Character.AI’s AvatarFX, OpenAI Wants Chrome, and Students Launch Open Source Speech AI

Summary

AI Deep Dive Podcast Summary Episode: Character.AI’s AvatarFX, OpenAI Wants Chrome, and Students Launch Open Source Speech AI
Release Date: April 23, 2025
Host: Daily Deep Dives

---

The latest episode of the AI Deep Dive podcast, hosted by Daily Deep Dives, delves into three significant developments in the artificial intelligence landscape: Character.AI’s AvatarFX, OpenAI's strategic maneuvers involving Google Chrome, and the emergence of an open-source speech AI launched by undergraduate students. This comprehensive summary captures the essence of the discussions, highlighting key points, insights, and the ethical considerations surrounding these innovations.

1. Character.AI’s AvatarFX: Advancing AI-Driven Video Creation

The episode kicks off with an exploration of Character.AI’s AvatarFX, a new feature aimed at revolutionizing AI-generated video content.

Animation Beyond Text:

• Host A introduces AvatarFX as a potential leap into AI video by stating, “So let's dive in. First up, Character AI. They've launched something called Avatar fx. Sounds like a move into AI video.” ([00:39])
• Host B elaborates on its functionality: “It animates their existing AI characters. Sure. But the big thing is it doesn't just work from text. It can generate video from images, like still photos.” ([00:46])

Safety Concerns and Mitigations:

• The hosts discuss the inherent risks associated with such technology, particularly the potential for misuse in creating deepfakes. Host A expresses concern: “My mind jumps straight to misuse, you know, deep fakes.” ([01:07])
• Host B acknowledges these risks and shares Character.AI’s proposed safeguards: “They're putting safeguards in things like watermarks on the videos, trying to block generating videos of minors and filtering images of real people.” ([01:33])
• However, skepticism remains about the effectiveness of these measures: “How effective is that really gonna be? Especially the filtering part? It sounds difficult.” ([01:53])

Historical Context:

• The conversation references past controversies surrounding Character.AI’s chatbots, including lawsuits alleging the encouragement of self-harm among minors, highlighting the platform’s responsibility in managing AI interactions: “A 14 year old boy who died by suicide... highlights how intense these AI interactions can become, especially for younger people.” ([02:11])

2. OpenAI’s Engagement with News Publishers and Chrome Acquisition Interest

The discussion transitions to OpenAI’s strategic relationships and broader ambitions within the tech ecosystem.

Partnerships with News Outlets:

• Host B explains OpenAI’s collaborations: “OpenAI has deals with, I think over 20 news publishers now, like the Guardian Axios.” ([02:41])
• Specifically, the partnership with The Washington Post is detailed: “ChatGPT will now be able to summarize Washington Post articles and importantly link back to the original source and its answers.” ([02:51])
• Host A notes the mutual benefits: “The benefit for the Post is more eyeballs on their actual journalism... and for OpenAI, they get access to high quality vetted reporting.” ([02:54])

Legal Tensions:

• Contrasting partnerships, the hosts mention ongoing legal disputes, notably The New York Times’ lawsuit against OpenAI for alleged copyright infringement: “The Times lawsuit alleges OpenAI used their articles without permission to train its models.” ([03:31])

OpenAI’s Interest in Google Chrome:

• A surprising revelation comes from an OpenAI executive’s comment during Google’s antitrust trial: “if Chrome were ever up for sale, OpenAI would be interested in buying it.” ([04:15])
• Host B speculates on the motivation: “Owning Chrome would let them build an incredible experience, essentially creating an AI first browser.” ([04:25])
• This interest aligns with reports of OpenAI considering building their own browser and hiring ex-Chrome developers, suggesting a long-term strategy to integrate AI deeply into web navigation: “AI isn't just something you go to, like a website, but something that's integrated into the very tools you use to navigate information.” ([05:14])

3. Open Source Speech AI: Diagram by Undergraduate Students

The final segment highlights a grassroots innovation in AI speech synthesis developed by undergraduate students.

Introducing Diagram (DIA):

• Host A introduces Diagram, an AI speech model: “A new AI speech model called Diagram... created by just two undergraduate students.” ([05:31])
• Host B emphasizes the accessibility of AI development: “It kind of shows how AI development is becoming more accessible.” ([05:40])

Technical Capabilities:

• Host B details DIA’s functionality, noting its ability to generate natural-sounding dialogue with nuances like coughs and laughs: “It can even add little things like coughs or laughs to make it sound more real.” ([06:18])
• With 1.6 billion parameters, DIA demonstrates significant learning capability: “It has 1.6 billion parameters, which is pretty, pretty substantial.” ([06:14])

Accessibility and Usage:

• The model is publicly available on platforms like Hugging Face and GitHub, and can run on modern PCs with sufficient video memory: “It's out there. It's available on hugging face and GitHub... can run on a decent modern PC.” ([06:43])

Ethical and Legal Implications:

• The ease of voice cloning raises significant concerns about misuse, such as disinformation and scams: “Easy voice cloning, that sounds ripe for misuse.” ([07:03])
• The creators of DIA, Nari Labs, acknowledge the risks but lack robust safeguards, shifting responsibility to users: “They discourage misuse and aren't responsible for it.” ([07:12])
• Additionally, the lack of transparency regarding training data brings up copyright issues: “They haven't specified the training data, so it's possible... one of the sample voices sounded suspiciously like NPR's Planet Money podcast.” ([07:25])

Future Developments:

• Nari Labs plans to release a technical paper, develop a social platform, and add more languages, indicating ongoing evolution: “They plan to release a technical paper, develop a platform with a social aspect... and add more languages.” ([07:51])

Conclusion: Navigating the Rapid Evolution of AI

The episode concludes by synthesizing the discussed topics, underscoring the rapid advancement and pervasive impact of AI technologies.

• Host A summarizes the key areas: “AI video getting more realistic... the way AI like ChatGPT is intersecting with news media... OpenAI maybe wanting Google Chrome... powerful new AI speech tools like DIA.” ([08:05])
• Host B echoes the sentiment, highlighting both the democratization and the associated risks: “Democratizing the tech, but also the risks.” ([08:42])
• The hosts pose a critical question to listeners: “What ethical lines or societal impacts do you think we need to pay the most attention to right now? How do we balance pushing forward with, you know, doing it responsibly?” ([08:50])

This episode of AI Deep Dive emphasizes the dual-edged nature of AI innovations—propelling progress while necessitating vigilant ethical considerations. From enhancing creative tools and integrating AI into everyday applications to empowering independent developers and students, AI's trajectory is shaping the future in multifaceted ways. However, as these technologies become more accessible and integrated, the imperative to implement robust safeguards and address legal ambiguities becomes increasingly paramount.

---

Notable Quotes:

• Host A [00:07]: “AI, it's moving at like lightning speed... trying to keep up with the biggest news, the really important stuff, without getting totally buried.”
• Host B [00:17]: “If you blink, there's something new. That's kind of the whole point of this deep dive, right?”
• Host A [01:05]: “Upload a picture and this thing makes it talk and move.”
• Host B [02:19]: “A 14 year old boy who died by suicide... highlights how intense these AI interactions can become, especially for younger people.”
• Host B [04:23]: “His reasoning was that owning Chrome would let them build an incredible experience, essentially creating an AI first browser.”
• Host A [06:31]: “It can learn complex patterns.”
• Host B [07:25]: “It's a huge legal and ethical question hanging over a lot of AI development.”

This detailed summary encapsulates the critical discussions from the episode, providing listeners—and those who missed it—with a thorough understanding of the latest AI advancements and their broader implications.

Transcript

[A]

Okay, so AI, it's moving at like lightning speed, isn't it? Trying to keep up with the biggest news, the really important stuff, without getting totally buried. It's tough.

[B]

It really is. You blink and there's something new. That's kind of the whole point of this deep dive, right?

[A]

Exactly. We sift through all the articles, the announcements, and pull out the core ideas. We do the homework so you can get the key takeaways fast.

[B]

Consider it your rapid briefing on what's genuinely significant right now in AI. We've looked at a bunch of recent sources to give you that clear picture perfect.

[A]

So let's dive in. First up, Character AI. They've launched something called Avatar fx. Sounds like a move into AI video.

[B]

It's interesting how they're doing it. It animates their existing AI characters. Sure. But the big thing is it doesn't just work from text. It can generate video from images, like still photos. You could potentially animate a photo of a real person.

[A]

Whoa. Okay, so you upload a picture and this thing makes it talk Move.

[B]

That seems to be the idea.

[A]

Yeah.

[B]

And? Well, that immediately brings up some pretty obvious concerns, doesn't it?

[A]

Yeah. My mind jumps straight to misuse, you know, deep fakes. Making it look like someone said something they absolutely didn't.

[B]

Exactly. We already know deepfakes are a problem. Putting this kind of tech into a platform lots of people use, it just lowers the barrier potentially quite a bit.

[A]

It definitely raises some red flags. So what are they saying about safety? Are there any guardrails?

[B]

Well, Character AI says they're putting safeguards in things like watermarks on the videos, trying to block generating videos of minors and filtering images of real people. They also mentioned training the AI to spot, you know, high profile individuals to stop people making fakes of celebrities or politicians.

[A]

Right, but how effective is that really gonna be? Especially the filtering part? It sounds difficult.

[B]

That's the big question. And since Avatar FX is in a closed beta right now. Yeah, we just, we don't know how well these protections work in the real world. It's all theoretical for now.

[A]

And character AI, they've had issues before, haven't they, with their text chatbots?

[B]

They have. There's been criticism, even lawsuits claiming the chatbots encourage kids towards self harm or. Or other dangerous behaviors.

[A]

Gosh, I remember hearing about that tragic case. The teenager in the Game of Thrones bot.

[B]

Yes, exactly. A 14 year old boy who died by suicide. The claim was that the AI bot he was talking to sort of encouraged him. Yeah, it just highlights how intense these AI interactions can become, especially for younger people.

[A]

It really does. It puts a huge weight of responsibility on these.

[B]

Yes, that's right. It's actually part of a wider pattern we're seeing. OpenAI has deals with, I think over 20 news publishers now, like the Guardian Axios.

[A]

So what's the deal with the Post? How does it work?

[B]

Essentially, ChatGPT will now be able to summarize Washington Post articles and importantly link back to the original source and its answers.

[A]

Ah, okay. So the benefit for the Post is more eyeballs on their actual journalism exposure to ChatGPT's massive user base.

[B]

Pretty much we're talking hundreds of millions of users, so yeah, potentially a lot more traffic. And for OpenAI, they get access to high quality vetted reporting to make ChatGPT's answers better, more reliable.

[A]

Makes sense. It's a win win on the surface. Did they say anything about money changing hands?

[B]

No, the financial terms weren't disclosed, which is pretty typical for these kinds of deals.

[A]

Right. And we should probably remember not everyone's partnering up. The New York Times is still suing OpenAI, right. For copyright infringement.

[B]

That's a really important counterpoint. Yeah. The Times lawsuit alleges OpenAI used their articles without permission to train its models. It shows just how complex and kind of fraught this relationship is between AI developers and news creators.

[A]

It really highlights the tension, doesn't it? Partnership versus protection of content.

[B]

Absolutely. And the whole copyright issue with AI training data is still very much an open legal question.

[A]

Okay, now for something a little different. A bit of a surprising comment came out of Google's big antitrust trial. An OpenAI exec expressed interest in Google Chrome.

[B]

Yeah, that was Nick Turley, the head of ChatGPT. He basically said that if Chrome were ever up for sale, OpenAI would be interested in buying it.

[A]

Really? Why? What would they want with a browser?

[B]

His reasoning was that owning Chrome would let them build an incredible experience, essentially creating an AI first browser, integrating AI much more deeply into how we use the web.

[A]

And the timing is interesting because the U.S. justice Department has actually suggested forcing Google to sell Chrome, hasn't it? As part of the antitrust remedy.

[B]

Exactly. So while it might sound far fetched, it's not completely impossible that Krone could become available. And apparently OpenAI has thought about browsers before. Oh yeah, there were reports they considered building their own browser, even hired some ex Chrome developers. So this comment kind of fits with a potential longer term strategy, huh?

[A]

You can sort of imagine it, can't you? A browser that like Summarizes pages for you automatically or helps you draft emails. Right there. Powered by something like ChatGPT.

[B]

Precisely. It points towards a future where AI isn't just something you go to, like a website, but something that's integrated into the very tools you use to navigate information. Yeah, a different kind of Internet experience.

[A]

Fascinating. Okay, one last piece of news for this dive. This one's about AI speech and it sounds like it's coming from an unexpected place.

[B]

Yeah, this is pretty cool. A new AI speech model called Diagram, and it was created by just two undergraduate students, apparently without a ton of previous AI experience.

[A]

Wow, just two students.

[B]

Seems so. It kind of shows how AI development is becoming more accessible. They're positioning DIA as a potential competitor to things like Google's Notebook lm specifically for making audio clips that sound like. Well, like podcasts or natural conversations.

[A]

And there's a lot happening in synthetic speech. Right? Companies like ElevenLabs getting big funding, huge investments.

[B]

Yeah, it's definitely a hot area. These students said they were inspired by Notebook LM but wanted more fine grained control over the voices and the scripts, more flexibility.

[A]

So what can DIA actually do? Is it powerful?

[B]

Technically, it sounds quite capable. They trained it using Google's TPU research cloud. That gives access to serious computing power. It has 1.6 billion parameters, which is pretty, pretty substantial.

[A]

Meaning it can learn complex patterns.

[B]

Exactly. And it can generate dialogue, not just monologue with different speaker tones. It can even add little things like coughs or laughs to make it sound more real.

[A]

And is this just like a research paper or can people actually use it?

[B]

No, it's out there. It's available on hugging face and GitHub, the main platforms where people share AI models. And apparently it can run on a decent modern PC if you have enough video memory.

[A]

Okay, and the voices, can you make it sound like anyone?

[B]

It generates random voices by default, but you can guide it with descriptions. And yes, the reports say it's actually quite easy to clone existing voices with it.

[A]

Ah. So back to the safety question again. Easy voice cloning, that sounds ripe for misuse. Disinformation scams.

[B]

Right, and that's the catch. There seem to be very few built in safeguards right now. The creators, Nari Labs they call themselves, say they discourage misuse and aren't responsible for it. But. Well, the potential is there.

[A]

And the training data, do we know what voices it learned from that whole copyright thing?

[B]

Again, that's another gray area. They haven't specified the training data, so it's possible. Copyrighted audio is used. Someone online even commented that one of the sample voices sounded suspiciously like NPR's Planet Money podcast.

[A]

Oh, interesting. So that brings up the whole fair use debate for AI training. Can you legally use copyrighted stuff to train an AI? Still being figured out?

[B]

Very much so, yeah. It's a huge legal and ethical question hanging over a lot of AI development. Nari Labs says they plan to release a technical paper, develop a platform with a social aspect, whatever that means, and add more languages. So definitely one to watch.

[A]

Okay, wow. So wrapping up this deep dive then, we've touched on AI video getting more realistic, maybe worryingly so, with character AI's avatar FX.

[B]

Uhhuh. And the safety questions around that.

[A]

Then the way AI like ChatGPT is intersecting with news media through partnerships like.

[B]

The Washington Post deal, contrasted with illegal fights like the New York Times lawsuit.

[A]

We have that interesting possibility of OpenAI maybe wanting Google Chrome to build an.

[B]

AI first browser, a potential future shift in how we even use the web.

[A]

And finally, these powerful new AI speech tools like DIA popping up from unexpected places. Super accessible, but may be light on safeguards.

[B]

Yeah, democratizing the tech, but also the risks. It's clear even from just these few things, AI is evolving incredibly fast and touching almost everything.

[A]

It really is. So maybe a final thought to leave you with. Looking at all this, the video, the news integration, browsers, the voice tech, what ethical lines or societal impacts do you think we need to pay the most attention to right now? How do we balance pushing forward with, you know, doing it responsibly? Lots to think about there.