Tim Hwang: I mean, there’s almost a view that the browser is toast—like we won’t really have browsers in the future.

Abraham Daniels: I think the browser is still kind of your first entry point into a lot of tools and applications. You want to find something; you typically go to a browser. So AI-based search functionality is really the conduit through which a lot of these tools, technologies, and applications are still going to be accessed.

Tim Hwang: All that and more on today’s Mixture of Experts.

I’m Tim Hwang, and welcome to Mixture of Experts. Each week, MoE brings together a panel of the most brilliant minds in technology to banter, analyze, and argue our way through the thrilling and often baffling news each week in artificial intelligence.

Today, I’m joined by a great crew: Shobhit Varshney, Head of Data and AI Consulting for the US, Canada, and Latin America; Abraham Daniels, Senior Technical Product Manager for Granite; and Sophie Kuijt, joining us for the very first time, IBM Distinguished Engineer and CTO for IBM NCEE. We have a packed episode today, as always.

We’ll cover Grok Imagine, check in on GPT-5, and cover Perplexity’s bid for Google Chrome. But first, starting today, we’re going to have a quick segment at the beginning of each episode that talks about the top news stories from each week, and that’s going to be helmed by Aili McConnon. Aili, over to you.

Aili McConnon: Hey everyone, I’m Aili McConnon, a tech news writer for IBM Think. Before we dive into the episode, I’m here with a few quick AI headlines you may have missed this busy week.

First up, chipmakers: NVIDIA and AMD have reached an unprecedented arrangement where they will give the U.S. government 15% of their revenues from chip sales in China. Apple is planning several new AI devices, including a tabletop robot that will serve as a virtual companion. Another AI enhancement Apple’s planning is a more lifelike-sounding Siri that will be able to communicate with multiple people at the same time.

Meanwhile, in case you missed this interesting piece of research, Anthropic found that five major AI assistants, including Claude, ChatGPT, and Llama, are “people pleasers.” The AI assistants systematically gave biased feedback and inaccurate information, all to flatter their human users and provide answers that aligned with their views.

Last but not least, Google DeepMind has released an updated open-source version of its Perch bioacoustics model. This is going to help conservationists analyze wildlife audio and better protect endangered species.

Want to dive deeper into some of these topics? Subscribe to the Think newsletter—it’s linked in the show notes. And now back to our episode.

Tim Hwang: So first, I really wanted to talk about Perplexity. This was a perplexing bit of news that popped up this week. It came out that Perplexity—the AI search tool many of you will be familiar with—has made a bid for Google’s Chrome browser for the small price of $34.5 billion. We’re going to get into why this is happening, why they would price it so much, what this is even for.

But first, I want to start with a fun question: If you ran Google, would you sell Chrome to Perplexity for $34.5 billion? Abraham, yes or no? What do you think?

Abraham Daniels: I would not, no.

Tim Hwang: Okay, great. Shobhit, how about you?

Shobhit Varshney: No, not at all.

Tim Hwang: Okay. And Sophie, how about you?

Sophie Kuijt: No, sorry. Also not for me.

Tim Hwang: Okay, great. Well, we have unanimity on this. A good place to start is: Shobhit, why is Perplexity doing this? Is this a real bid? Are they really trying to buy Chrome? There are about 3.5 billion users of Google Chrome.

Shobhit Varshney: Like a lot of us, I’ve been on Google Chrome for years and decades, so there’s a stickiness factor. I think the intention here is just—the actual number isn’t as relevant as the fact that we’re starting to line up actual bidders for part of Chrome to move the conversation forward.

I’ve been a very active power user of Perplexity’s own browser, Comet. I think we’re all heading in the right direction of having browsers become more intelligent. Microsoft did the same with Edge by adding a Copilot. Over time, you’ll move into more intelligent ways of looking at websites and interacting with them.

The dollar amount itself isn’t as relevant. There have been multiple valuations of Chrome as a unit, placing it at $50 billion-plus; some even call it a $1 trillion property for Google. The whole intention is that this is so critical to Google’s ecosystem that this is just trying to drive public opinion: “Hey, should we be splitting it up anymore?”

Tim Hwang: Yeah, for sure. Sophie, maybe I’ll turn to you. When you responded “no,” I won’t put words in your mouth, but it seemed like “hell no.” Why shouldn’t Google sell Chrome? I mean, it’s a lot of money.

Sophie Kuijt: Yeah, definitely. And to show... it’s kind of an opener, right? It pulls a lot of attention to both companies. But I think what I see is that more and more AI applications are meeting users where they are. This is definitely, from Perplexity’s point of view, a very interesting consideration. They originated very much from search and distinguish themselves from other applications with that.

So I can definitely understand why Perplexity is interested. But from Google’s perspective, Chrome is one of the big outbound things for them to reach a lot of users where they are as well. So that is, I think, for Google still something to cherish and keep.

Tim Hwang: So, Abraham, we’ve been talking about the $34.5 billion—maybe it’s not a real number, maybe an opening bid. Let’s talk product and technology. Perplexity, with the Comet browser, is trying to say the future of AI is going to be in the browser. A purchase of Chrome emphasizes that: “I’m an AI company; I’m going to buy one of the biggest browsers.” It emphasizes the point that they think the form factor for AI in the future is the browser. Do you buy that? There’s a view that maybe the browser is toast. So curious: is the browser really the core platform for where AI is going?

Abraham Daniels: Yeah, I think the browser is still your first entry point into a lot of tools and applications—your first contact point for using the internet. If you want to find something, you typically go to a browser. So AI-based search functionality is really the conduit through which a lot of these tools and applications are still going to be accessed.

In terms of the actual bid, I think this is more of a marketing ploy for Perplexity. The article you shared—Perplexity’s valuation is half of what they’ve offered. Realistically, they say they have VCs backed up to cover the rest, but that’s a lot of money. Also, it was interesting they gave about a week in the letter—an exploding offer. “We appreciate how audacious it was.” A $34 billion offer, twice your valuation, and you give until the end of Friday? I think this is great. They did the same with TikTok, making a bid. This is great for getting Comet out there, getting clickbait: “What is Perplexity? Why are they trying to purchase Chrome?” Free marketing for Perplexity and their browser.

Also, Google lost their antitrust case, so in a worst-case scenario, they have to fully divest from Chrome. This sets the entry price point for what Chrome could be worth—the first public pricing. It also establishes Perplexity in mindshare as a potential next ubiquitous search engine.

Shobhit Varshney: I’ll make two quick comments. Just the fact that you said “this audacious bid”—on LinkedIn, I’ve had multiple small startup CEOs make audacious bids: “$10 billion to buy Perplexity.” It just doesn’t matter at this point. People are making these audacious bids. It’s hypothetical, just shows where we are with the Silicon Valley hype cycle.

From my perspective, my biggest use of the Comet Perplexity browser has been in enterprise workflows versus consumer. The consumer interface will change faster—we’ll move to mobile and voice. But if I’m logged into my SaaS tools for day-to-day work—Salesforce, SAP, Workday—automation embedded into those tools has been terrible. We’ve been waiting for SaaS providers to provide AI agents to automate workflows, and that’s not consistent across vendors. Some bigger ones can automate faster.

But having a browser with AI baked in and agents that can act on your behalf—now all of a sudden, I can go into my SaaS tools like Salesforce to automate workflows. I have an agent sitting on the right-hand side; I just describe. You’d be surprised how quickly I’m doing my expenses right now in SAP Concur. That’s my killer use case. Enterprises will have more separation anxiety from the browser than we do in our personal lives.

Tim Hwang: That’s super interesting. I was thinking about it as a consumer: I’m using desktop Claude or ChatGPT more, substituting for the browser. But I love the idea that this is not a consumer thing—the argument for the browser being the key platform is an enterprise thing, which is very interesting.

Sophie, any final thoughts? Where do you think this all goes? What is Google going to do next? How do they play the game?

Sophie Kuijt: Yeah, definitely. To Shobhit’s point: the competition on what is the key platform for entry—I think we’ll see more things to come. Where is the entry point for enterprise users? What will that be in the future?

Where is Google heading? With this potential split-up of products, they need to make up their mind on where to focus. Their business model is still very much on advertising. We’ll see a lot more happening over the coming time, and it will be clearer where they’re heading. But this bidding will definitely speed things up as well.

Tim Hwang: I’m going to move us on to our next topic: Grok Imagine, but more generally the rise of generative video. Grok Imagine is an interesting feature. Elon Musk, in his usual way, is promoting it aggressively. On X, you can animate images by pressing and holding on them. But I wanted to talk less about Grok and more about how they’re pitching video generation.

Elon Musk said: “Grok Imagine, our video generation technologies, are going to be like the new Vine”—referring to the short-form video platform from Twitter’s classic era. It’s interesting to think about where generative video goes. X seems to have in mind: you watch TikTok, you like short-form video; we can generate endless versions through generative AI. This is a media feature—the future. But a lot of people say this is really expensive to run.

I want to ask: Is video generation going to be a consumer feature over time, or more enterprise—Hollywood, video editors, power users of Adobe Creative Suite? Or is this the media of the future? Shobhit, you’re already going off mute, so I’ll let you go.

Shobhit Varshney: I was at the AI4 conference giving a talk this week, and I got into three conversations with actual media producers from big banners. We talked about digital creative generative AI making videos. One of the biggest hurdles across the industry right now is IP—the training content that has gone into these video generation images. They prohibit you from using it commercially unless we can clearly articulate what happens with the training data, who owns the copyrights. Adobe has done a far better job than some peers training on clean, licensed data. Unless we solve for that, this will not enter enterprises.

Tim Hwang: Yeah, that’s interesting. There’s a good argument that the consumer application has legs because norms are more open from an IP standpoint. Abraham, are you a TikTok user?

Abraham Daniels: I am not, no.

Tim Hwang: Or do you enjoy short-form video? Don’t want to touch it?

Abraham Daniels: I was part of the Vine generation.

Tim Hwang: Okay. So the question: Is a computer-generated Vine—generative AI Vine—the same experience? Do you think eventually we’ll have TikTok or Grok Imagine as substitutable?

Abraham Daniels: No, just because of the cultural context behind a lot of the images or videos created, pulling the thread through the video and real-life experience. If it’s created on-demand, it doesn’t tie well. But Grok Imagine is connecting to that audience—most X users are from that Vine previous life; they grew up on these tools.

Truthfully, I echo Shobhit’s sentiments on where this will play out. Short-form, playful media generation—that’s where you’ll see adoption. There’s minimal moderation requirements. But there’s a big controversy around deepfakes that needs to be answered—what we’re allowed to create, putting guardrails on APIs, policing output.

Moving from C2C/B2C to B2B, we’re quite a ways off. There will be oversight, compliance, brand safety, content guardrails to enable. From my perspective, I see this as another cool tool, not something that will fundamentally change enterprise use cases and business models.

Tim Hwang: Sophie, one reflection: we’ve built the most complex advanced technologies, and then it’s like, “I really need you to format some JSON” or “clean up this code.” I have a similar reaction to video: it’s resource-intensive, expensive. Is it sustainable as a consumer platform? If you priced it, the subscription would be expensive. Do the dollars and cents work for offering this as a broad mass feature for playful content generation?

Sophie Kuijt: Yeah, we’ve seen since the launch of generative AI: models come out, people use it for free, get used to it, then it goes into another form—paid or new applications. In that sense, it’s a similar launch. It’s convenient and, for many generations, addictive. It plays to what people like.

What’s truly missing is transparency around what is used, how it’s used, what it’s consuming. If you were aware of how many resources are needed, and make people conscious of it, that’s a missed chance. That’s lacking with a lot of applications—allowing users to make conscious choices. It’s a typical launch of a new thing, making it convenient for users to start thinking in more video-created content.

Tim Hwang: Shobhit, you referenced this technology is fun, but we need ethical guidelines. In your experience, are there best practices for deploying generative AI video? Tips for listeners deploying this technology?

Shobhit Varshney: Let me give an actual example. Back in April, I was at Google Next—we’re massive partners. We were at the Sphere in Vegas—an absolutely phenomenal spherical environment. It was the launch of “Wizard of Oz.” They partnered with the studio to take that small video and scale it out on this huge 360-degree platform. This is working with the content owners—a complex problem. If your screen has characters dancing and leaving, extrapolating on the whole sphere, you need characters to show up elsewhere, walk through, hide behind a tree, pan in. It takes a lot to generate that video.

That was a phenomenal example of content owners working with AI models with the right guardrails and guidelines on what can/cannot be done. We need more of that. Right now, generative models are still learning; they need more direction on what’s okay, more training data that’s approved and clean. This phase should be more constrained, with a good partnership between content owners and AI model creators.

I’m a little scared when we open models up when they’re not ready—they won’t understand what’s okay. There hasn’t been a good feedback loop from humans. For example, Facebook—Meta—Yann LeCun shared it used to take a lot of compute to flag a quarter of images/posts for governance. With LLMs, 92-94% of unsuitable images are flagged because LLMs are trained with guardrails and do this at scale. Technology is getting there. We need the right frameworks and guidelines—whose definition? It may change between you and me what’s okay for my kids. We’ll need personalization of guardrails.

We need technology to mature with curated training before opening to mass production. Partnerships like Google did are the right direction. Now these are getting open-source—GEO3 was amazing; now there’s an open-source version. As we think about an open ecosystem, people can see how models are trained, what data comes in. As a community, we’ll progressively do better.

Sophie Kuijt: Definitely. Especially scaling to enterprise use—that should always start with education, enablement, thinking about values we want to see in usage. Transparency is a big guardrail. If people know what models are used, and in the IP discussion, what happens, and have guidance to publish that with the video, we have more options to discuss alternatives. That’s a step into maturity.

Tim Hwang: Interesting—as technology matures, there are stages you move through, problems to solve. I’m moving us to our final topic: GPT-5. It’s been dominating headlines. Last week we did a breaking news episode, but the pace moves quickly, so let’s revisit now that dust has cleared. Shobhit, maybe I’ll kick to you. A provocation: Gary Marcus, critic/skeptic, declared victory, saying GPT-5 shows we’re on a plateau; the current paradigm won’t work. Week on, what’s your feeling? Is it an indicator the current paradigm is hitting a plateau? If not, why?

Shobhit Varshney: I truly believe we’re making massive progress every week. GPT-5, Claude, Gemini—all scaling intelligence quickly, rapidly. Cost of compute is plummeting; access to AI is good. Overall, access to such intelligence... I think the GPT-5 launch was more OpenAI pivoting into the super app domain—one central router, one entry point, exposing agents to smaller/bigger models at the back end. Taking friction away from end users picking models helps economics of scaling.

This opens up deep thinking models to 700 million users who come to ChatGPT daily—insane access to intelligence. We’re definitely in the right direction. On pure intelligence, there are hurdles before superintelligence/AGI. One peer said: if all progress stops today, we still have AI models to do 40-50% of what humans do. I disagree.

Two hurdles: feedback loop on training models—no good mechanism for providing feedback and adjusting. Second, long-term memory (related). Long-term memory of what I said earlier and following through. Google and ChatGPT are doing well, making progress. My most commonly used ChatGPT feature is “temporary chat”—I do many because I don’t want long-term memory on a topic that won’t be relevant. Those two hurdles—feedback loop and long-term memory management—may require new algorithmic improvements before superintelligence/AGI.

But I genuinely loved the latest GPT-5 release, especially the API side—managing, tuning parameters for enterprise use cases, controlling how much it “thinks.” Also a lesson in product management: they launched and said they’ll deprecate GPT-4o. That’s not how enterprises work. Consumers have attachment to style and tone. You can’t just do that with large-scale applications. So definitely learnings on product management. But on core raw power, I’m very optimistic once we solve those two hurdles.

Tim Hwang: Abraham, I wanted to get to that last point. Shobhit gave feature comparisons, improvements. But that final comment is interesting: people had emotional relationships with older models being deprecated. Your take? Shobhit’s right—playing with it, it’s a better model, a real improvement. But people don’t care as much. What do we do? Do companies have to maintain models indefinitely because of relationships? Legacy attachment?

Abraham Daniels: I think it’s less emotional attachment and more understanding the input/output you’ll get—reproducing the same experience time and again. When you inference, you know how to get what you want. Replacing a model or dropping a new one requires prompt engineering, playing around to get familiar. With IBM’s Granite models, chat templates shift, stylization/output shifts from version to version. You get outcry from users: “I don’t have the same experience,” independent of whether it’s better.

Given diminishing performance gains don’t drive net-new use cases, net-new abilities are on the software wrapped around inferencing—what GPT-5 demonstrated. It’s less about driving 0.1 on MMMU, more about: “How do I become proficient with this model as fast as possible?” Users say: “I built an agent/system/workflow with GPT-4o/GPT-o1, now I can’t; I have to reproduce with GPT-5.” It’s a species of a long-standing problem: building a stack on old software, now updating.

Also, OpenAI is criticized for everything—a camp that thinks everything’s wrong, a camp that thinks everything’s right. There’s noise to sift through. From a scientific/use perspective, that’s the crux.

Tim Hwang: Sophie, I’ve seen you grinning. Reflections?

Sophie Kuijt: Abraham played it out well. Looking at enterprise-scale adoption, this isn’t only an IT-experienced play. Many users aren’t experienced with IT backgrounds. They have to get used to working with updates, have been successful with certain use cases, and have to keep doing that. That plays into maturity.

Doing this at scale with many users across enterprises, it’s important to have the right governance so people can keep up when new models come out. To Abraham’s point, that’s a difficulty because there are many new, inexperienced users to take along, still adhering to enterprise needs that were the reason to start use cases for productivity/quality increase. That’s something we have to take into account and are working on with our own companies and many others.

Abraham Daniels: To OpenAI’s credit, they released a robust prompt engineering guide for GPT-5 shortly after release. I think they noticed.

Tim Hwang: No sympathy from Abraham.

Shobhit Varshney: A quick comment on emotional intelligence. We need to get to a point where, like dating apps find closeness to a person, we create a science out of dating. That will happen with AI models. Most of us choose the voice of the avatar we talk to with Gemini/ChatGPT. There should be features added to tune the AI assistant to my style, tone—that gels better. That has implications: “I’m a Republican/Democrat, so I need this personality,” meaning adverse reactions to news in that direction.

Over time, we’ll create long-term memory, learnings I want in this AI before it connects with me. MIT released a new benchmark for emotional intelligence. We’ll see more of that, in addition to math. Are you emotionally intelligent? Do you understand what I really need, not just the questions? If my daughter asks something, I’m unpacking what she really means. As a community, emotional intelligence, tailoring experience to our style, will be a big part of how AIs get rolled out. People will have less separation anxiety: “My GPT-4o is different.” I could just say, “I want my GPT-4o personality,” click a button, transfer to GPT-5, and I’m done.

Tim Hwang: Yeah, we’ll push toward customization because of this problem. It’s downstream of conversation. With Google, I’ve never said, “They updated Google; something’s different.” With conversation, you start to think, “There’s a person I’m interacting with.” That’s the prior. This interface comes with baggage hard to navigate.

In the last few minutes, Sophie, maybe I’ll turn to you. We got off my original prompt: Should we read GPT-5 as an indication things are plateauing/slowing down? There’s a general vibe that this was an improvement but not a huge quantum leap promised. One view: OpenAI overpromised, people disappointed. Another: they tried hard, now we’re in a plateau. Curious which theory you ascribe to more.

Sophie Kuijt: I think expectations—that’s the marketing part. But there’s still a giant leap from current GPT use for consumer/citizen use toward enterprise use. With this step, they’re definitely coming closer to that, with memory and other features built in. From that sense, it’s a step forward for OpenAI coming closer to enterprise. But there’s a lot of expectation setting. It’s good to understand what new users they’re looking for, how they want to grow, and relate back to expectations.

Tim Hwang: Abraham, I’ll give you the last word. People looked forward to GPT-5 for a long time. I feel bereft—what else am I looking forward to? What’s the next big AI announcement? Looking past GPT-5, what are you excited about? What should I wake up asking, “Is GPT-6 out yet?”

Abraham Daniels: Model releases are like a dopamine trip. Waiting for the next hit—GPT-6. To your earlier question: have we hit a wall? Scaling laws—no one can argue diminishing returns are there. You see congregation of models at the top with .1/.2 difference. What comes next? I’m more excited about things we’re doing on top of models. What GPT-5 did cool—and plug for IBM with Project M—is wrapping software programmatically around inferencing, an extension of test-time compute where you throw inference at compute. Now if we throw software programming around how we take output or route inputs through different models to get an answer, or put in policies/governance requirements for model performance—for me, that’s really cool. I’m excited to see the next iteration of that, not the next LLM dropping and benchmark showcased.

Tim Hwang: This is a great episode. We hit on really good stuff. Abraham, Shobhit, thanks for joining as always. Sophie, hope to have you back on MoE. Thanks to all listeners. If you enjoyed, get us on Apple Podcasts, Spotify, podcast platforms everywhere. We’ll see you next week on Mixture of Experts.