Gabe Goodhart: There are those that project into a utopian future where AI really plays the role of humans in roles that humans don’t like to play. There’s those that also see that exact same future and get extremely pessimistic and worried about it.

Bryan Casey: All that and more on Mixture of Experts.

Alright. Hello everyone, I am Bryan Casey. It’s back-to-school week this week, and Tim’s kid gave him a fever immediately. So you are all stuck with me. I think that’s a situation we can all, as parents—at least for those of you that are—understand and get ready for in terms of the fall season. So you’re all stuck with me today.

Welcome to Mixture of Experts. Every week we bring together a panel of experts, technologists, and product leaders to talk about the latest news in AI.

Today, I’m joined by a great crew: Kaoutar El Maghraoui, Principal Research Scientist and Manager for the Hybrid Cloud Platform; Gabe Goodhart, Chief Architect for AI Open Innovation; and Mihai Criveti, Distinguished Engineer for Agentic AI.

As always, we have a packed episode this week. We’ll be talking about Anthropic’s recent raise, skepticism, AI bears, and an AI winter. But we’re going to start off with the biggest story of the week, which was we finally got some clarity on where Google’s antitrust case was going to land.

Before we get into that, I’m going to turn things over to Aili McConnon, who’s going to take us through just a couple of other top stories of the week. So, Aili, over to you.

Aili McConnon: Hey everyone, I’m Aili McConnon, a Tech News Writer with IBM Think. Before we dive into today’s main episode, I’m here with a few AI headlines you may have missed this busy week.

First up: OpenAI has added new safeguards to ChatGPT so it can better detect emotional distress in teens and other users in crisis and guide them to the real-world support they need.

Next: IBM and the semiconductor company AMD are teaming up. They’re combining the power of quantum computing with traditional computing and AI to create quantum-centric supercomputing.

Meanwhile, your back-to-school shopping may look a little different. Amazon has just launched Amazon Lens Live, so next time you’re out shopping, you can hold your phone, point it at an object, and see matching products you can swipe through.

Last but not least: You may never need to worry again that Starbucks is going to run out of your favorite vanilla cold cream or caramel drizzle. Starbucks baristas will soon only need a tablet to scan, store, and supply shelves, and the built-in AI tools will identify which items are running low and then automatically reorder them.

Want to dive deeper into some or all of these topics? Subscribe to the IBM Think newsletter linked in the show notes. And now back to our episode.

Bryan Casey: We’ll kick things off today by talking about the Google antitrust case. For those following closely, many consider it the biggest antitrust case in tech since the ‘98 case, U.S. vs. Microsoft. Some call it a blockbuster case.

This case actually kicked off in 2020—five years ago, which feels like an eternity. One of the most interesting aspects is how much has changed over the last five years in this space. The reason this show exists is because of AI, and it turns out AI played a significant role in the ruling.

The ruling was described as a more conservative, market-based approach than more dramatic results could have been. There was discussion about whether Google would be forced to divest things like Chrome or Android, or whether they could still pay to be the default in browsers and devices.

The ruling ended up: Google keeps Chrome and Android, can continue to pay to be the default search engine for platforms like browsers and devices. But a few things changed:

1.     Some exclusivity agreements Google has can’t be pursued the same way.

2.     Limited data sharing with potential competitors to make it easier for other entrants—a more dynamic marketplace.

Back to the point: the case originated in 2020, now 2025. One big reason for the conservative ruling was AI. I’ll read a couple lines from the document:

·       “Generative AI technologies pose a threat to the primacy of traditional internet search.”

·       “The money flowing into the space and how quickly it has arrived is astonishing. These companies are already in a better position, both financially and technologically, to compete with Google than any traditional search company has been in decades.”

The whole category changed over five years. That’s what I want to talk about: the convergence of search and AI markets.

Let’s start with defaults and how important they are. A counterpoint: ChatGPT has hundreds of millions of users, one of the fastest-growing products ever. If defaults were that big a deal, how is it growing so quickly? Are defaults as important as people say, or is this happening anyway? Gabe, maybe start with you.

Gabe Goodhart: Let’s start with defaults. How many applications did you use today before logging into this session? Probably a couple hundred. How many individual settings in each? Tens of thousands total. Defaults in software engineering are unsung heroes enabling us to use software daily.

Something as big as your default entry point into the internet is a big deal. For enthusiasts, you might change that default, but for the vast majority who don’t join AI podcasts: if it works, it works. Don’t mess with it. If the button is there when I grab my phone, I click it until it stops working.

Defaults are extremely important. This ruling is clearly a win for Google—they can still throw financial muscle around to remain default. But there’s a win for the enthusiast consumer willing to change a default, especially around blurring lines between generative AI and search.

The best use of generative AI is good search. Until now, ChatGPT lives in a different bucket than the Google icon—maybe labeled “AI.” To truly become the default, it must blur into the utility people go to for knowledge or finding things online.

This ruling may allow that blurring by allowing AI tools to be framed as search options on mobile devices, changing the landscape from a one-horse race to a mini horse race. AI tools challenging Google’s role as default hasn’t happened yet. On platforms like Twitter, people say they don’t use Google anymore—but Twitter is an echo chamber, not reflective of the entire consumer market.

I don’t see chatter about browsers/devices flipping from Google/search as the primary web interface to AI tools. I’m curious: What’s the barrier? Is it consumer behavior, technological limitations, or both? They’re not defaults yet.

Kaoutar El Maghraoui: That’s a great question. With AI, there’s a twist. I agree with Gabe: defaults are important. Research shows most users stick with defaults—they don’t switch settings. Google paid Apple over $20 billion annually to secure Safari’s defaults—there’s real strength there.

With AI, the importance of defaults intensifies because assistants don’t just route you to links; they shape answers. With agentic AI, routing to many tools underneath—users aren’t even aware—defaults become even more important.

Platform partnerships are key. Apple and other platform makers are becoming kingmakers. If Siri defaults to OpenAI or Anthropic instead of Google, the balance of power in AI search shifts overnight. It’s important to understand how these partnerships shape the market.

For users, initial experience is sticky. Platforms presenting these products—and with agents, where they default under the hood—is crucial. How do we break into those partnerships? New startups or companies wanting to enter face a monopoly still holding big power. Who pays most in AI partnerships?

The ruling is a mixed bag: a short-term win for Google avoiding breakup and protecting its core business model. For competitors, including AI startups, it’s mixed. The judge didn’t break up Google (many rivals hoped for that), but the data-sharing mandate is a lifeline. AI companies like OpenAI, Perplexity stand to benefit—accessing trove information to improve answer engines and compete more directly with Google. But penetrating partnerships with platform makers like Apple is also key.

Mihai Criveti: Google wins because they keep Chrome. Apple wins because they keep the $20 billion from Google and open up for additional revenue from OpenAI, Anthropic, other AI/search engines. Some not winning: Mozilla, Firefox, smaller browsers—they still get revenue from Google (paid to be default), but they don’t have an avenue to penetrate this market.

We’re consolidating search and AI capabilities in the hands of maybe 3–4 large organizations: Google, OpenAI, Anthropic, frontier models, maybe Perplexity. They’ll have funding to get into default or at least second/third option. Smaller providers won’t have a mechanism.

Second: What are you searching for on phone vs. tablet vs. computer? On phone, I’m out of house—searching for nearby restaurants open. Still a necessity for traditional search engines—they have data from maps, businesses, reviews. Even with AI on top, it’s still Google’s or Bing’s data. OpenAI and other AI providers don’t have that current data availability. They can build agents on top, but for phone use, not as relevant for desktop/other applications.

Nobody on phone asks, “How do I create a Python program to do this?” The answer might be better than top StackOverflow hits, but you’d pick that answer. Watch this space—it’ll take more years before major impact from AI search engines. Search engines themselves now prioritize AI answers (Google Search, Bing AI summary) to compete with AI vendors.

Gabe Goodhart: One more thing on why AI apps aren’t replacing browser apps: technology. The web is visual—websites have banners, visually appealing elements. AI assistant UX is primarily text-based. There’s novelty in text framed directly to you, driving the AI revolution, but that UX layer hasn’t incorporated generated visual output.

Nothing fundamental about generative AI prevents that—it’s just a level of generated output not yet incorporated into AI platform UX. This ruling sets grounds for those technologies to come together. Exclusivity agreements have been a barrier to blurring lines between informational answers (AI best now) and experiential results with visual/interactive elements beyond text/chat.

I’m curious where this goes from a UX perspective.

Bryan Casey: I always think about the browser as a portal to the internet. AI is only kind of that, where search is explicitly that now. I go back and forth: Is the browser the right way to think about the main distribution model? I’m a heavy user of both traditional search and AI tools, with slightly different ways I use both, moving back and forth.

To me, they feel like portals into two different worlds: one to interact with the web, the other to talk to an assistant. They feel less like converging into one thing vs. two distinct portals with overlapping Venn diagrams.

Do you ultimately imagine the web and AI converging into one giant blob (hopefully with better UX)? Or living in parallel with obvious intersections but fundamentally different?

Kaoutar El Maghraoui: User experience is already changing with AI-native assistants, chatbots, agents. I expect convergence at some point. We may see new user interfaces not browser-based—neuromorphic interfaces, brain interfaces, voice assistants—things surrounding us. We’re in a time of the biggest revival/revolution in user interfaces not seen in decades. Biggest changes were iPhone and before that graphical interfaces. Now another big change with chatbots.

It won’t be switching back and forth—more converged experience. Today’s interfaces become back-end systems; front-end systems will be new ways of interacting. Back-ends route to traditional APIs or agentic AI. It’ll be a mix, hopefully converging into different interfaces with voice, touch, brain-based interactions.

Mihai Criveti: Mark Zuckerberg is thinking “Metaverse!” Finally, it’s time.

Gabe Goodhart: I have a thought on this—curious if Mihai would jump in. Our mental model of two separate channels is technological—we’re not quite there yet. It’s around the interface. I had a conversation with a colleague deeply in this space: models don’t call tools. A model can only produce a token. A system wrapped around a model can do more—invoke arbitrary code, grab things off the internet, frame things as tools the model generates tokens to suggest steps.

We’re in infancy expanding beyond textual UX for tokens in/out. “Agentics” means evolution of models with expanded conventions for input/output tokens. Right now, no convention for output tokens implying visual/interactive experiences. Models have no inherent knowledge that for quantity, the right output isn’t a number with percent sign—it’s a visual dial or bar chart.

Continued evolution: I’ve seen articles about MCP extending into GUI space—if that happens, model authors can train conventions to generate visual components, bringing AI app UX closer to interactive UI on a web page.

Mihai Criveti: This is exciting. I’ve looked at MCP UI and what Block is doing. We’re doing similar in our project: calling AI agents that render UI components based on results—documents, links, images. Visual representation isn’t standardized—maybe MCP one way, but far from a standard RFC every browser, AI app, phone implements.

Web succeeded due to open standards—my webpage renders in your browser regardless. We’re way away from that with AI apps. AI agents interacting can’t display visual elements. There’s also security risk—you don’t want everyone rendering UI components looking like your bank for login.

Watch this space—lots of evolution in UX. Much built in last six years of compute based on interfaces from early mainframes (teletype terminals in Unix, QWERTY keyboards). All evolutions of previous systems.

I’m eager to see something developed from scratch. AI will enable innovations: what if UI wasn’t keyboard/mouse? Different ways of engaging? AI progress in visual, voice recognition/generation will give those options.

Bryan Casey: Moving to our second topic: Anthropic’s Series F raise. This is a trend—Databricks did Series K. Large, fast-growing companies staying private longer at higher valuations is increasingly the norm.

I want to talk about where Anthropic is in the market. Impressive number/valuation—they’re a big leader. Two ways to think about Anthropic:

1.     Known among AI players as more focused—really going after code use case. OpenAI, Google going after full stack of what models can do. Anthropic carved out a lucrative niche with durable advantage (long in AI years—dog years).

2.     Commercial success of AI so far: two killer use cases—chat and code. That’s where money is. Enterprise use cases spinning up, but big valuations ride on those two. Anthropic has leadership in one.

Mihai, start: Do you see Anthropic as a focused player or a leader in one of two markets?

Mihai Criveti: Also a question of cost-effectiveness. Anthropic has best bang-for-buck model for code. Opus 4.1 is the best planner for AI agents—really good at focused use case but not cost-effective. For general chat, great results but ten times more expensive than smaller, cheaper models.

That’s where Anthropic shines: expensive, complex use cases with high value vs. general chat. They’re great at everything but too expensive for everything. We’ll see niche players carve market: small, tiny model really good at one thing, cost-effective.

Bryan Casey: Makes sense. Gabe: GPT-5 came out—initial reaction bearish, expecting more, didn’t deliver. Over time, sentiment online turned more positive—pretty good model. On MoE podcast, not default coding platform, but people switched to GPT-5.

How durable is Anthropic’s advantage in code space? Lots of speculation on secret sauce—nobody totally agrees. What about that market, stickiness, durability of their position?

Gabe Goodhart: Answer in two rounds:

1.     Meta: How durable is AI in development market, Anthropic or not?

2.     Anthropic relative to overall AI in development.

On first: Maybe foreshadowing next segment. Once you start, you don’t go back. As a developer, hard to get brain plugged into “Does AI fit my workflow?” Once you find a small piece AI slots into, hard to go back. AI durable in development space—removes blockers/friction.

Anthropic in developer space: Model can’t do anything—needs patterns trained to perform in a system. Anthropic building models working well with systems they build around. Durability pinned on those systems wrapping the model.

Back to UX: Claude Code excellent UX resonating with developers. Many other options, but Claude Code sits in default position for terminal AI assistant. They figured out right tools, prompts for tools working well with models—puzzle comes together providing fuzzy but tangible advantage.

They must stay on ball about ecosystem to stay in position. Models’ benchmark scores, fundamental capabilities will level playing field—especially with smaller, specialized models cheaper for specific tasks. Overall tooling/UX is where they’ll hold advantage.

Bryan Casey: Another angle: Anthropic growth—Claude Code primary, but new AI-powered dev tools like Cursor, Windsurf, Replit (vibe coding) successful/fast-growing for prosumer space—not seasoned devs like Gabe, but people getting into coding. AI tools help over hump of initial productivity, environment setup.

As models/systems improve, how large opportunity expanding code market? Two patterns:

1.     Heading to place where anyone can build/prototype anything—standard practice. Forget traditional dev process—get 80% there, throw to professional.

2.     Speed-ups fake—not as big due to refactoring, figuring out what code does, security implications. Path to productivity for vibe coding not straightforward.

If it worked, TAM would go way up. Kaoutar, start: How think about vibe coding, accessibility of programming as major component of market long-term?

Kaoutar El Maghraoui: Huge opportunity for players getting it right. Vibe coding for proof points, small prototyping. Serious coding integrated in enterprise settings, real workflows with safety, compliance, bug-free, Speedos (speed?). Product vs. experimentation, proof of concepts, research.

Claude family of models for code—coding LLMs quietly gaining attraction in enterprise workflows, marketed safer/more reliable for professional environments (legal, healthcare, finance). Coding LLMs major part of product edge.

Anthropic position broader: safe, enterprise-ready, aligned AI. Hype we’re seeing—article discussed hype—now over big hype, entering pragmatic phase: “Can this make money? ROI? Safe? Integrate? Maintain?” More to unpack than vibe coding—efficient, safe, ROI, enterprise-ready.

Space growing—not just one model. How use, wrap, maintain—whole lifecycle. Code generated pragmatic, efficient—integrate in whole stack, run efficiently on hardware, manage compliance, safety. More complex than generating lines of code—integrating in serious workflow takes more work than quick vibe coding.

Who gets that right—huge opportunity. Not completely replacing programmers—boost productivity. Developers needed to understand output of code LLMs, integrate, orchestrate, utilize as part of big design. Other skills required/acquired. Platforms allowing test, debug—maybe other LLMs do reasoning/debugging.

Whole workflow needs done right efficiently, cost-effectively. Anthropic making headways with safety motto, enterprise focus. Others may emerge.

Mihai Criveti: All we need is one field release, developers switch. Last week, three days I switched from Claude Code to Codex. Using Opus 4.1 to write test cases—it looked at code, said test cases failing, removed test cases. I said, “Fix code.” It said easier to fix test cases. Repeated.

Looked at incident report: six days ago, three days Opus 4.1 degraded quality—lower intelligence, tool calling issues with Claude Code. I experienced directly—immediately switched to next model/tool. Lots of choice. If Google Chrome not working, switch to Edge, Firefox, find another browser—not “for three days, I won’t use internet.”

Bryan Casey: Great transition to final segment: ROI, hype cycle, plateau of productivity. Still in hype cycle, but aspects transitioning to real productivity—how get value. But online, another segment in totally different place—market for AI cynicism.

Not sober analysis (“too hyped, still a fan, not quite there”)—much further: rooting for failure, talking bubbles, economy collapse, destroying planet. What motivating? Simple: fear of change? Beyond traditional tech skepticism.

GPT-5 flashpoint—people: “Oh, it’s over. Progress stalling.” Progress curves over time seem normal. Use cases valuable. How read situation? Any plausible universe fourth AI winter in this cycle? Or permanent productivity from here? Gabe, looked winding up.

Gabe Goodhart: Teased last one. Why? Expectations game. Certain AI community members fond of projecting future—built brand on it. Some project utopian future: AI plays roles humans don’t like. Others same future—pessimistic, worried. Don’t rule out either—love sci-fi.

Place today: technology phase change. Nothing novel. Mihai said: If Chrome update segfaults, you don’t stop using internet—find different browser. Same rapidly becoming true for AI in daily lives, even ways not realizing. If today lost access to all AI models, we’d feel pain—even AI skeptics, somewhere crept into daily workflows.

From technology phase change: no. Article spoke previous AI winters—commonality: funding/usage isolated in small number specialized users oversold on potential capability, disillusioned. With technology now ubiquitous, generative AI not going away. Funding for elements looking at further-reaching futures may go away.

Reason appetite for skeptical view: human nature evolutionary, competitive survival of fittest perspective. Pre-internet, not deep literature imagining future humans communicate instantly. AI occupied place in human psyche: what if? How much of me reproducible in machine? Books, stories timeless.

Skepticism keys in dystopian view if achieve that. From technology perspective, here to stay. Steady investment not going away soon.

Kaoutar El Maghraoui: Totally agree—not heading for full-blown AI winter. Industry experiencing slowdown, but more reality check. Debate over GPT-5 highlights critical gap between optimistic promises (bulls) and practical concerns. Period of slowdown—opportunity for industry focus on real-world problems, sustainable business models.

Historically, past AI winters occurred when overhyped promises failed to materialize, leading to sharp drop in funding/interest. First AI winter: failures to achieve human-level intelligence. Second: limitations of expert systems. Lesson: unrealistic expectations biggest threat to AI progress.

Today: tons of investment made. AI embedded in billion-dollar products (ChatGPT, Claude, Copilot). Past AI winters didn’t have real market traction. Huge infrastructure built—GPUs, cloud data scales—innovation doesn’t vanish, compounds.

Similarities with past: hype vs. reality. Investors expect AGI leaps, progress more incremental (GPT-5 slower vs. GPT-4, but incremental). Overreliance on benchmarks—like expert systems failed out of lab, LLMs show brittleness in wild. Friction applying to real problems—parts fail, parts work, refinements, reality check.

Today’s ecosystem has sustained revenue stream—financial base didn’t exist in past ventures. Here to stay. Keep making progress—ups/downs like any big technology shift. Industrial revolution, printing press—storytellers replaced by print, many unhappy. Prints/publishing did to industry/lives. Shift and reality check for AI, not winter.

Mihai Criveti: Different take: progress tremendous but invisible due to shift in model cost—“intelligence too cheap to meter.” Models integrated more, becoming transparent—part of application workflow, behaves better, fast, cheap, effective—you don’t see it like ChatGPT.

Shift focus from building models for raw performance to models like GPT-5 with router routing to different models. Not perfected yet, but direction positive. Price changes: Nano $0.05 for 1 million tokens—insignificant, amazing.

Continue direction: AI embedded more into application workflows, real-world systems beyond consumer typing box for answer.

Bryan Casey: Great place to end. Mihai, Kaoutar, Gabe—thank you for joining. Listeners: like, subscribe if fan of pod. See you back next time on Mixture of Experts.