Guang Mi on the Diverging Paths of Large Models, AI Products, and the Startup Window

Guang Mi opens with his conflicted feelings about AI investing and startups. The technology makes visible progress every month, which is exciting, but leading model companies are moving so fast that they keep squeezing the opportunities available to venture-backed startups. A small company facing giants often has few strong cards to play.

Zhang Xiaojun responds that his assessment sounds “pessimistic and painful.” Guang adds that leading foundation-model companies are all starting to build products, so AI founders have to think about how to avoid competing directly with giants.

Guang says the first six episodes focused on the upper limits of model intelligence, while the past two or three months have shifted his attention toward products. He also cites the phrase “finance for Jewish people, AGI for Chinese people” to summarize his view of opportunities for Chinese founders.

Zhang introduces the seventh episode of The Global Large Model Quarterly: this season, Silicon Valley model companies are beginning to diverge. Google Gemini and OpenAI still pursue general capabilities; Anthropic focuses on coding and agentic systems; and Thinking Machines Lab, founded by Mira Murati, explores multimodality and next-generation interaction.

Zhang welcomes Guang and explains the delay: he has just returned to Beijing from a business trip. They look back on the previous six episodes and agree that recording their thinking at each stage is valuable, like an archive of ideas that evolves over time.

Guang hopes listeners will remember three key ideas: model companies are both diverging and converging; product competition takes the form of horizontal bundles and vertical integration; and AI products need to be judged by whether they can deliver a Level 4 experience.

Zhang asks whether the quarter was uneventful. Guang says the opposite: the industry is more turbulent than ever. Google and OpenAI keep expanding across multimodality, reasoning, coding, agents, and consumer products, and Google is even exploring world models, while other labs begin focusing on specific areas.

Guang says the global AI story is not quiet at all; it has entered a period of explosive growth. The clearest signal is that model companies are specializing according to their strengths: a few teams continue to pursue general capabilities, while others choose a more focused field.

Zhang asks when Anthropic decided to bet on coding and agents. Guang thinks the shift became clearer after Claude 3.5 launched in the summer of 2024, when positive developer feedback gave the team powerful reinforcement.

Developer feedback on Claude may have led Anthropic to focus fully on coding and agents, gradually stepping back from multimodality, consumer products, and standalone reasoning models. Guang compares this intense positive reinforcement to a team doubling down under “reward hacking.”

Guang cites public revenue estimates to illustrate Anthropic's growth: its annualized revenue may have been below $100 million at the end of 2023, around $950 million in 2024, and could exceed $12 billion in 2025, perhaps reaching tens of billions in 2026. Even a $170 billion valuation may not look expensive.

He adds that token consumption and revenue at OpenAI and Anthropic are both growing rapidly, with month-over-month increases perhaps reaching 20–30%. Industry data he has seen suggests that AI labs may be adding more annualized revenue than publicly traded SaaS companies are adding.

Guang sees Anthropic's core strategy as focus on coding and agents. If it tried to cover multimodality and consumer products at the same time, like Google and OpenAI, it would spread itself too thin. But if those two giants put all their weight behind coding, Anthropic would face pressure too.

Zhang asks why Anthropic's focus on coding is a non-consensus bet. Guang suggests that the long-term value of coding and agents may far exceed expectations; like world models, they could be important paths toward broader intelligence.

Guang compares coding and agents to tools that expand what people can do in the digital world. Humans use language and tools to expand their sphere of activity; agents that master language, code, and tool use could perform all kinds of operations people can do on computers and phones.

He gives creating a website or a PowerPoint deck as examples. In the past, that might require hundreds or even thousands of back-and-forth steps with a studio. An agent can use reasoning and coding to generate an interactive webpage, one-off software, or a PDF. The output format is not the main point; coding and agent capabilities are.

Guang also believes Anthropic's team is stronger than people outside the company realize. In earlier episodes, they discussed how Kimi, Qwen, and DeepSeek were no longer following OpenAI in some capabilities and were instead drawing on Anthropic's approach.

Zhang mentions that some people describe Kimi's move from “China's OpenAI” toward “China's Anthropic.” Guang thinks that is smart and sees DeepSeek and Qwen moving in similar directions in some areas.

Zhang asks Guang to introduce Mira Murati and Thinking Machines Lab. He calls it one of Silicon Valley's strongest recent startups, valued at about $10 billion and having raised $2 billion, and says the team is highly cohesive.

Guang describes Thinking Machines Lab as a composite spinout of core OpenAI product and infrastructure talent. For example, Meta offered an enormous amount to recruit its infrastructure lead, Andrew, but was unable to bring the rest of the team along.

Mira was OpenAI's CTO, so she understands what OpenAI does well and where it falls short, and has insight into what the company may do next. Guang speculates that she may steer Thinking Machines Lab toward multimodal-native systems and next-generation consumer interaction.

Guang thinks ChatGPT has built a substantial lead in chat products, so Thinking Machines Lab may explore a new kind of interaction instead of copying the same product. The new form may still rely on the phone, since a great deal of personal context is already stored there.

He does not believe AI will make phones less important. Phones have evolved from devices for calls and texts into hubs for entertainment, connectivity, and data storage. The next generation of smartphones may be more proactive and have new interaction properties.

Guang calls OpenAI, Gemini, and Anthropic “intelligence-first” labs, then raises another question: might Mira adopt a “product-first” culture, given that existing intelligence is already sufficient for exploring excellent products?

Guang further speculates that Mira's product-first thinking may align with Apple's direction and says it will be worth watching whether Apple's board invites her in the next year or two. If Apple lacks a team that truly understands AI, it will be hard to plan the core experience of its next-generation phone.

Zhang asks what “multimodal-native” means. Guang explains that the context people enter into ChatGPT is far richer than the few keywords they type into a search box; voice, dynamic input, and agents operating software bring even more context.

As users express themselves more naturally, use voice, or let agents operate computers and integrate software, interaction will go beyond plain text. Zhang adds that voice may capture more context; Guang says it is not just voice, but more natural human-computer interaction overall.

Zhang turns to the choices made by xAI, Grok, and Meta. Guang says Musk's bet on a gigantic compute cluster is significant, but so far its compute advantage has not produced a certain payoff. A massive cluster is the baseline for not falling behind, not a guarantee of overtaking competitors.

Grok's position in the ecosystem remains unclear: chat puts it up against ChatGPT, search against Google and Perplexity, coding against Anthropic, while it has yet to make a clear choice in multimodality or world models. Guang worries that Musk wants too many things and lacks focus.

Guang says Grok may consume far more resources than other models without delivering a proportional return; even Doubao may provide a better end-to-end voice experience. He also worries that xAI could fall behind without a larger organization to support it.

If xAI eventually has hundreds of thousands of next-generation GPUs, it may achieve a step change in its models through sheer cluster scale. Musk is very good at building data centers. Until then, however, scale by itself does not prove an advantage.

On Meta, Guang thinks its Superintelligence team is unlikely to reach the global top three in the short term; even xAI, which Musk has pushed aggressively, has not done so yet. It remains to be seen whether Meta is intelligence-first or product-first and whether its team can build something from zero to one.

Zhang mentions Meta's large-scale recruiting. Guang thinks this is a pivotal moment for Mark Zuckerberg to deploy his personal resources and execution ability. But even if the top three Silicon Valley model companies are close in capability, it is still difficult for others to break into that group.

Guang summarizes the two key paradigms of the last couple of years as pretraining scaling and reinforcement-learning scaling, and notes that OpenAI drove both paths. The model roadmap is becoming clearer, but companies may still diverge in their products and ecosystems.

Zhang asks where each company will fit in the long-term ecosystem. Guang says that today they are all called tech companies, but Google does search, Meta social media, Amazon e-commerce, and Apple phones. AI companies may all be pursuing AGI now, but they will differentiate further over time.

Using what he calls a deliberately rough analogy, Guang casts ChatGPT as the next-generation Google and Anthropic as the next-generation Windows: coding provides APIs through which applications can expand almost without limit. There are only a few operating systems and platforms that can truly support such expansion.

He expects Google Gemini to remain strong, with possible advantages from economies of scale. Zhang asks how an investor or job seeker should choose a company, and Guang cautions that the end state of models changes too quickly to know for sure which one will win.

A leading model's capabilities may become the market average in just three to six months; the competition resembles Formula One, with teams constantly passing one another. Model generations used to arrive every six or twelve months, but now the cycle is two or three months, and a single mistake can make a company fall behind.

Asked how he would allocate $100 in stock investments, Guang gives his current subjective allocation: $40 to OpenAI, $40 to ByteDance, $10 to Anthropic, and $10 to Google. He stresses that this is only his judgment at this particular point in time.

Guang assigns a larger share to OpenAI because ChatGPT has nontechnical moats far beyond model capability alone. Zhang notices that this differs from his allocation in the previous episode, showing that his view of product value is changing.

Guang says OpenAI's advantage is not only its model. Over the past quarter, model capabilities have gradually converged, but ChatGPT's reputation and brand have continued to strengthen, with clear consolidation among consumer products at the top.

ChatGPT may become the default AI choice for many users, much like Google became the default search engine. New users no longer compare models constantly, and ChatGPT or Doubao may sit in the easiest-to-reach spot on their phone's home screen.

ChatGPT has become the default AI tool for many new users; some even put it or Doubao in the most convenient position on their phone's home screen. Guang thinks brand recognition and user habits are now taking root quickly.

Guang compares model training to Netflix's investment in content: spending on a leading model becomes product experience, a place in users' minds, and brand, unlike traditional paid acquisition, which disappears once the money is spent. ChatGPT did not rely on massive advertising campaigns to acquire users.

As model capabilities keep improving, they can also bring back users who once canceled their subscriptions. Guang thinks ChatGPT's long-term retention could be exceptional, perhaps approaching a product like WeChat; the best model will still capture the most traffic.

Guang compares it to Coca-Cola: even if someone comes up with a slightly better-tasting formula, consumers may still think of Coca-Cola first. ChatGPT's place in the brand consciousness can form a similar nontechnical moat.

On revenue scale, Guang cites public estimates of about $12 billion ARR for OpenAI and $5–6 billion for Anthropic, saying there is already a clear gap between them and other AI products, and that they are still growing quickly.

Guang expects Claude Code could soon reach $1 billion in ARR and might even surpass Cursor. Products launched by model companies are widening the revenue gap, again illustrating how the strongest models can capture substantial value.

OpenAI may soon become one of a new generation of tech giants. For investors and founders, the growth of leading companies is both exciting and anxiety-inducing: they are continually consuming opportunities for startups and venture capital.

Zhang asks why foundation-model companies build agents themselves rather than just charge for API access. Guang explains with his “mining” analogy: model companies spend tens of billions of dollars to mine intelligence, so if applications capture most of the proceeds, the model companies naturally move up the stack into products.

Claude Code launched eight or nine months after Cursor, yet caught up and even overtook it in just a few months. Guang believes foundation-model companies are shifting from relying solely on an API ecosystem to building agents end to end, because APIs do not provide a sufficiently strong moat or retention.

Guang believes there is still room for independent agent startups because they can start with APIs and deliver an experience quickly with a small team. Claude Code's success also shows how much value an application team can create, although model companies' internal advantages and speed of response will keep putting pressure on them.

Guang expects Anthropic's experience in coding to transfer to other fields. For founders, the window is very short: today's tech giants understand new technologies and are willing to commit company-wide resources to them, unlike the old giants of the mobile-internet era.

Zhang asks whether agents using external models still have an advantage. Guang says of course they do, and identifies two trends: horizontally, bundling many capabilities into an all-in-one suite; vertically, integrating chips, models, agents, and an application ecosystem.

ChatGPT's suite uses a single subscription to bundle chat, search, coding, agents, Workspace, PowerPoint, and more. This puts pressure on standalone vertical tools because mainstream users do not want to pay repeatedly for each capability.

Guang thinks an all-in-one suite will meet many general-purpose needs and may limit the market for specialized agents. He further imagines that PowerPoint could merge with web coding, perhaps making traditional file formats unnecessary.

Google Gemini demonstrates vertical integration: from TPU chips to the Gemini model and agents, then onward to Docs, Chrome, Android, and YouTube. Even if its model temporarily lags by a few months, Google's end-to-end scale could give it staying power.

Guang admits it is difficult for startups to go toe-to-toe with giants in their core markets, but remains optimistic. When the industry structure seems settled, new structural opportunities and non-consensus founders can still appear.

He recalls that in 2015 many people thought Taobao's e-commerce scale effect could not be challenged, and therefore missed Pinduoduo and Shein. Today, too, investors need to find founders like Huang Zheng who understand the industry structure deeply and can break through where giants are not looking.

Guang believes investors should spend enough time understanding a founder's underlying logic rather than just looking at the surface-level product. He is still searching for “today's Huang Zheng,” someone who can identify a new opportunity in a market full of giants.

Zhang asks whether the model or the product matters more. Guang says the first six quarterly reports focused heavily on the intelligence ceiling, but over the last two or three months the focus has clearly shifted to product, because technical gains from models must become a commercial moat.

Two observations partly changed Guang's view: ChatGPT has built powerful traffic and brand moats, and in coding, even though Anthropic leads, Gemini 2.5 Pro can catch up quickly. A lead in model capability may not last; a product moat is more durable.

Progress by DeepSeek, Gemini, and Claude has not fundamentally shaken ChatGPT. Guang thinks ChatGPT has accumulated many nontechnical advantages, while coding tools and model companies often have only technical advantages.

He notes that Sam Altman once articulated two parallel ambitions: bring ChatGPT to one billion daily active users and achieve AGI. Many labs may achieve AGI, but not every company can build a product with one billion daily users.

Guang thinks OpenAI strikes the best balance between exploring intelligence and running a product business. Its research is in the global top tier, and it turns the model dividend into product traffic and brand recognition. Mass distribution to consumers has enormous commercial value.

Even if Anthropic, Gemini, or Grok has a good model, that does not mean it will be equally strong with consumers. ChatGPT's brand value grew out of a unique social window and accumulated users; even if its model stopped advancing temporarily, its distribution advantage would remain strong.

If models continue improving, OpenAI is also among the top tier in research. Guang believes the gap between leading models may be only three to six months, making a technical experience short-lived; that is why product competition matters more.

Zhang sums up that investment in model capabilities also recharges the brand. Guang agrees that ChatGPT's biggest moat is not technology but brand; even if someone has a better “formula,” consumers may not change their first choice.

Zhang asks when Thinking Machines Lab might launch a product. Guang says the team is strong, but a new company lacks a mature distribution channel. It may be able to use Apple in the United States, but independently building a consumer product on ChatGPT's scale is difficult; Doubao may have a better chance.

Guang compares model companies to miners and product companies to refiners who process a discovered vein. Timing is critical: even with high token costs, the first product to deliver a magic moment that amazes users may earn attention equivalent to $500 million to $1 billion in advertising.

Perplexity, Cursor, Manus, and DeepSeek each mined important model-driven advantages, such as information retrieval, coding, or agents. Guang points out that as model companies gradually catch up in these capabilities, the startup window is shrinking.

Perplexity once had a window of nearly two years because model companies took a long time to make web search work well. Cursor went from gaining traction to Claude Code's launch in about nine months; Manus had only three or four months before similar products followed.

A shrinking window does not mean application startups are doomed; it means competition is more intense. Product companies can still start with APIs. Using Claude Code as an example, Guang believes a small team can still build a product worth hundreds of millions of dollars or more.

Guang thinks in-house teams at model companies understand model tools and training details better, which can give them an advantage in how they call models. But if agents perform equally well, an independent startup's token costs are burdened by API margins—a structural disadvantage.

Guang says cost is an important source of unfairness: even if a startup's agent performs just as well, its subscription revenue may not cover high token consumption. The foundation-model company providing the API also needs a gross margin, and the standalone app can be crushed by costs.

AI products' success so far has largely come from the research dividend of models. Chat, search, coding, tool use, and PowerPoint are each like a sensor on a smartphone, enabling new kinds of internet products.

Zhang asks Guang to describe a great AI product manager. Guang thinks the next generation of product leaders will likely come from algorithms or models, but will also need product sense and business judgment to capitalize on model capabilities and anticipate technical changes over the next six to twelve months.

Guang defines a product's north-star metric as whether it can deliver a Level 4 experience at scale: users are impressed from start to finish, and the product truly completes a task rather than merely displaying model capability.

He says the two best agents at the time had reached Level 4: ChatGPT Deep Research for information search and Anthropic's Claude Code for software development. Once users get that experience, they are willing to upgrade from $20 a month to $200.

Search and coding reached Level 4 first because language and code are today's biggest model dividend, and because white-collar work is valuable, programmers are highly paid, and software development is easier to close the loop on. Robotics also has to solve the added challenge of hardware iteration.

White-collar work is more likely than blue-collar work to reach a Level 4 experience first, and programmers are among the most valuable white-collar workers. Robotics still needs hardware iterations; if coding has not yet reached Level 4 at scale, other physical-world scenarios will take even longer.

Guang attributes the success to accumulated data: the internet has decades of web pages, and GitHub has decades of code. This data abstracts many kinds of human activity; once you enter the physical, atomic world, the speed of iteration slows down.

Zhang asks whether Claude Code succeeded because of its product or its model. Guang estimates that about 90% came from model capability and 10% from product and engineering, especially its management and compression of context for large codebases.

On the competition between Claude Code and Cursor, Guang says Claude Code's revenue is growing extremely quickly. If Cursor copies the same experience, it still has to bear token costs several times—or even ten times—as high under a fixed subscription fee, which is difficult to sustain as a long-term loss.

For example, with a subscription of about $200 per month, some users can consume thousands of dollars' worth of tokens in just a few days. Costs may be several times, or even ten times, revenue. Guang thinks Cursor could easily be dragged down if it has to sustain those losses over the long term.

Guang adds that Claude Code continually reads the codebase and compresses the context while preserving important information; Anthropic takes long context seriously. Internal tool naming and training experience may also make model companies better at using tools.

Guang even thinks coding companies that do not build their own models may lack a long-term advantage, with competition eventually shifting to cost. Cursor may need to join forces with a major company such as Microsoft, Google, Grok, or Meta. Still, the model landscape is changing.

Zhang points out that the boundary between models and applications will keep changing. Guang says AI products are still at a very early stage and ChatGPT and Google may gradually converge into all-in-one suites, much as Xiaohongshu has combined search, short video, and a social feed.

Guang uses Xiaohongshu as an example: many young users already treat it as both a search engine and a social feed, with search, short videos, and social networking merging. He thinks ChatGPT and Google may evolve in a similar all-in-one direction.

Asked what might be the next Level 4 experience, Guang is optimistic about Office, Workspace, PowerPoint, and Excel data analysis. They involve structured content and verifiable results, and overlap with coding; but there is no clear reward signal for what makes a “better PowerPoint deck.”

Workspace may have a much larger user base than coding: its audience is no longer just tens of millions of programmers but hundreds of millions of office workers. Guang then talks about how his view of Google has changed, and expects ChatGPT and Google eventually to converge on similar product forms.

Guang speculates that ChatGPT will eventually build an advertising platform because OpenAI hired a CEO experienced in ad monetization. A small group of heavy users may pay hundreds or even thousands of dollars per month, while most users use it for free and see ads.

Google still has the world's strongest advertising platform. If products converge into all-in-one suites, search and ad matching could become more effective. Google is also launching ad products to stimulate long-tail demand, and its scale effects remain powerful.

Guang says ChatGPT currently leads through brand recognition, while Google may eventually catch up through scale. ChatGPT has already changed user habits: from entering two or three keywords in a search box to describing dozens of keywords or a whole request to AI.

More user context can help match needs and improve inference efficiency, but whether Google can change user habits remains the biggest uncertainty. Google has several strong cards, including its own chips and models, voice technology, and operating system.

The exact form of the next main interface is still hard to imagine, but Guang thinks services will certainly converge. Agents may break down the end-to-end silos between apps and call different services directly; the interface could lie somewhere between a phone screen and an app.

On the question of an AGI bubble, Guang suggests looking at what the smartest people are using: the widespread adoption of Claude among quantitative finance professionals and excellent programmers is a sign that demand is real. He is more concerned about whether the Silicon Valley giants' combined capital spending of hundreds of billions of dollars this year is creating value at the application layer.

He worries that capital may circulate between chip companies and model companies, creating “nested revenue,” while front-end applications create no real value. Zhang asks who would be hit first if the bubble burst; Guang thinks capital would concentrate in the leading companies.

On a new topic from the Bay Area, Guang says America's core assets are the dollar's role in global finance and Silicon Valley's technological lead. China's industrialization capabilities are eroding U.S. advantages in other areas, while American finance and technology sit at the top of the social pyramid.

Guang describes a cycle of capital: newly issued dollars are diluted in value, and money flows into stocks of technology giants. Chinese people used to buy homes; Americans now buy the “Magnificent Seven” stocks. He worries about a system dominated by capitalists, tech entrepreneurs, and merchants.

He is skeptical about the revival of U.S. manufacturing, believing many young people do not want to do dirty, tiring work and would rather invest or make money from technology. If Silicon Valley no longer leads, can the dollar's financial advantage still hold?

Guang speculates: if China makes a breakthrough in advanced lithography within two years, would global capital flow back toward China? It shows the connection between technological leadership and the status of a financial center, although the outcome remains hard to predict.

Zhang mentions the influence of Chinese founders in global AI. Guang cites the saying, “Jewish finance, Chinese AGI,” and notes that China has ample domestic supply and strong manufacturing capabilities, but comparatively limited domestic demand.

He thinks Chinese teams need to actively expand into global markets, especially North America, where willingness to pay is strong. After the Industrial Revolution, Britain developed the East India Company; after the Second Industrial Revolution, the U.S. produced multinational corporations. This may now be an important moment for Chinese founders to build global companies.

Guang says Chinese founders' key strengths are fast iteration and learning, while their weaknesses must be overcome in global competition. The key to entering North America is to establish roots in Silicon Valley and first join the network that spreads through YC and San Francisco's AI community.

He lists several non-consensus views: even when the landscape seems settled, “the Huang Zheng of this era” may emerge; the value of teams such as Anthropic and Thinking Machines Lab may be underestimated; and coding's contribution to industry may also be underrated.

Guang believes the first breakthroughs in robotics may not come from teams with today's dominant computer-vision backgrounds; language, code, and pattern capabilities may matter more. Robotics must cross several levels of maturity comparable to GPT-4, so it may not become widespread quickly in the near term.

He is very optimistic about Chinese founders over the next three to five years. Comparing human and gorilla intelligence, Guang says humans have language, tool use, and the ability to create tools, which gives them a much larger space to explore.

Zhang asks Guang how to invest in AI and build a fund with 30x DPI. Guang notes that technology changes every month and leading companies are converging; OpenAI itself is starting to look like a VC that allocates capital and research resources internally.

OpenAI and other leading model companies are squeezing out agent-startup opportunities, making AI investing very difficult. Guang advises continually updating one's investment strategy and spending enough time understanding the logic and opportunities of a new generation of founders.

Zhang asks whether traditional VC's army-of-analysts approach and diversified investing have stopped working. Guang thinks concentrated investment is more effective at this stage. If you miss companies such as OpenAI, Anthropic, xAI, Perplexity, or Cursor, the whole investment cycle may have been for nothing.

Looking back at the mobile internet, the major opportunities were WeChat, short video, information distribution, food delivery, e-commerce, and ride-hailing. Investors need to catch a few big winners—for example, Hillhouse invested in Pinduoduo, 5Y Capital in Xiaomi and Kuaishou, and Sequoia in ByteDance.

Guang does the math: a $300 million fund would need to create about $10 billion in value to deliver a 30x return. Even so, he remains optimistic about AI, since annual AI capital expenditure in the U.S. is around $300 billion and could reach a trillion over three years.

He compares the enormous AI investment to making compost for the future. Mobile-internet infrastructure once prepared the ground for apps such as WeChat and Pinduoduo; today's trillions of dollars in AI investment may likewise nourish a future boom in products.

Guang cautions that investment strategies must continually be challenged by technical changes and convergence among leading companies. Concentrated investing may work today, but that could change. Since ChatGPT launched, OpenAI and Anthropic have been the biggest winners.

Many Silicon Valley VCs passed on OpenAI and Anthropic because AI capital expenditure was so large and gross margins seemed insufficient; some also missed Mira Murati's high-valuation startup. Guang thinks identifying excellent founders is itself a non-consensus opportunity.

He advises investors to focus on supporting the fastest-growing and strongest founders and to have in-depth conversations with them every quarter. As examples, Guang cites Zhang Yiming's deep insight into information distribution and Huang Zheng's into social commerce.

Zhang concludes that new opportunities can still arise when the landscape appears settled. She asks listeners to share which company they would join and which CEO they favor over the next four years. The hosts thank the audience and preview the next episode.

Zhang Xiaojun closes the show, inviting listeners to subscribe to Zhang Xiaojun's Business Interview on Xiaoyuzhou, Apple Podcasts, Ximalaya, and QQ Music. She also welcomes comments recommending guests and topics, and says they will meet again next time.