it? The content of the interactions and the frequency of use seem to play a role in the psychological outcomes.

[B]

And of course, everyone's different. People who tend to get attached easily, who see the AI as a friend and who use it all the time, those are the ones who might have more negative experiences.

[A]

It really highlights the importance of considering individual differences and psychological predispositions when we talk about the impact of AI.

[B]

Oh, and the researchers were really careful to point out that this is just the beginning of their research. It hasn't been like officially reviewed by other scientists yet, and they were mainly looking at ChatGPT and people in the US who speak English. A lot of their findings were just correlations too, meaning they don't prove that one thing causes another. Plus, they were relying on people to honestly report their experiences. And their methods for figuring out like, the emotional tone of the conversations might not be perfect.

[A]

It's important to acknowledge those limitations, of course. But even with those caveats, the findings are still really valuable and provide a lot of food for thought as we navigate this new landscape of human AI interaction.

[B]

So what do we learn today? We've talked about totally reimagining a huge social media platform using AI. We've seen how even the smartest AIs can be biased depending on the language they use. We've talked about how the whole AI industry might be changing, with open source becoming a big deal. And finally, we've gotten a glimpse into how using these AI chatbots can actually affect how we feel.

[A]

It's amazing to see how all these threads are connected, how the development of AI is not just a technological story, but a social, cultural, and even psychological one.

[B]

Exactly. And it's moving so fast. So here's something to think about. What are the most important questions we should be asking right now? About how AI is going to change our lives and how we get information.

[A]

It's a conversation that we all need to be a part of. The future of AI is something we're all shaping together, whether we realize it or not.

[B]

Thanks for joining us for this deep dive. We'll be back soon with more ways to help you stay ahead of the curve in this crazy world of AI.