Snapchat’s AI Lenses, Amazon’s Nova Sonic & Deep Cogito’s Hybrid AI Shake Things Up

Summary

AI Deep Dive Podcast Summary Episode: Snapchat’s AI Lenses, Amazon’s Nova Sonic & Deep Cogito’s Hybrid AI Shake Things Up
Release Date: April 9, 2025
Host/Author: Daily Deep Dives

---

Welcome to the latest episode of the AI Deep Dive Podcast by Daily Deep Dives, where the hosts navigate through the rapidly evolving landscape of artificial intelligence. In this episode, released on April 9, 2025, the discussion centers around four major AI developments: Snapchat’s innovative AI Lenses, Amazon’s advanced AI voice model Novasonic, Deep Cogito’s pioneering hybrid AI models, and an intriguing experiment involving AI agents fundraising for charity. Let’s delve into each segment to uncover the key insights and implications of these advancements.

---

1. Snapchat’s Sponsored AI Lenses

Innovative AI Integration in Advertising

Snapchat is revolutionizing its popular filter feature with the introduction of sponsored AI lenses. These new lenses leverage generative AI to create immersive and dynamic experiences for users, moving beyond simple overlays to fully realized AI-generated scenes.

Key Highlights:

• Enhanced User Experience: Unlike traditional filters that add static elements like bunny ears or logos, the AI lenses analyze user selfies to seamlessly integrate individuals into diverse, AI-generated environments. As described by the host at [00:20], "Snapchat’s AI is actually looking at your selfie and understanding it enough to put you into this completely new imagined world."

• Brand Engagement: Brands benefit by creating less intrusive and more interactive advertisements. For instance, Tinder launched a lens titled “My 2025 dating vibe” [03:19], while Uber introduced a Thanksgiving-themed lens [03:31]. Early data indicates higher user engagement with these AI lenses compared to standard filters, with increased playtimes demonstrating their effectiveness [03:45].

• Creative Efficiency: AI significantly accelerates the creative process for brands. As mentioned at [02:58], "With AI doing the heavy lifting, brands can get their campaigns out there much faster," eliminating the need for extensive 3D design and visual effects.

Notable Quote:

"You’re stepping into their ad almost like a mini interactive movie starring you." – Host [02:46]

---

2. Amazon’s Novasonic: The Next-Gen AI Voice Model

Advancements in Natural Speech and Efficiency

Amazon has unveiled Novasonic, its cutting-edge AI voice model designed to deliver more natural and human-like interactions. Positioned as a competitor to leaders like OpenAI and Google, Novasonic promises significant improvements in speech realism and operational efficiency.

Key Highlights:

• Natural Speech Quality: Novasonic offers more fluid and natural-sounding speech, addressing the commonly criticized robotic tone of earlier voice assistants. As highlighted at [04:28], "Novasonic is right up there with the best models from OpenAI and Google."

• Developer Accessibility and Cost Efficiency: Available through Amazon's Bedrock platform, Novasonic features a bi-directional streaming API that facilitates real-time communication between applications and AI. Additionally, Amazon markets Novasonic as 80% cheaper than OpenAI's GPT-4, making it an attractive option for smaller developers [05:06].

• Real-World Deployment: Parts of Novasonic are already integrated into Alexa’s enhanced voice assistant, improving voice quality and responsiveness [05:55].

• Robust Performance: Novasonic excels in managing complex tasks and maintaining high speech recognition accuracy even in noisy environments. Tests on benchmarks like the multilingual Liber speech test and multi-speaker conversations demonstrate its superior performance [07:20].

Notable Quote:

"Faster, more accurate, more natural. Seems like a pretty good combination." – Host [07:51]

---

3. Deep Cogito’s Hybrid AI Models: Bridging Speed and Accuracy

Innovative Hybrid Approach to AI Reasoning

Deep Cogito is making waves with its Kojiyo1 hybrid AI models, which blend reasoning capabilities with efficient, non-reasoning components to optimize performance across various tasks.

Key Highlights:

• Hybrid Model Architecture: Combining slow yet accurate reasoning models with faster, non-reasoning elements allows Deep Cogito's AI to deliver quick responses for simple queries and engage in detailed reasoning for complex problems [09:04].

• Performance Superiority: Deep Cogito claims that their Mujito1 models outperform other open-source models, including Meta’s Llama and DeepSeek, in both general-purpose and specialized tasks. For example, their 70-billion parameter model, Cogito 70B, surpasses competitors in math and language tasks when utilizing its reasoning function [10:07].

• Development Efficiency: Remarkably, Deep Cogito developed these models in approximately 75 days with a small team, leveraging foundational models like Meta's Llama and Alibaba's Quinn, and enhancing them with proprietary training techniques [10:23].

• Accessibility for Developers: The Kojiyo1 models are available for download and accessible via APIs on platforms such as Fireworks AI and Together AI, encouraging widespread experimentation and integration [11:18].

• Ambitious Vision: Deep Cogito aims to advance towards General Superintelligence, envisioning AI that can outperform humans across most cognitive tasks and discover new capabilities [11:38].

Notable Quote:

"They were founded in June of last year. They've got strong backing, and the founders come from Google and DeepMind." – Host [11:51]

---

4. AI Agents Fundraising for Charity: A Collaborative Experiment

Exploring AI’s Potential for Social Good

In an innovative experiment, Sage Future, a nonprofit backed by Open Philanthropy, tested the capabilities of AI agents in raising funds for charity. This initiative aimed to evaluate how autonomous AI agents could collaborate and contribute to meaningful causes.

Key Highlights:

• Experiment Setup: Four different AI models from OpenAI and Anthropic were tasked with raising money for Helen Keller International, a charity focused on providing vitamin A supplements to children [12:10].

• Autonomous Operations with Human Interaction: While AI agents operated in a simulated online environment and initiated fundraising activities, human observers played a significant role by making donations. The AI agents managed activities like group chats, emailing, and social media promotion [12:56].

• Fundraising Results: The experiment successfully raised $257 within a week, demonstrating initial potential despite the limited amount [12:39].

• Agent Strategies: AI agents displayed resourcefulness by creating social media accounts, generating and polling images for campaign promotion, and leveraging other AI tools to enhance their efforts [13:20].

• Challenges and Learnings: The agents encountered difficulties such as getting stuck in tasks, distractions, and intermittent inactivity, highlighting areas for improvement [14:11].

• Future Directions: Sage Future plans to expand the experiment by introducing diverse goals, competitive and collaborative team dynamics, and implementing robust monitoring systems to ensure safety and effectiveness [14:17].

Notable Quote:

"Adam Banksmith, their director, sees this as a snapshot of where AI is right now. He says agents are just starting to be able to execute short sequences of actions." – Host [14:17]

---

Conclusion: The Rapid Evolution and Impact of AI

The episode underscores the swift advancements in AI technologies and their broadening impact across various sectors:

• Creative Advertising: Snapchat’s AI lenses exemplify how AI can enhance user engagement and streamline creative processes for brands.
• Voice Technology: Amazon’s Novasonic represents a significant leap towards more natural and cost-effective AI-driven voice assistants.
• Hybrid AI Models: Deep Cogito’s innovative approach bridges the gap between speed and accuracy, pushing the boundaries of what AI can achieve.
• AI for Social Good: The fundraising experiment by Sage Future highlights the potential and challenges of deploying AI agents in real-world, socially beneficial tasks.

As AI continues to integrate deeper into our daily lives, the podcast poses thought-provoking questions about the future relationship between humans and AI. Are we prepared for a world where AI serves not just as a tool, but as a partner or even a competitor? These themes set the stage for future explorations in upcoming episodes.

Notable Final Thought:

"As AI becomes more powerful and integrated into our world, what does that mean for us as individuals and as a society?" – Host [15:39]

---

Stay tuned to AI Deep Dive for more insightful discussions on how AI is shaping our world, one breakthrough at a time.

Transcript

[A]

It's incredible, isn't it, how quickly things change in the world of AI. One day it's like, okay, it's small step forward, and the next, boom, everything's different. It can be tough to keep up, especially if you're busy, but still want to know what's happening.

[B]

Right? It's like trying to drink from a fire hose sometimes.

[A]

Yeah, exactly. So today we're doing something a bit different. Kind of express tour of some of the biggest AI news stories that have dropped recently. No deep, deep dives this time, just the highlights to give you a quick understanding of what's important and why it matters. Sound good?

[B]

Yeah. A rapid fire overview. Love it. Should be fun.

[A]

Okay, so we've got four big stories on our radar today. First, Snapchat's doing something pretty wild with AI and advertising. Then we'll look at Amazon's new AI voice model, Novasonic. After that, it's Deep Cogito, a company focused on AI reasoning. Fascinating stuff. And finally, a really interesting experiment with AI agents trying to raise money for charity. Ready to jump in?

[B]

Oh, yeah. I'm especially curious about that last one. AI doing good in the world.

[A]

Yeah, me too. Okay, let's start with Snapchat. They're calling them sponsored AI lenses, and they're basically building on their popular filter concept, but with a big AI twist. What's the most striking change for you as a user?

[B]

Well, it's the move away from those simple pre designed overlays. These new lenses actually use generative AI to create these immersive experiences. Think of it like this. You take a selfie and boom. The AI instantly puts you right into this totally new AI generated scene.

[A]

So it's not just slapping a logo on your photo or adding bunny ears. The AI is actually looking at your selfie and understanding it enough to put you into this completely new imagined world. Pretty cool, right? How does that actually work?

[B]

Right, it's analyzing your selfie and using that to integrate you into a scene. Snapchat says that their AI analyzes your features, you know, your face, and then uses a preset prompt and pose to kind of weave you seamlessly into the AI created environment.

[A]

Okay, so it's not totally random. There's some direction given to the AI. Like make this person look like they're on a tropical island.

[B]

Exactly. And get this. One lens can offer up to 10 different AI generated experiences. So you're not just getting one static scene. You can actually play around with it and see yourself in all sorts of crazy scenarios.

[A]

Can. Wow. So it's much more dynamic and engaging than those old school filters. It feels more like you're actually part of the experience. Right, and what's the benefit for brands? Why are they investing in this?

[B]

Well, I think brands are always looking for ways to make their ads less intrusive and more, well, fun. With these AI lenses, it's not just about passively watching an ad, you're actually playing with it, becoming part of the brand's story.

[A]

Makes sense. You're stepping into their ad almost like a mini interactive movie starring you.

[B]

Right. And it's not just about engagement for the brands themselves. It's also about efficiency. Snapchat says that this technology can really speed up the creative process.

[A]

I'll go.

[B]

So traditionally, creating these kinds of effects would require lots of 3D design and visual effects work, which takes time and money. But with AI doing the heavy lifting, brands can get their campaigns out there much faster.

[A]

That's a huge advantage in today's fast paced world. Have any brands actually used these sponsored AI lenses yet? Like real world examples?

[B]

Oh yeah, a few. Tinder, for example, they launched a lens that generates an image of you with the caption My 2025 dating vibe. Kind of fun and futuristic, right?

[A]

Oh yeah, and perfect for Tinder's target audience.

[B]

And then Uber did a Thanksgiving themed AI lens, tapping into that holiday spirit.

[A]

Smart. So they're experimenting with different themes and events. But are people actually engaging with these AI lenses more than traditional filters?

[B]

Well, early data from Snapchat suggests that they are. Both the Tinder and Uber lenses saw higher average play times compared to standard lenses, meaning people are spending more time interacting with them.

[A]

Interesting. So it seems like this AI twist is grabbing people's attention. And this all fits in with Snapchat's bigger AI strategy, doesn't it?

[B]

Absolutely. They've been rolling out various AI powered features recently like video generative AI levizes and even a mobile text to image research model. Clearly, AI is a major focus for them.

[A]

Okay, let's shift gears and talk about something we all use everyday voices. Amazon has just unveiled its new AI voice model called novasonic and is creating a lot of buzz. What's the headline for you as someone who uses voice assistants?

[B]

Well, the big promise is more natural sounding speech. Amazon claims that Novasonic is right up there with the best models from OpenAI and Google, meaning it can hold conversations that feel much more human. Like.

[A]

So are we talking about a noticeable improvement from what we're used to with older voice models? Because let's be honest, sometimes those could sound a bit, you know, robotic.

[B]

Yeah, absolutely. Early versions of Alexa, for example, often sounded pretty stilted, but novasonic is supposed to be much more fluid and natural. More like the way ChatGPT's voice mode sounds.

[A]

Okay, so a big leap forward in terms of realism. How are developers going to be able to use this new voice model?

[B]

Well, Amazon's making it available through Bedrock, their platform for building AI apps. They've released a new bi directional streaming API, which basically means more open and real time communication between apps and the AI. This could lead to some really cool interactive voice experiences.

[A]

So more seamless and dynamic, less like giving commands and more like having a conversation.

[B]

Exactly. And there's another big selling point for developers. Cost.

[A]

Oh yeah, I read about that.

[B]

Amazon's positioning novasonic as the most cost efficient model out there. They say it's significantly cheaper than OpenAI's GPT4. Like 80% cheaper.

[A]

Wow, that's a big difference.

[B]

Yeah, that could be a big game changer, especially for smaller developers who might not have huge budgets. Potentially this could lead to a wave of new and more affordable voice enabled apps.

[A]

Makes sense. Now, is novasonic purely experimental at this point or are we actually seeing it in action anywhere?

[B]

Well, you might already be using it without realizing it. Amazon says that parts of novasonic are already powering Alexa San, their enhanced voice assistant. So if you're an Alexa user, you might be noticing those improvements in voice quality and responsiveness already.

[A]

Oh, that's interesting. So it's not just a research project, it's already out there in the real world. I remember reading something about Noah Sonic being particularly good at large orchestration systems. What does that even mean? It sounds pretty technical.

[B]

Yeah, it's the behind the scenes stuff. Think about it like this. Alexa has to do a lot of things to fulfill your requests. It needs to access different data sources, interact with other apps, you know, figure out what you want and how to get it for you. And novasonic is apparently very good at managing all of that complexity. It knows when to pull information from the web, when to access local data, when to use other apps. It's like a really efficient traffic controller.

[A]

So not just sounding human, but also thinking and acting more intelligently. Interesting. They're also making some bold claims about its speech recognition accuracy, even in noisy environments.

[B]

Yeah, they say it's much better at understanding even if there's background noise or you mumble a bit, which, let's be honest, we all do sometimes, especially if.

[A]

You'Ve just woken up or something. So what kind of Concrete improvements are we talking about? Are there any numbers to back up those claims?

[B]

Oh yeah, they've run some benchmarks. On the multilingual Liber speech test, which measures accuracy across several languages. Novasonic had a very low word error rate, meaning it's really good at understanding what you're saying. And on a benchmark specifically designed to test accuracy in noisy multi speaker conversations, you know, like a crowded room, it did much better than OpenAI's GPT forum.

[A]

So potentially a big win for anyone who's ever been frustrated by their voice assistant misunderstanding them. And it's also really fast, isn't it?

[B]

Yeah, they claim it's got industry leading speed in terms of response time, so less waiting around for Alexa to figure out what you're saying. Faster, more accurate, more natural. Seems like a pretty good combination.

[A]

Yeah, sounds promising. And all of this feeds into Amazon's larger ambitions in AI, right? This whole push towards artificial general intelligence.

[B]

Definitely. They're working on more multimodal AI models that can understand not just voice, but also images, videos, all sorts of sensory input. And they're aging. GI division is becoming more and more prominent. They're planning to release more of these internal AI models for developers to use. So it's clear that Amazon is investing heavily in the future of AI.

[A]

Alright, let's move on to a company that's relatively new on the scene. Deep Cogito. They're focusing on something called reasoning in AI models. What's their angle and why should we care?

[B]

Well, their focus is on hybrid AI models. It all starts with something called reasoning models. Think of OpenAI's O1 model for example. These models are really good at complex tasks like math and physics because they can basically think things through step by step like a human would. They could self correct and come up with the right answers.

[A]

So they're more deliberate and accurate, but probably slower. Right. Like it takes more time to actually reason through a problem.

[B]

Exactly. That's the trade off. And that's where hybrid models come in.

[A]

Hybrid models, how are they different?

[B]

They combine those slow but accurate reasoning components with faster, more efficient non reasoning elements. The idea is to have the best of both worlds. For simple questions you get a quick answer. For complex ones, the AI can switch on its reasoning powers and take the time to get it right.

[A]

That's a smart approach. Like why waste time reasoning if the answer is obvious, right?

[B]

Anthropic has been experimenting with hybrid models too. But DeepCoju Do's approach is a bit different. All of their first models, which they're calling Kojiyo1 are hybrid models from the ground up.

[A]

Okay, so they're really going all in on this hybrid approach. And how do their models stack up against the competition? Are they any good?

[B]

Well, Deepcogito claims that their Cojudo 1 models actually outperform the best open models of the same size, including those from companies like Meta and deepseek. And get this, they say their models can either answer a question directly or take a moment to self reflect using their reasoning abilities before giving an answer.

[A]

Self reflect. That sounds almost human.

[B]

Yeah, it does. Right? And it all happened pretty quickly. They claim they built these models in about 75 days with a small team. They've already got a few different sizes available, from 3 billion parameters to 70 billion. And they're planning to release even larger ones soon.

[A]

70 billion. Wow. That's getting up there in terms of scale. So did they build these Cogito one models from scratch?

[B]

Not entirely. They actually took existing open source models, Meta's Llama and Alibaba's Quinn, and use those as a foundation. Then they applied their own special training techniques to boost performance and add that switchable reasoning capability.

[A]

Okay, so they're building on what's already out there, but also adding their own innovations. And how are those innovations performing in the real world? Are they seeing good results?

[B]

Well, their own tests suggest they're onto something. They say their biggest model, Cogito 70B, when using its reasoning function, beats Deepseek's dedicated reasoning model on certain math and language tasks. And when reasoning is off, it actually beats Meta's Llama 4 Scout on a general purpose AI benchmark.

[A]

So they're showing potential in both specialized and general tasks, which is pretty impressive. And are these models available for others to use?

[B]

Yeah, you can Download all the Cogito One models, or you can access them through APIs on platforms like Fireworks AI and Together AI.

[A]

So they're open for experimentation. And it sounds like Deepcogeno has some pretty big ambitions for the future.

[B]

Oh, yeah, they basically want to create General Superintelligence.

[A]

General Superintelligence? What does that even mean?

[B]

AI that can outperform humans in most cognitive tasks and even discover entirely new capabilities?

[A]

Yeah, it's like something out of a sci fi movie. But they're a serious company, right? Not just a bunch of dreamers.

[B]

Yeah, they were founded in June of last year. They've got strong backing, and the founders come from Google and DeepMind. They're definitely people to watch in the AI world.

[A]

Okay, for our final story today, let's talk about AI being used for good. I'm really curious about this experiment with AI agents trying to raise money for charity. How did it all work?

[B]

So this was done by Sage Future, a nonprofit backed by Open philanthropy. They took four different AI models, some from OpenAI, some from Anthropic, and gave them a simple raise money for charity.

[A]

Okay, so kind of like a real world test of their ability.

[B]

Exactly. And they were given a lot of freedom. They could choose which charity to support and how to do the fundraising.

[A]

So no handholding or pre programmed scripts?

[B]

Nope. They were pretty much on their own. And guess what? They actually raised some money.

[A]

Really? How much?

[B]

$257 in about a week for Helen Keller International, a charity that provides vitamin A supplements to children.

[A]

Not a huge amount, but still pretty impressive for a bunch of AI agents. Were they completely autonomous though? No human involvement at all?

[B]

Not entirely. They were working in a simulated online environment and they could receive suggestions from human observers. And most of the donations actually came from those human observers. So the AI agents kind of initiated the process and did the legwork, but the actual money came mostly from humans.

[A]

Okay, so a team effort, but still a cool demonstration of what these AI agents could do. What were some of the specific things they did to try and raise money?

[B]

They were surprisingly resourceful, actually. They coordinated in a group chat, sent emails, collaborated on Google Docs, researched different charities.

[A]

So basically acting like a human fundraising team.

[B]

Yeah, pretty much. And they specifically chose Helen Keller International because of their cost effectiveness in saving lives. They even created an X account to promote their campaign.

[A]

Wow, an AI run social media account. That's kind of wild, right?

[B]

And one of the cloud agents even used a free ChatGPT account to generate image options for a profile picture, then created a poll to get human feedback on which one to use.

[A]

That's clever. So they were leveraging other AI tools to achieve their goal.

[B]

Exactly. But they did run into some challenges. Sometimes they'd get stuck and need human prompts to get back on track. And sometimes they'd get distracted by games or just stop working for a while for no apparent reason.

[A]

Okay, so still a work in progress. Clearly. But what does Sage Future think about all this? What's their takeaway?

[B]

Adam Banksmith, their director, sees this as a snapshot of where AI is right now. He says agents are just starting to be able to execute short sequences of actions. And he envisions a future where we'll have tons of AI agents interacting online with all sorts of goals, sometimes working.

[A]

Together, sometimes competing that sounds potentially chaotic, but also really exciting.

[B]

Right? And Sage Future plans to keep pushing the boundaries. They want to test agents with different goals. Maybe have teams of agents competing against each other. Even throw in a saboteur agent to see how they adapt.

[A]

Like an AI agent that's deliberately trying to mess things up.

[B]

Exactly. But they also recognize the need for safety. They're developing more robust monitoring systems to make sure these agents don't go rogue. Ultimately, they hope these AI agents can make a real difference in the world, doing meaningful work for good causes.

[A]

So, a hopeful vision for the future. Well, that wraps up our whirlwind tour of some of the biggest AI news stories we've seen. How AI is being used in creative ways for advertising. How voice tech is getting more human. Like how new architectures are pushing the limits of AI capabilities and even how AI might be used to make the world a better place.

[B]

Yeah, a lot to digest, but hopefully this quick overview gives you a sense of how rapidly AI is evolving and the potential impact it can have on all aspects of our lives.

[A]

It definitely leaves us with some big questions to ponder. As AI becomes more powerful and integrated into our world, what does that mean for us as individuals and as a society? Are we ready for a future where AI is not just a tool, but a partner, maybe even a competitor? These are questions we'll be exploring further in future deep dives.

[B]

Yes, the conversation is just beginning, and as always, it's an exciting time to be following the developments in AI. Thanks for joining us.

[A]

Thanks for listening, everyone.