The next Google isn't search, it's a task engine
ChatGPT, Anthropic and Perplexity look nothing alike, but they are all fighting over the same thing: reorganising information and tasks. Search and recommendation merge into a task engine, and whoever holds context wins.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
ChatGPT is growing faster than TikTok did
ChatGPT has passed 300 million weekly actives, corresponding to 500-600 million monthly actives, and by this time next year it could be over 1 billion monthly actives. That growth rate is the fastest of any tech product globally, beating TikTok's earlier record of reaching 500 million-plus monthly actives in four or five years. Guangmi thinks hype is a factor, but the more fundamental point is that ChatGPT is heading in the direction of being ‘the next Google’ — its opponent and its goal are not just ChatGPT, but how to beat Google.
— Guang MiGreat companies come up from the edge market
Yahoo's portal model could only list the head of the web; Google used keyword search to index the entire internet and reorganised information distribution. Sequoia US backed Yahoo first, and its Google investment was meant to help Yahoo cover long-tail web content, because the long tail can only be triggered through the keyword model. People building portals at the time all thought keyword search like Google's could not get big, but Google grew up from Yahoo's edge market. Guangmi stresses that you must take companies that rise from edge markets seriously.
— Guang MiAnthropic is taking the AI-era OS path
OpenAI is turning itself into the only K-Lab, running away with the consumer side; Anthropic is heading down the AI-era OS operating-system path, occupying Coding and protocols, especially the recently released MCP. Perplexity's biggest innovation is redefining the form of AI search — it is AI using the search engine for you, and it is also an Agent, where users can keep asking and interacting around a single topic. Products will all eventually become task engines or task containers.
— Guang MiChatGPT's business model is 10-20x less efficient than the old internet
ChatGPT is a classic tool-type product, and a 5% paid rate would already be excellent, but Guangmi guesses it may only be 2%-3%, corresponding to roughly $0.5-0.6 per MAU. Compare Douyin, Taobao and WeChat: close to several hundred billion or even a trillion dollars in revenue, 1 billion monthly actives, $100 per MAU per year, $8-10 per month. Traditional internet products monetise 10-20x more efficiently than ChatGPT. Backend costs are surging 5-10x a year, the front end cannot squeeze out revenue, and it is not sustainable.
— Guang MiContext is the payment of the new era
Guangmi thinks 99% of practitioners today are staring at the model's generation capability or coding capability, but the more core thing is the ability to collect context. Without context synchronisation, the success rate of the vast majority of tasks is very low. He uses the analogy that context is the payment of the new era — e-commerce's two wings are logistics and payment, and without those two pieces of infrastructure the online shopping experience is terrible; when an AI model does tasks for you, success does not depend entirely on generation or coding capability, but more on whether context is fully synchronised. Context acquisition should be automated, not manually prompted.
— Guang MiOpenAI's organisational problems are no smaller than Google's
OpenAI's success today rests heavily on the mindshare and brand dividend from very strong early research and a far-ahead technology lead, but over the past year or two it has not caught that technology dividend well. The two most typical places are search and coding, where it is clearly not number one today. Search left Perplexity a full two-year window, and coding was overtaken by Anthropic's Claude Sonnet, with a large number of developers migrating to the Claude Sonnet ecosystem. Guangmi thinks it is still an organisational problem, or something somewhere is wrong — that many old-timers leaving is not a good thing.
— Guang MiLong Horizon Task is the next focus
Guangmi thinks the biggest thing to look forward to in 2025-2026 is Agent deployment, especially long-distance multi-step tasks — Long Horizon Task. Several big-name figures in Silicon Valley's core circle are working on it, and OpenAI CTO Mira leaving to do Long Horizon Agents is another example. It is comparable to Perplexity's Aravind leaving back then to start a company doing RAG-based search. Long Horizon Task deployment may be the most core direction for Agent startups, and the leading model companies OpenAI and Anthropic have both put a lot of effort into it.
— Guang MiIs the pretraining data wall permanent or temporary? There is disagreement
The Data Wall Ilya raised is a pretraining data plateau: internet data grows linearly, adding less than 1T of effective tokens a month, but the data model pretraining needs grows exponentially. Guangmi thinks pretraining has hit trouble today, one hundred percent, and the bottleneck is not compute or architecture but data. The large-scale training data on the internet that can raise intelligence may only be twenty or thirty T. Some are pessimistic, some optimistic; the optimists think there is still a lot of room for data mining, and the key is raising Data Efficiency — a human needs five to ten samples to learn a piece of knowledge, while a model needs thousands or tens of thousands.
— Guang MiIn their own words · checked verbatim
I do feel you absolutely have to take companies that rise from edge markets seriously — great companies, more often than not, grew big from edge markets.
我是感觉一定要重视边缘市场起来的 往往伟大公司 都还是从边缘市场做大做起来的
Guang Mi5:02
I think 99% of practitioners today are fixed on the model's generation capability or coding capability; I think a more core thing may be the ability to collect context. Without context synchronisation, the success rate of the vast majority of tasks is very low.
我觉得99%的人从业者 今天都指定着模型的生成能力 或者coding能力 我觉得更加核心的一个 可能是context的采集能力 如果没有context的同步 其实绝大多数的任务 成功率都非常低的
Guang Mi20:13
There is a simple analogy: context may be the payment of the new era.
有一个简单的比喻 就是context可能就是新时代的支付
Guang Mi21:15
Everyone says Google's organisational problems are huge; looked at this way, OpenAI's organisational problems are actually not small either.
其实大家都说 Google的组织问题很大 其实这样看 OpenA的组织问题 其实也不小
Guang Mi28:21
I think pretraining today — I think looked at today, it has hit trouble, one hundred percent.
我觉得今天预训练 我觉得今天看应该是百分百 遇到困难了
Guang Mi1:11:49
Figures
| ChatGPT weekly actives | over 300 million | 3:02 |
| ChatGPT paid rate | 2%-3% | 12:10 |
| ChatGPT revenue per MAU per month | $0.5-0.6 | 12:10 |
| Traditional internet product revenue per MAU per month | $8-10 | 13:10 |
| Cursor's new-round valuation | $2.5 billion | 59:45 |
| Limit of trainable internet data | twenty or thirty T | 1:11:49 |
| GPU scale of the world's first-tier companies | 100,000 cards | 14:10 |
Glossary
- Long Horizon Task
- A complex task requiring multi-step, long-horizon planning to complete; the core direction for Agent deployment.
- MCP / Model Context Protocol
- An open-source protocol standard from Anthropic that connects models to external tools and data sources, with large long-term impact.
- Data Wall
- The pretraining data plateau raised by Ilya: internet data grows linearly, but model pretraining needs exponential data.
- Reward Model
- A model used to evaluate the quality of model outputs; the O1 route depends on it, but general generalisation remains a hard problem.
- K-Lab
- Guangmi's term for consumer-facing AI labs like OpenAI, as distinct from the OS path.
How to listen
Founders and investors watching AI product form and investment judgement, especially those who want to understand 2025-2026 Agent deployment and the context race.
The recap and keyword summary after 1:22:19 can be skipped quickly; everything before it already covers the ground.