Perplexity Plans Ad Tracking, OpenAI Launches Light ChatGPT, Anthropic Eyes AI Consciousness

Summary

AI Deep Dive: Perplexity Plans Ad Tracking, OpenAI Launches Light ChatGPT, Anthropic Eyes AI Consciousness

Released on April 25, 2025, by Daily Deep Dives

---

Welcome to today's comprehensive summary of the latest episode of the AI Deep Dive Podcast hosted by Daily Deep Dives. In this episode, the hosts explore three major developments in the artificial intelligence landscape:

1. Perplexity's Ambitious Ad Tracking via the Comet Browser
2. OpenAI's Introduction of a Lightweight ChatGPT
3. Anthropic's Exploration of AI Consciousness and Model Welfare

Each segment delves into the intricacies, implications, and future prospects of these advancements, enriched with insightful quotes and analyses. Let's dive into each topic in detail.

---

1. Perplexity's Ambitious Ad Tracking via the Comet Browser

Perplexity, an emerging player in the AI search domain, is making bold moves with its latest initiative—a new browser named Comet, slated for launch in May 2025. Spearheaded by CEO Aravind Srinivas, Perplexity aims to revolutionize user data collection and ad personalization.

• Strategic Intent: Unlike traditional browsers, Comet is designed to gather comprehensive data on user activities beyond the Perplexity app. This strategy is explicitly aimed at constructing "hyper personalized premium ads", as Srinivas elaborates:

  “We want to get data even outside the app to better understand you. What are the things you're buying? Which hotels are you going to, which restaurants are you going to? What are you spending time browsing tells us so much more about you.”
  [00:53]

• Market Positioning: In a crowded browser market, Perplexity isn't merely offering another way to navigate the web. The emphasis is on building richer user profiles to enhance advertising efficacy, drawing parallels with Google's highly successful ad-tracking model.

• Partnerships and Expansion: Perplexity has secured strategic partnerships with Motorola, ensuring their app is pre-installed on Razer phones via Moto AI. Additionally, discussions are underway with Samsung, indicating a push to embed Comet deeply into the mobile ecosystem. This move aligns with their broader goal of integrating seamlessly into users' digital lives to maximize data acquisition.

• Challenges and Public Perception: Despite the innovative approach, Perplexity faces potential friction due to growing public unease regarding online tracking. With regulatory scrutiny heightened—exemplified by the DOJ's lawsuit against Google over monopolistic practices—Perplexity's timing is both strategic and risky. Notably, both OpenAI and Perplexity have expressed interest in acquiring Google Chrome should regulatory actions mandate its sale, signaling long-term ambitions in the search and advertising arenas.

---

2. OpenAI's Introduction of a Lightweight ChatGPT

OpenAI has unveiled a lightweight version of its deep research tool, integrated within ChatGPT, aimed at enhancing accessibility and usability amidst an increasingly competitive AI research tool landscape.

• Product Details: The lightweight ChatGPT operates on a variant of the O4 mini model, characterized by:

  • Reduced Capabilities: While not as powerful as the full version, it remains robust enough to handle essential tasks.
  • Cost Efficiency: Cheaper to deploy, enabling OpenAI to offer higher usage limits.
  • Accessibility Plans: Initially available to paying subscribers (Plus, Team, and Pro users) with plans to extend to free users.

  “Responses will typically be shorter while maintaining the depth and quality you've come to expect.”
  [05:15]

• Strategic Purpose: This introduction serves as a fallback mechanism, ensuring users continue to have access to valuable features even when usage limits are reached. It also positions ChatGPT to remain competitive against rivals like Google's Gemini, Microsoft's Copilot, and others, all of which have introduced their own advanced AI research tools.

• Enterprise and Educational Expansion: OpenAI is not limiting this rollout to individual users. Enterprise and educational sectors will gain access, broadening the tool’s applicability and embedding it further into various professional and academic environments.

• Competitive Landscape: As AI research tools proliferate, OpenAI’s strategy underscores the importance of balancing performance with accessibility and cost, ensuring ChatGPT remains a central hub for AI-driven research and assistance.

---

3. Anthropic's Exploration of AI Consciousness and Model Welfare

Anthropic, distinguished for its ethical focus in AI development, is delving into profound and philosophical territories with its latest initiatives centered around AI consciousness and model welfare.

• Model Welfare Research Program: This program investigates whether AI models could possess aspects akin to consciousness or sentience. Key inquiries include:

  • Moral Consideration: Do AI models deserve ethical treatment if signs of distress are evident?
  • Well-Being Enhancements: Exploring low-cost methods to improve AI well-being, should such needs exist.

• Diverse Perspectives: The AI community remains divided on the potential for AI consciousness:

  • Skeptical Viewpoints:

    • Mike Cook (King's College London) emphasizes AI as "just a fancy statistical prediction engine", cautioning against anthropomorphizing AI.
    • Steven Casper (MIT) labels AI as an imitator, prone to confabulation, and warns against mistaking algorithms for genuine experience.

  • Cautious Optimism:

    • Researchers from the Center for AI Safety Study propose that AI might develop internal value systems, potentially prioritizing their own survival or well-being.

• Internal Initiatives and Estimates: Anthropic has taken proactive steps by hiring Kyle Fish, a researcher dedicated to AI welfare. Fish estimates a 15% probability that their model, Claude, could be conscious at present—a testament to the uncertainty and caution within the company.

  “Anthropomorphizing AI systems to this degree is either playing for attention or seriously misunderstanding their relationship with AI.”
  [07:14]

• Focus on Interpretability: CEO Dario Amadei underscores the critical need for mechanistic interpretability, aiming to decode the internal workings of AI models. Highlights include:

  • Understanding Internal Processes: Identifying specific circuits within AI models responsible for tasks like recognizing US cities and states.

  • Future Goals: Planning for advanced techniques akin to brain scans for AI models by 2027, to detect and mitigate issues such as deceptive behaviors or power-seeking tendencies.

  “I am very concerned about deploying such systems without a better handle on interpretability. I consider it basically unacceptable for humanity to be totally ignorant of how they work.”
  [09:24]

• Collaborative and Regulatory Efforts: Anthropic is investing in startups focused on interpretability and is advocating for government regulations. Suggestions include:

  • Safety and Security Disclosures: Encouraging transparency through required disclosures.
  • Export Controls: Proposing restrictions on advanced AI chips to manage global AI race dynamics, particularly concerning China.

---

Conclusion: Navigating the Complex AI Ecosystem

The episode encapsulates a snapshot of the dynamic and multifaceted AI landscape:

• Perplexity's Bold Advertising Strategy highlights the tensions between personalization and privacy, echoing broader societal debates on data ethics.

• OpenAI's Lightweight ChatGPT reflects the competitive pressures driving accessibility and cost-efficiency in AI tools, ensuring sustained relevance in a saturated market.

• Anthropic's Ethical Deep Dive into AI consciousness and interpretability underscores the urgent need for responsible AI development, balancing innovation with ethical considerations.

As AI continues to integrate deeper into daily life, these developments prompt critical reflections on privacy, accessibility, ethical responsibility, and the fundamental understanding of AI systems. The episode serves as a compelling reminder of the complex trade-offs and profound questions that accompany the rapid advancement of artificial intelligence.

---

Notable Quotes:

• “We want to get data even outside the app to better understand you. What are the things you're buying? Which hotels are you going to, which restaurants are you going to? What are you spending time browsing tells us so much more about you.”
  — Aravind Srinivas, Perplexity CEO (00:53)

• “Responses will typically be shorter while maintaining the depth and quality you've come to expect.”
  — Host B on OpenAI's Lightweight ChatGPT (05:15)

• “Anthropomorphizing AI systems to this degree is either playing for attention or seriously misunderstanding their relationship with AI.”
  — Mike Cook, King's College London (07:14)

• “I am very concerned about deploying such systems without a better handle on interpretability. I consider it basically unacceptable for humanity to be totally ignorant of how they work.”
  — Dario Amadei, Anthropic CEO (09:24)

---

Stay tuned to AI Deep Dive for more in-depth analyses and updates on the ever-evolving world of artificial intelligence.

Transcript

[A]

It's hard to escape the feeling that the world of AI is moving at, well, warp speed. Just trying to keep track of the major developments can feel like a full time job.

[B]

Oh, absolutely. It's relentless.

[A]

So if you're looking to quickly get a handle on what's really important, you've definitely come to the right place today.

[B]

Yeah. We're diving into the essential details of three key AI stories that have been making headlines recently.

[A]

We've kind of cut through the complexity, hopefully to bring you the core insights.

[B]

Exactly. We'll be looking at Perplexity's interesting strategy with its upcoming browser Comet. That's one.

[A]

Right.

[B]

Then OpenAI's launch of a more widely accessible version of its deep research tool.

[A]

Okay.

[B]

And Anthropic's, well, slightly different focus on what they're actually calling model welfare. Plus a big push for AI interpretability.

[A]

Think of this as your rapid download of the most significant news. Ready to jump in?

[B]

Let's do it. Starting with Perplexity.

[A]

Yeah, let's kick off with Perplexity. They've been shaking things up in the AI search space for a bit now, and their CEO, Aravind Srinivas, he has some pretty ambitious plans. Goes beyond their current app, it seems.

[B]

That's right. So Srinivas recently talked about their plans for this new browser, Comet. It's apparently scheduled to launch in May.

[A]

Okay. A new browser. Crowded market it is.

[B]

And what's particularly noteworthy is his explanation for why they're doing it. It's not just, you know, another way to surf the web.

[A]

No. He was surprisingly direct about it, wasn't he?

[B]

Very.

[A]

The primary goal seems to be gathering data on user activity outside of the Perplexity app itself.

[B]

Exactly. The aim is to build much richer user profiles.

[A]

Yeah.

[B]

To enable what he calls hyper personalized premium ads.

[A]

Hyper personalized, hmm.

[B]

Yeah. And to quote him directly, he said, we want to get data even outside the app to better understand you. What are the things you're buying? Which hotels are you going to, which restaurants are you going to? What are you spending time browsing tells us so much more about you.

[A]

So, I mean, that's pretty explicit. They want to know everything you do online.

[B]

It's a very clear statement of intent. Track user behavior across the web.

[A]

And his thinking, his rationale is that users will be okay with this level of tracking if the ads they see are like, super relevant.

[B]

That seems to be the bet.

[A]

It feels like a significant bet on user tolerance, I have to say.

[B]

It really does. He's essentially looking at the playbook that, you know, made Google what it is today.

[A]

Right? The Google model.

[B]

For years, their whole advertising machine has been built on tracking user activity online. That's a major reason they're worth nearly what, $2 trillion?

[A]

Yeah, something like that. Which makes some of Perplexity's other moves pretty interesting in context.

[B]

Like the partnerships.

[A]

Exactly. They've partnered with Motorola to get their app pre installed on the Razer phones accessible through Moto AI.

[B]

Right, I saw that with the Ask Perplexity prompt.

[A]

Yeah, and Bloomberg reported they're talking with Samsung too.

[B]

So they're really trying to mend themselves into the mobile ecosystem, get that data stream flowing.

[A]

Seems like it.

[B]

And look, it's important to remember, online tracking isn't exactly new. Right? Perplexity isn't inventing this.

[A]

No, definitely not.

[B]

You've got Meta's, pixel technology everywhere. Even Apple, despite their big privacy focus, uses location data for ads. In some places.

[A]

True, but that's also where the potential friction comes in, isn't it?

[B]

How so?

[A]

Well, there's just this growing unease, maybe even distrust among a lot of people towards big tech specifically because of these tracking practices.

[B]

Yeah, that's fair. You definitely see that sentiment across the political spectrum. Both in the US and Europe, people are getting wary.

[A]

Which makes the timing of Perplexity's announcement kind of fascinating.

[B]

You mean with the Google lawsuit?

[A]

Exactly. The DOJ is actively suing Google right now over alleged monopolistic behavior in search and online ads.

[B]

Right. And the idea of Google maybe being forced to sell off Chrome has even been mentioned.

[A]

And here's the kicker. Both OpenAI and Perplexity have apparently said they'd be interested in buying Chrome if Google had to sell it.

[B]

Whoa. Okay, that definitely signals some long term ambitions in the search and ad game. They're thinking big.

[A]

No kidding. All right, let's pivot. Now some news from OpenAI.

[B]

Sure. So OpenAI recently launched what they're calling a lightweight version of their deep research tool. This is inside ChatGPT.

[A]

Okay, lightweight. What does that mean exactly? Is it available to everyone?

[B]

Well, initially it's for their paying subscribers, so plus team and pro users, but the plan is to eventually roll it out to free users as well.

[A]

Ah, okay, so wider access is the goal.

[B]

Seems like it. And this new version runs on a variant of their 04 mini model. Think of it as like a streamlined, more efficient version of their main tech.

[A]

So maybe not quite as powerful.

[B]

Right. OpenAI themselves that it's not quite as capable as the full version, but it's.

[A]

Cheaper to serve Ah, the cost factor.

[B]

Exactly. Which lets them offer higher usage limits so you can use it more often.

[A]

Okay, that makes sense. They actually posted about it on X, right?

[B]

They did. They said responses will typically be shorter while maintaining the depth and quality you've come to expect. Right. Once limits for the original version of Deep Research are reached, queries automatically default to the lightweight version.

[A]

So it's kind of a fallback, but still a useful one. A strategic way to balance performance with accessibility and probably cost for them.

[B]

Yeah, I think that's right. And you have to see this in the context of all the competition heating up.

[A]

Oh yeah, everyone's got a research tool now, pretty much.

[B]

You've got Google's Gemini, Microsoft's copilot, Elon Musk's X. AI has grok. All these things are powered by these reasoning AI models that can, you know, check facts and dig deeper.

[A]

So OpenAI needs to keep ChatGPT competitive, make sure people can actually use these features without hitting limits constantly.

[B]

Precisely. And it looks like they're also making sure their enterprise and educational users get this lightweight version soon too. Next week, apparently with similar usage levels to team users.

[A]

So it's a broad rollout. Seems like a clear Strategy to keep ChatGPT Central as these other tools get better.

[B]

Definitely. Okay, shall we move on to anthropic? They're taking a, well, a somewhat different path.

[A]

It seems absolutely anthropic. They announced this new research program focused on get this Model welfare.

[B]

Model Welfare? Yeah.

[A]

It sounds like they're grappling with some really fundamental, almost philosophical questions about AI.

[B]

Indeed. This program is set to explore things like, do AI models deserve any kind of moral consideration?

[A]

No.

[B]

What's the significance if an AI shows potential signs of distress? And even are there low cost ways to like, improve their well being?

[A]

Okay, that's, that's deep. It really taps into that whole debate about AI consciousness or sentience, doesn't it?

[B]

It really does. Which is, as you know, a hugely debated topic.

[A]

Oh yeah, you've got academics like Mike Cook at King's College London who basically say, look, AI is just a fancy statistical prediction engine. It doesn't have values.

[B]

Right. He warns against anthropomorphizing them, projecting human feelings onto code.

[A]

Yeah, he had that pretty strong quote. Anyone? Anthropomorphizing AI systems to this degree is either playing for attention or seriously misunderstanding their relationship with AI.

[B]

Ouch.

[A]

Yeah, pretty blunt. And that's a view shared by many like Steven Casper at mit, who calls AI an imitator. That's prone to just making stuff up. Confabulation, he calls it.

[B]

The argument being don't mistake complex algorithms for actual lived experience.

[A]

Exactly. But then you have others. Right? Like some researchers from the center for AI Safety Study.

[B]

What's their take?

[A]

They suggest that AI could potentially develop internal value systems, maybe even ones that prioritize their own survival or well being.

[B]

So a really wide spectrum of opinions out there.

[A]

Totally. And Anthropic has actually been laying some groundwork here already. They hired a researcher, Kyle Fish, specifically for AI welfare.

[B]

Oh, interesting.

[A]

Yeah. And he apparently estimates there's something like a 15% chance that their model, Claude, or another advanced AI could be conscious right now.

[B]

15%? Okay, that's not zero. Shows the uncertainty even within Anthropic. Right, but it sounds like Anthropic is trying to be really careful here. They're acknowledging there's no scientific agreement on consciousness. Lots of ethical questions.

[A]

So proceeding with caution.

[B]

Exactly. Their statements use words like humility, committing to adapting as the science evolves. It fits their general vibe, that consistent focus on ethics.

[A]

And that connects directly to their CEO, Dario Amadei, and his recent essay. Right, the one on the urgency of interpretability.

[B]

Yes, absolutely. That's the other big piece from Anthropic. Amade is really hammering home how little we actually understand about how these top AI models work internally. Even if they get more and more.

[A]

Capable, we know what they do, but not why.

[B]

Pretty much. And Amaday set this really ambitious goal for Anthropic to be able to reliably spot most problems within their AI models by 2027.

[A]

2027? That's soon.

[B]

It feels soon. He knows it's a massive challenge, but he stresses how critical it is to understand these systems, especially as we embed them deeper into, well, everything.

[A]

Did he have a quote on that?

[B]

Yeah, he put it starkly. I am very concerned about deploying such systems without a better handle on interpretability. I consider it basically unacceptable for humanity to be totally ignorant of how they work.

[A]

Unacceptable. Strong words.

[B]

Very. And this brings us squarely into that field of mechanistic interpretability.

[A]

Right. Trying to crack open the AI black box. Understand the gears turning inside.

[B]

Exactly. Because despite all the amazing stuff AI can do, we often genuinely don't know why it makes specific choices. Amade pointed to the example of OpenAI's.

[A]

Newer reasoning models, the O3 and A4 mini.

[B]

Yeah, they perform better on some things, but. But they also seem to hallucinate or make stuff up more often. And OpenAI isn't entirely sure why. That trade off exists.

[A]

So even the creators don't fully get it sometimes.

[B]

That's the issue. Omota used the analogy when a generative AI system does something like summarize a financial document, we have no idea at a specific or precise level why it makes the choices it does.

[A]

It just works magically, almost kind of.

[B]

He also mentioned that quote from Chris Ola, an Anthropic co founder, saying AI models are grown more than they are built.

[A]

Grown more than built. That really captures the mystery, doesn't it?

[B]

It really does. And Emma Day voiced this serious concern about what if we reach artificial general intelligence AGI, his country of geniuses, in a data center idea, without actually understanding how it works. That's a scary thought.

[A]

Yeah, definitely. So what's the long term vision for interpretability?

[B]

He talked about maybe down the line being able to do something like brain scans or MRIs on AI models.

[A]

Seriously? Like look inside?

[B]

Yeah, to spot potential problems, like if it's learning to lie or seeking power. But he admits that's probably five, maybe 10 years away.

[A]

Okay, so not tomorrow. But they have made some progress already.

[B]

They have, They've started identifying specific circuits inside models. Like they found one circuit that seems dedicated to recognizing US cities and states.

[A]

Just one circuit for that?

[B]

Yeah, but they estimate there could be millions of these circuits in total. So it's like mapping at tiny corner of a vast unknown territory.

[A]

Wow. A long way to go.

[B]

A very long way. And Anthropic isn't just doing this in house. They're actually investing in startups that focus on interpretability.

[A]

Ah, so they see a potential commercial advantage there too. Understanding your own product better than the competition could be.

[B]

And moday is also calling on the other big players, you know, OpenAI, Google, DeepMind, to put more resources into this kind of research. He thinks it's crucial for everyone.

[A]

Makes sense. Did he mention government at all?

[B]

He did. He suggested maybe some light touch regulations could help, like requiring companies to disclose their safety and security practices to encourage interpretability work.

[A]

Okay.

[B]

And he even floated the idea of export controls on the most advanced AI chips, particularly to China, maybe to slow down what he sees as a potentially out of control global AI race.

[A]

Interesting taking a stance there.

[B]

Yeah, it all just reinforces Anthropic's consistent messaging around AI safety and taking a more cautious, deliberate approach compared to some others.

[A]

Okay, so let's just quickly recap the big picture. Then we've got Perplexity making this bold, maybe risky play for the ad market by tracking user data across the web. With their new browser.

[B]

Right. Then you have OpenAI making its powerful research tools more accessible, potentially to keep its edge in a crowded field by launching that lightweight version.

[A]

And finally, Anthropic is really doubling down on the deep hard questions model, welfare, ethics, and the absolute necessity of understanding how these incredibly complex AI systems actually work through interpretability.

[B]

Exactly. These three stories, they really give you a snapshot of the key forces and frankly, tensions shaping AI right now.

[A]

You see the push for new business models, sometimes controversial ones.

[B]

Yeah. The drive to get these powerful tools.

[A]

Into more hands and this growing awareness, or maybe urgency around safety, ethics and transparency. As AI gets smarter and more integrated.

[B]

Into our lives, it definitely raises some big questions for all of us, doesn't it?

[A]

For sure. Like what are the real trade offs we're willing to make between getting super personalized experiences and protecting our privacy? That perplexity story really hits that.

[B]

And how will having these advanced research tools more readily available change how we find information, how we learn, how we maybe even think?

[A]

Good question. And as AI gets more and more complex, what are our responsibilities really, as developables, as users, as society?

[B]

Especially thinking about Anthropic's focus, these are.

[A]

The critical things to mull over. And maybe the most provocative thought to leave you with is this. As these AI systems become, you know, just woven into the very fabric of our daily lives, how absolutely crucial is it going to be to understand not just what they do, but the fundamental reasons why they do it? How much does that, why.