Summary

The technology consensus at mid-2025 is that the industry is moving from L2 reasoning models to L3 Agents, but “standard equipment does not mean the end.” Intelligence in both base and reasoning models continues to improve in parallel. In the U.S., OpenAI, Anthropic, Google, xAI, and Meta are taking turns resetting the frontier; in China, DeepSeek R1, Qwen3, and Doubao 1.6 are closing in on the leaders in benchmark rankings. Coding is the first major battleground for Agents because code exists in a purely online environment where “right is right and wrong is wrong,” making it naturally suited to reinforcement learning.
The Agent capability metric worth tracking is not chat scores, but how long a task it can complete continuously. From 2019 to 2025, the duration of tasks models could complete roughly doubled every 7 months; looking only at 2024–2025, after reasoning models emerged, the interval shortened to about 4 months. That makes the shift from minutes to hours and even days less distant than it sounds. “Agents working for several hours” will push the Token consumption of a single model call up by several orders of magnitude.
Falling unit inference costs will not automatically reduce total industry spending; AI is replaying “Andy gives, Bill takes away.” The chart may show GPT-4-related capabilities rising roughly 100x while costs fell roughly 100x. But reasoning models can consume 5–10x more than base models for thinking and answering, while Agents stretch tasks from seconds to hours. 庄明浩’s conclusion is that software will keep “eating up this additional multiple of capability,” so demand for compute, Tokens, and inference infrastructure will continue rising.
There is still no consensus on whether the Scaling Law has failed; the industry is testing synthetic data, reinforcement learning, and “having models grade models.” AlphaGo Zero-style self-generated data and using another model as a reward model both resemble a self-lifting loop. DeepSeek R1-0528 is a concrete example, with markedly stronger math and Coding capabilities. The market is watching V4 and the unreleased R2 because if V4 is stronger than V3 and is trained along the R1-Zero-to-R1 reinforcement-learning path, R2 “will definitely be very strong”; the only uncertainty is how strong.
Rapid intelligence gains do not mean Agents will quickly replace white-collar workers; L3 may become a protracted campaign. Flow, a Kimi technical staffer, estimates that an Agent capable of replacing 80% of white-collar work may take 5–10 years or longer unless a disruptive technology emerges; others see 2–3 years, and 庄明浩 preserves the disagreement. Meanwhile, China’s open-source models have become a more certain industrial variable: DeepSeek and Qwen are pushing the capability curve higher, while interest in open-source projects around Agent architecture and workflows has grown by more than 70%.
Product competition is converging on three things: the browser as the Agent’s vessel, visible execution, and pay-for-results. Browsers can carry memory and planning, invoke tools, and execute actions, which is why Chrome and Safari, along with Quark and QQ Browser, are again becoming AI entry points. Interfaces are converging on “chat on the left, watch it work on the right,” because “seeing, seeing, seeing” helps users understand the process, tolerate failures, and build trust.
Capital markets have simultaneously entered an ARR acceleration, valuation expansion, M&A, and early-stage Infra cycle. OpenAI revenue rose from $1B in 2023 to $4B in 2024, with its 2025 forecast raised from $10B to $13B; Anthropic’s forecast rose from $2B to $3B. Cursor went from roughly $100M to $500M ARR in less than a year, then raised $900M at a valuation of about $10B. Meta’s investment in Scale AI is described elsewhere in the source as both $14.3B and $14.8B, in each case for 49%; “the late stage is already too expensive for everyone to invest,” pushing early capital toward Agent environments, memory, execution, search, payments, permissions, and Agent-to-Agent and human-to-Agent protocol layers.

Deep dive

1. This “consensus” is only a snapshot of the industry in June 2025

庄明浩 began by putting an expiration date on the conclusion: AI is changing so quickly that when he revisited his roundup from late 2024, it already “felt like something from another era.” What follows is therefore not permanent truth, but a set of provisional judgments the industry appeared to share as of one or two days before the livestream.
He also explicitly left room for future falsification: looking back several years from now, today’s “consensus” may prove wrong. The framework still runs through technology, products, and capital, then ties the three together with trends that may emerge over the next six months.

2. Agent becomes the theme of the year, while intelligence remains the core R&D line

Under OpenAI’s definition, borrowing autonomous driving’s level system, L1 is Chatbot, L2 is a reasoner, and L3 is an Agent; 2025 marks the shift from L2 toward L3. But 庄明浩 repeatedly stresses that “standard equipment does not mean the end”: progression does not mean the next layer replaces the previous one, but that overall capability keeps moving higher.
DeepSeek, Doubao, Kimi, Qwen, and the leading U.S. companies have all made reasoning models standard, while continuing to push both base and reasoning models. Reasoning models are produced through reinforcement learning on top of a Base Model, so continued improvement in base-model capability will also lift final reasoning performance.
The U.S. frontier race is a relay: whichever of OpenAI, Anthropic, Google, xAI, or Meta has just released a new model temporarily takes the lead. Recent baton-passers include Gemini, o4, and o3-pro; no company stays ahead permanently.
At the end of 2022, the industry believed Chinese models lagged the U.S. by roughly 1–1.5 years. By DeepSeek R1, Qwen3, and Doubao 1.6, the benchmark gap had narrowed to roughly the distance from the world’s best models. 庄明浩 therefore sees continued head-to-head competition between the leading Chinese and U.S. models as the next major storyline.

3. Coding becomes the first task category Agents fight over

As Agents move from language to action, they first need a setting where they can work reliably. 庄明浩 considers Coding relatively straightforward because it takes place in an online virtual world with tight logic and clear feedback.
Whether code runs provides a natural reward signal: “right is right and wrong is wrong.” That explains both the rapid progress of reasoning models in math, physics, chemistry, and programming, and why every leading model maker treats AI Coding as a battleground that must be won.
Coding is not free of complexity; it is simply better suited to current reinforcement-learning paradigms. Its results have strict standards and its incentive system is comparatively easy to construct, making it the priority setting for transferring model capability into Agent products.

4. Task duration is replacing traditional scores as the Agent capability yardstick

The industry is beginning to measure Agents by how long a task they can execute: from several minutes to 10 or 20 minutes, then to hours and even days. Duration is becoming a measurable range of capability.
The statistics cited by 庄明浩 show this metric roughly doubling every 7 months from 2019 to 2025, close to doubling twice a year. Restricting the window to 2024–2025, after reasoning models appeared, shortens the interval to about 4 months, with the curve steepening visibly.
The metric has limitations, but once traditional benchmarks approach full marks, measuring Agent improvement becomes a problem in its own right. At the current pace, Agents working continuously for more than several hours “should be visible soon.”

5. Lower inference costs open the door to larger Token consumption

庄明浩 uses an Andy–Bill-style law to describe AI: “Andy gives, Bill takes away.” Hardware, model routing, and smaller models lower unit costs; software then absorbs the new capability through more complex reasoning, longer answers, and more tasks.
The chart may show GPT-4-related capabilities rising roughly 100x while costs fell roughly 100x. But a basic Chatbot might think for only 2 seconds and finish its answer in a few seconds, whereas a reasoning model may think for 30 seconds before producing a longer answer, raising per-call consumption by roughly 5–10x.
If an Agent moves from responding in seconds to working continuously for several hours or a day, it crosses another order of magnitude. Even if Token prices fall by 1–2 orders of magnitude, usage could rise by 3 orders of magnitude. Without reasoning models and Agents, Token costs in the Base Model era might already have become “negligible.”

6. Benchmark saturation turns “how to write the test” into the new bottleneck

Traditional tests such as legal exams, the gaokao, and CPA questions are losing their ability to discriminate. Old leaderboards are repeatedly pushed toward 100, while new ones are quickly caught up with by models. “If a model scores 100 on every question,” further capability gains become difficult to measure.
Evaluation failure feeds back into R&D: teams cannot determine what a model can currently do, making it difficult to design the next round of improvements. How to evaluate, evolve, and further improve Agent capability has therefore become a new problem.
Investment firms are being forced into the arena as well. Technical staff at Sequoia China launched X-Bench, building their own questions to evaluate model capabilities. 庄明浩 summed it up as “even investors have to jump in and build a Benchmark,” showing that the industry is being forced to fill the evaluation gap.

7. Synthetic data and model judges offer new directions for Scaling

After September 2024, “human data is almost exhausted” and the possibility that the Scaling Law had failed became popular views. Yet base-model capabilities continued improving in 2025, shifting the question to whether articles, images, videos, and information generated by models could in turn train models. 庄明浩 stresses that this is not yet a consensus, only an emerging line of experimentation.
He compares it with AlphaGo Zero: AlphaGo first learned from human game records, while Zero did not rely on human records and generated previously nonexistent data through self-play. If synthetic data from large models can produce effective feedback, it would be like “stepping on the left foot with the right, then the right with the left, and flying upward.”
Another difficulty in reinforcement learning is the reward. Math and code have standard answers, but questions such as writing style or which word fits better are difficult to score mechanically. Annotation work at companies such as Scale AI has therefore moved from ordinary labelers to PhDs in law, physics, finance, medicine, and other fields, yet many questions remain difficult even for experts to evaluate.
The new experiment is to have another model act as judge, then use its evaluations to stimulate the original model. “Using one model’s capability to judge whether another model got the answer right” creates another self-lifting loop. 庄明浩 believes that anything producing this kind of self-reinforcing effect may continue moving upward.

8. DeepSeek turns an abstract training path into an observable sample

DeepSeek R1 comes from the V3 base model: pure reinforcement learning first produced R1-Zero, after which post-training corrected its flaws to create the user-facing R1. The chain makes the relationship among the Base Model, reinforcement learning, and reasoning model layers unusually clear.
R1-0528, released on May 28, showed especially sharp gains in math and Coding; other fields improved as well, but less visibly. 庄明浩 attributes the gains to using another model to score outputs and act as the reward model. The new version has nearly caught up with the leading OpenAI and Claude models, despite starting from V3, which was released roughly six months earlier.
For the upcoming V4 and the unreleased R2, he offers conditional reasoning rather than a precise forecast: if V4 is stronger than V3 and reinforcement learning follows the same path, R2 should be stronger still. “Definitely very strong” is his directional call; “how strong, I don’t know” is the explicit uncertainty he preserves.

9. L3 will not replicate the straight-line rise in model capability; Agents may fight a protracted campaign

Pretraining, post-training, and reinforcement learning mainly raise the model itself. Moving from language to behavior requires memory, planning, tools, permissions, environments, and error-correction systems to work together; the leap cannot be completed simply by having the model “step twice on the ladder.”
OpenAI uses autonomous driving to illustrate AI’s levels, and autonomous driving has precisely remained at L3 for a long time because L4 essentially requires no human. 庄明浩’s point is that extremely rapid model progress does not imply that fully autonomous Agents will arrive at the same speed.
In “The Protracted Campaign of Agents,” Flow, a technical staffer at Kimi, estimates that an Agent capable of replacing 80% of white-collar work may take 5–10 years or longer unless an entirely new disruptive technology appears. Others predict 2–3 years, while some believe it will take longer than 5–10 years. 庄明浩 leaves the disagreement intact rather than choosing a timetable for the audience.

10. China’s open-source models have shifted from a backup route to an industrial force

庄明浩 attributes the rapid rise of China’s open-source capabilities mainly to DeepSeek and Alibaba’s Qwen. In his view, Meta Llama, once the strongest U.S. open-source representative, has largely fallen behind, while Meta has repeatedly adjusted its model teams and organization.
Leading projects have in turn prompted Xiaohongshu, Zhipu, MiniMax, Kimi, and others to release open-source models. The old view that large models cost too much and open source could not mobilize enough resources is, at least in light of current results, “not necessarily true.”
The statistical definitions behind the two curves are not the same: the U.S. curve does not distinguish open from closed source, while the China curve includes only open-source models. Interest in open-source projects also maps onto the same industrial storyline: Agent architecture and workflows grew by more than 70%, followed by Coding, Inference, API Management, Post-training, AI Search, and Agent Framework. 庄明浩 cautions that attention is only a third-party metric, but the direction is highly consistent.

11. The browser becomes the Agent’s connection layer and runtime

In less than 2 months, the Chinese internet has produced a large number of Agent products, including Coze Space, Baidu Xinxiang, Lovart, Fellou, Tiangong, and Flowith. At the same time, the old browser form has once again become a target for large-model, AI-search, and Agent companies.
Reports that Chrome could be broken up over antitrust concerns have drawn attention, and companies including OpenAI have reportedly made offers. Google sees Chrome as a core weapon in the AI battle. Changes in Safari search triggered stock-market adjustments, while Perplexity, Genspark, and Arc have also built around AI browsers. Quark’s strategic position has risen in China, and QQ Browser has begun frequent updates to its AI features.
The reason is not that browsers have suddenly become novel, but that Agents need “somewhere to be”: a place to hold memory and planning, invoke tools, execute actions, and maintain a stable, extensible connection to the outside world. The browser may therefore become the Agent’s vessel and runtime.
Product interaction is converging on “a chat box on the left, with the user watching it work on the right.” 庄明浩 emphasizes “seeing” 3 times: when hallucinations cannot disappear, visible execution lets users understand what the Agent did, award failures a “process score,” and remain willing to try.

12. Invite codes, SOTA, and 0.x versions form the new-product SOP

New-product launch videos are highly standardized: the founder sits and introduces the product, then demonstrates it. Invite codes control early traffic and costs while creating viral distribution; some are even resold, spawning platforms such as Ma Dou Lai that upload, request, and automatically distribute invite codes.
A formula summarized by friends is: “Launch the first AI Agent product in a defined field, achieve SOTA on a certain Benchmark, and require an invite code to use it.” Add enough qualifiers and anyone can become “the first.” Another comment was more direct: “The Invite-Code Rebellion, as history will call it.”
Version numbers no longer represent maturity in the traditional sense. Cursor only moved to 1.0 after reaching roughly $500M ARR. 玉伯 previously ran Yuque at Alibaba, then spent a year working at Feishu at ByteDance before starting a company; his product Uman has already been in development for a year and is still at version 0.3.
庄明浩’s explanation is that the AI era allows products to reach users in an extremely early form, stabilizing gradually as models, teams, and users iterate together. Commercial scale and product completeness can therefore become misaligned in ways rarely seen in the previous software industry.

13. AI products are not about adding features, but carving stable outcomes out of “omnipotence”

Traditional car manufacturing assembles the wheels, chassis, and frame piece by piece. Internet agile development iterates from skateboard to scooter, bicycle, and motorcycle before arriving at the car. AI starts from the opposite end: a large model with more functions than anyone can reasonably use, forcing product teams to subtract.
Large models hallucinate, have blurry boundaries, and lack many industry-specific details, so they cannot be shipped directly. The real work of a vertical Agent is to “cut”: define task boundaries, remove unstable capabilities, and ensure repeatable output.
Measurement ultimately comes down to revenue, gross margin, and the data flywheel. Meitu’s ID-photo product is the clearest example: roughly RMB15.9 per use, another RMB4 for an outfit change, and another RMB3 for a larger size. Users in a hurry who only need to upload one front-facing photo are willing to pay for a clear result.

14. ARR acceleration lifts valuations and pushes capital toward M&A and Agent Infra

OpenAI revenue rose from roughly $1B in 2023 to $4B in 2024, while its internal 2025 forecast was raised from $10B at the start of the year to $13B. Anthropic’s forecast for the year was likewise raised from $2B to $3B in roughly 5 months.
Cursor went from roughly $100M to $500M ARR in less than a year, then raised $900M at a valuation of about $10B. AI companies are often priced on ARR multiples; the very largest can typically command multiples in the tens, and at the peak even hundreds. Revenue and the multiple rising together creates another self-lifting loop.
On Meta’s deal with Scale AI, 庄明浩 says earlier that the investment was $14.3B for 49%, then later says $14.8B for 49%; the amount can only be treated as roughly $14.3B–$14.8B, with valuations in both cases above $30B. Meanwhile, 2025 has already seen a large number of AI M&A deals and transactions among private companies. 庄明浩 therefore called it early: “2025 will absolutely be a major M&A year.” He also believes China’s antitrust regime “seems to have turned the page,” citing transaction leads including Tencent Music–Ximalaya, Baidu–YY, and Xunlei–Hupu.
Late-stage projects have become too expensive for many funds to participate in, pushing early opportunities toward missing Agent infrastructure: browser or virtual-machine environments, long-context memory, execution, search, payments, identity and database permissions, and tool-calling protocols. Most Agents currently use Anthropic’s MCP to call tools; in the future, Agent-to-Agent and human-to-Agent connections may also need their own standards, much like HTTP, SMTP, and FTP.
The Gartner curve provides a position check on this wave: Cloud and AI have moved toward rational maturity, while generative images sit further along the curve. AI Coding is at the hottest point, near the “peak of inflated expectations”; Agents are approaching the top, while edge AI is still climbing. 庄明浩’s closing point is that AI is not at one single stage at any given moment: technology, products, and capital are distributed across the peak, trough, and maturity zones.

How AI Learned to Talk and What It Means - Prof. Christopher Summerfield Findell Capital's Brian Finn on Oportun $OPRT