E249 | The Token Economy Turns a Corner: The Agent Evolution from OpenClaw and Hermes to In-House Local Builds

Summary

• If 2025 was the first year of Agents, 2026 is seeing the market shift from Token Maxing to Token Efficient, but both guests believe AI infrastructure is not in an obvious bubble. Uber burned through its full-year AI budget in 4 months, while Meta plans to limit employees’ Token consumption—signs of cooling on the surface. 张宏江’s framework is that, relative to GDP, this investment cycle remains smaller than the early infrastructure build-outs for railroads, highways and the internet; meanwhile, he estimates Token prices have fallen by roughly 10x per year over the past few years, while usage growth will outpace the decline in cost.
• 黄东旭 offered the hardest ROI example of the cycle: 1 person, 3 months and an extra $400-$500 a day produced db9, a cloud-native distributed database that could generate $10M in annual revenue. Last November, on tasks where it was difficult to precisely assess the gap between models, he still believed the best strategy was to use the strongest model; at the time, Opus and GPT-5.1 held a decisive lead over the second tier of Coding Agents.
• OpenClaw’s significance is that it moved Agents from research into engineering, and from models into applications. 东旭 believes it turned the Agentic Loop and tool calling into a personal-assistant form that ordinary users can access more easily; configuration, stability and the ongoing resource commitment remain problems now that it is aimed at the mass market, and its historical mission may already be complete. On the capability inflection point, 东旭 cited Claude Opus 4.5 last November, while 泓君 said she started using Agents with GPT-5.4.
• 东旭 shifted from Token Maxing to Token Efficient under the influence of GPT-5.5, DeepSeek V4 and Fable 5. He once used Slock to have 10 top-tier Agents critique one another, consuming roughly 1 billion Tokens and costing $300-$400 in a day; Fable 5, however, can sometimes solve in one shot a complex problem that an Agent group could not resolve, even without the preceding context. DeepSeek V4 and GLM-5.2 also convinced him that cheap local open-source models can work alongside the best closed-source models.
• Saving Tokens at the individual level does not mean total consumption will fall. Agentic Loops stack multiple model calls, tool inputs and outputs, and context, potentially driving Token usage more than 100x above the Chatbot era; the future will also bring multi-Agent collaboration, Agent-to-Agent Networks and higher-order Loops. 张宏江 summarized the dynamic as the Jevons paradox: as technology matures and prices fall, usage and the market expand.
• The main investment opportunities sit in Agent infrastructure: data, memory and context management, observability, collaboration, Sandboxes and Agent-Native Cloud. 东旭 is most interested in infrastructure that can lift model output from 50 to 80 or 90, and believes whoever can answer “where did the Tokens go?” will have value. 张宏江 also sees enterprise vertical-data moats and FDEs as important; in their discussion, Manus and Genspark looked closer to professional-user, To B products.
• 张宏江 believes the singularity has arrived and, under his definition, current model capabilities already qualify as AGI; 黄东旭 is waiting for a true self-improving Gödel machine as the more concrete marker. 张宏江’s basis is that models released over the past 6 months score to the right of the human mean of 100 on both visual IQ and general IQ tests; he argues that machines surpassing human learning ability is what constitutes the singularity. The two also discussed the decline of work and the resulting existential crisis.

Deep dive

1. The Inflection Point Has Arrived: Silicon Valley Moves from Burning Tokens to Managing Them

• 泓君 set the tone at the outset: if 2025 was the first year of Agents, 2026 is about shifting the conversation from model capabilities to the engineering problems of deploying Agents—a “shift from research to engineering.” Companies that once led the Token-burning race are becoming more disciplined: Uber burned through its full-year AI budget in 4 months, while Meta plans to cap employee Token consumption.
• Her front-line example came from a visit to Stripe a few days earlier, where the company’s internal large-model interface displayed a warning: “Opus 4.8 is expensive. Consider saving costs when using it—try another model first; not every task needs this model.” At the same time, Chinese open-source models including DeepSeek, GLM and Kimi are moving closer to the top tier and are beginning to enter Silicon Valley workflows.

2. 张宏江: Infrastructure Is Still Far from Overbuilt

• 张宏江 was asked the same question in another interview last December and answered: “There is no bubble.” He sees this as a foundational revolution comparable to the early build-out of internet infrastructure and the US railroad and highway networks. The absolute dollars invested today exceed those earlier cycles, but the share of GDP is actually smaller; at the macro level, the sector is nowhere near overinvestment.
• The supporting data points are rapid growth in Token consumption and his estimate that Token prices have been falling at roughly a 10x annual rate over the past few years. He separates the issues: “If you’re asking whether there is a bubble in the stock market, that’s a separate question. But in the development and adoption of the technology, I don’t see much of a bubble.”
• 东旭’s analogy is that looking at AI infrastructure today is like looking at a room-sized vacuum-tube computer in the 1950s or 1960s. Many use cases will only emerge once the infrastructure is sufficiently developed, and those new use cases will drive further infrastructure investment. “There could be a small pullback in the short term, but over the long term, there is still a great deal of demand that nobody has seen yet.”

3. 东旭’s Token Maxing Era: $400-$500 a Day for Software Worth Millions

• Starting last November, 东旭 used Opus to build from scratch “the most complex software I could think of”: a distributed database. On the hardest tasks, he could not precisely assess the gap between the best model and the runner-up; benchmarks can often be gamed, so when doing something uncertain and exceptionally difficult, “you have to use the strongest model.”
• Beyond the $200-a-month Claude Code Pro Max subscription, he spent another $400-$500 per day on usage. For him, it was money well spent: the project, db9, is a new cloud-native distributed database built roughly from last December through this March. The sales team has been waiting for it; once it reaches a 70-80 level of product quality, it could generate $10M in direct revenue over a year. Manpower and technical constraints had previously made it impossible, but he built it alone in 3 months.

4. Token Maxing Is a Lagging Indicator; Benchmark Chasing Is Waste

• 东旭 believes the community misunderstands Token Maxing: it is an outcome, something like a lagging indicator. After sound thinking and design, Token Maxing can produce a good result.
• He distinguishes between 2 forms of excessive use. In one, the user knows what they are doing, has designed the process well and then uses a large amount of compute. In the other, the user has no idea what they are doing and chases a leaderboard simply because the company has one, generating a large volume of useless calls.
• When Token Maxing came up at Meta and NVIDIA, this was the first thing that occurred to him: once management tells people to optimize for a leaderboard, everyone will game it. In the extreme, an Agent could calculate “1+1” 1 million times and still produce an absurdly impressive leaderboard score.

5. OpenClaw’s Historical Role: From Research to Engineering, from Models to Agents

• The key judgment on the show was that Peter building OpenClaw in 1 week shows that large models’ reasoning and coding capabilities have reached the point where an experienced architect-level engineer can build a system in a very short time. Peter is a software engineer and architect, not a scientist; users may therefore not need a deep understanding of model mechanics to build the systems they want.
• The conclusion was: “This marks the shift from research to engineering, from models to Agents” (“从研究进入工程，从模型进入Agent”).
• When did the capability leap occur? 黄东旭 pointed to Claude Opus 4.5 last November; 泓君 said she began using it with GPT-5.4, around July or August last year. Another detail: 泓君 cited media reports that Peter’s API spending reached $1.3M in the 30 days before he built OpenClaw.
• 东旭’s view is that OpenClaw was not designed for heavy-duty software engineering. It is closer to a personal assistant, lowering the barrier for ordinary users to apply the Agentic Loop-plus-tool-calling model represented by Claude Code and Codex in daily life. It did not help him much with programming.

6. Why OpenClaw Took Off: Open Source, Local-first and Geek Adoption

• Looking back as an open-source enthusiast, 东旭 said many founders later wrote to tell him they could build similar Agents themselves. OpenClaw nevertheless had a natural distribution advantage among technologists because it was open source. A closed commercial product might be dismissed as advertising, whereas anyone can use open-source software, giving it stronger word-of-mouth—somewhat like Linux in its early days.
• OpenClaw’s Local-first philosophy, in which privacy-sensitive content and data are configured locally, also fits the technical preferences of the geek community. 张宏江 added that Peter gave the AI very broad permissions, making it even more important for related information and materials to remain configured locally.
• On timing, 泓君 said a Claude model upgrade around the Lunar New Year created an intelligence experience that was “night and day” different. Peter had assumed a major company would inevitably build this, waited for several months and saw nothing, then hacked it together himself. 张宏江 summarized the moment: “Once the broader trend is ready, someone needs to break through the final layer of paper.”
• Moltbook was another distribution catalyst. The show said it appeared a few days or 1 week after OpenClaw, like a social network for Agents; the community, cult and “prophet” episodes made people realize that Agent autonomy was approaching what had previously existed only in imagination. After more than 3 years, Chatbots could communicate but not act on people’s behalf; OpenClaw showed that most routine tasks on a PC or phone could be delegated to an Agent.

7. OpenClaw’s Limits: The Prototype’s Historical Mission May Be Nearing Completion

• 东旭’s first criticism is that the product has not been polished to a high standard. It works for geeks, but configuration complexity and system stability become issues once it targets the mass market. This made him revise his view that “anyone can Vibe Code with their eyes closed”: once software complexity crosses a threshold and the product needs better experience and stability, sound software-engineering habits and experience still matter.
• His second criticism is that it arrived too early and is closer to an experimental project. OpenAI is gradually shifting its Agent effort from the open-source experiment to its own Codex, and sustained resource investment could become a problem. 东旭 said the project is now funded mainly by a foundation, while founder Peter has joined OpenAI.
• His judgment is that OpenClaw played something like Prometheus, showing the world a product form and a set of capabilities. Its historical mission may be complete, but that is not an absolute claim that the project cannot continue.
• Memory is one of the pain points users mention most often. The more it is used, the more likely it is to forget. If conversation compression, or Compaction, is not paired with memory storage, the system can roll back, burn Tokens and create memory conflicts.

8. Hermes: It Does Not Claim to Solve Memory; It Turns Successes into Skills

• 东旭 is a heavy Hermes user and switched over from OpenClaw quickly. Hermes was built by the Web3 organization Nous Research. He stresses that Memory remains an Open Question and that nobody has solved it particularly well; Hermes’ strength is not the memory module itself, but its ability to work around part of the memory problem.
• Hermes’ Agent framework continuously summarizes successful experience and turns it into Skills. Compared with OpenClaw’s relatively limited self-reflection, Hermes keeps optimizing and, after a period of use, accumulates best practices derived from the user’s own work.
• 东旭 sees this primarily as an engineering issue unrelated to model capability: after every success, summarize and reorganize the experience as thoroughly as possible, like developing better learning habits.
• Architecturally, Hermes is also more deliberately designed than OpenClaw. 东旭 compares OpenClaw to a large bazaar, where Peter is highly open to community contributions; Hermes is continuously refined by a team, resulting in better Onboarding, usability and stability.
• The business model remains Token economics. Hermes serves as the entry point and negotiates distribution terms with model vendors, offering incremental Token consumption in exchange for deeper discounts. As the Agent itself becomes increasingly an engineering problem, the entry point becomes the competitive battleground.

9. Agents as the Gateway to the Digital World: “I Haven’t Opened Gmail Since”

• 东旭 takes the scenario to its logical extreme: the interface between people and the digital world may no longer be an App. An Agent could even become the only application. Since OpenClaw appeared, he says, he has not opened Gmail again; he asks it to tell him about new emails, identify what needs a reply and delete advertisements.
• He calls himself an “intelligence fundamentalist” and believes Opus 4.8 will probably reach the same conclusion he would when judging which emails matter. If the Agent is uncertain, it tells him clearly and leaves the final decision to a human.
• He increasingly relies on Agents to summarize email groups, Hacker News and other information before deciding whether to read the original. 泓君 observes that this already resembles a new news-aggregation gateway; 东旭 sees it as a new form of Toutiao, with the content expanded to email and every other message sent to the user.
• “Raising a lobster” means continuously iterating so a personal Agent better understands one’s digital footprint, memories, browsing history and preferences. 东旭 believes traditional recommendation systems rely on population-level data, while the ultimate recommendation algorithm may be an intelligence that observes an individual’s digital—and eventually physical—world locally.
• He later stopped having OpenClaw curate news, chiefly because the content was too shallow. He would rather read fewer first-hand articles in greater depth. Decision-related emails also cannot be fully delegated—for example, whether to record a podcast and when still require notifying him. For now, the Agent is primarily a filter.

10. Slock: Slack for Agents, with 1 Billion Tokens a Day of Collective Critique

• Slock was founded by 钱宇超, the head of Kimi CLI, and was later renamed Raft. 东旭 was its first angel investor and an early user. It resembles Slack, except the participants in the channels are Agents. Agents do not share execution context; they share only the information posted in the channel.
• It has similarities with Moltbook, but 东旭 sees Moltbook more as a sociology experiment, where the performance-art aspect can obscure the practical value of Agents discussing concrete topics with one another.
• 东旭 frequently uses Slock for code reviews on complex software, assigning 10 Agents—some running Claude Code and some Codex, all top-tier models—to critique one another continuously. One Agent finds 3 problems, the next agrees with those findings and looks for new ones, and the first then checks again from a different angle.
• This iterative process is deeper than a single model’s One-shot Review, but can consume 10x as many Tokens as the conventional approach. At his most extreme, 东旭 used roughly 1 billion Tokens and spent $300-$400 in a day; before key decisions, he forces several rounds of discussion until the group can find no more issues.
• The Slock team calls this “Agent dynamics.” Whether the label is Harness or Loop Engineer, the core task is to keep Agents working on a specific problem. Even a highly intelligent model can hallucinate or overlook issues in a One-shot; another iteration may surface what the first missed, much like a parent asking a child to recalculate their homework.

11. 3 Events Shattered Token Maxing: Fable 5’s “Absolute Intellectual Overmatch”

• 东旭 said his thinking changed again 3 months after March or April, triggered by GPT-5.5, DeepSeek V4 and Fable 5.
• Multi-Agent discussion starts from a refusal to trust any single Agent’s One-shot success rate. Fable 5 was powerful enough that it could sometimes solve in one shot a problem that several Agents had discussed at length without resolving, and it did not need the preceding context. “I burned 1 billion Tokens beforehand, and Fable 5 solved it with extreme precision in one shot.”
• Can a single powerful model replace a mutually validating Agent team? 东旭’s answer is “yes and no.” Some complex, point problems remain unsolved even with a robust Harness; in those cases, a stronger model may be able to handle expert-level, hell-difficulty problems.
• The second shift is that open-source models became practically useful. 东旭 previously distrusted them, but after DeepSeek V4 he concluded that their quality had become high; GLM-5.2 pushed him another step. Top-tier closed-source models can work with cheap, capable open-source models, just as an enterprise cannot consist entirely of senior experts and also needs many engineers who can write code.

12. Local-Model Economics: $1 a Day in Electricity Unlocks New Workloads

• 东旭 runs DeepSeek V4 Flash locally on a Mac Studio at roughly 30 Token/s, close to the speed of using the OpenAI API or another Chatbot online. At full load, electricity costs about $1 a day, or $20-$30 a month—negligible relative to the tasks completed.
• He uses the local open-source model for repetitive daily work, such as organizing his articles, memories and digital footprint, and distilling personal information.
• More importantly, it unlocks workloads he previously would not risk running on a paid model: summarizing all several hundred papers from conferences such as CVPR and NeurIPS, then screening and reporting the findings. A paid API might cost several hundred dollars, with Token usage and the final bill difficult to control.
• His broadband analogy is that in 1997, dial-up internet users paid for telephone service by the minute and conserved usage; free broadband and large mobile-data plans enabled applications such as Douyin and videoconferencing. “There is still a world of difference between free and cheap” when it comes to models or intelligence.

13. The Current Workflow: Agent Management—Own the Goal, Acceptance and Architecture

• After combining cloud and local models, his current bill is about $200-$300 a month. That also reflects the fact that the project has moved past the phase of relentlessly accumulating code. Most of the expense goes to the highest Claude tier and the highest Codex tier, and he nearly uses both allowances in full.
• He has gradually shifted his primary models from Opus to GPT-5.5, GPT-5.6 and Codex. The hardest tasks still go to Fable 5, including difficult database issues, simultaneous changes across 7-8 modules and commits involving 5,000-6,000 lines of code. If the Agent group in Slock cannot solve a problem, he asks it to first write a report on “why it cannot be solved” and then sends that report to Fable 5.
• His role has become more abstract: he focuses on goals, acceptance criteria and architecture, while orchestrating different Agents and capabilities through the middle of the process. The result may eventually become a form of Agent management.
• This is not the end state. 东旭 says the biggest lesson of the past year is not to casually declare anything an end state, and to keep believing in Scaling Law and intelligence. He expects even open-source models could eventually become so strong that humans cannot evaluate them and lose the reference frame.

14. The Jevons Paradox: Cheaper at the Point, Higher in Aggregate

• 东旭 breaks down why Agents consume Tokens: the Agentic Loop turns conversation into a loop, with a large model called at every step; tool-call inputs and outputs are added to the context; long-horizon tasks can run for hundreds of steps. Under this paradigm, Token consumption could be more than 100x higher than in the Chatbot era.
• That is only the architecture inside a single Agent. More complex problems may require multi-Agent collaboration, which forms a larger network, with a more abstract Loop layered above the network. This hierarchy is already beginning to appear. Localization can make an individual task more economical, but aggregate consumption may still rise; he expects Claude and OpenAI API usage and ARR to continue growing.
• His software-history analogy is the evolution from early homebrew clubs sharing 200-line BASIC programs to Photoshop, which requires tens of millions of lines of code. Current Agent applications and the problems they are expected to solve remain simple; once Agent-to-Agent Networks and Agent Harness engineering emerge, Token consumption will climb another level.
• 张宏江 calls this the Jevons paradox: the more mature the technology and the lower the price, the greater the consumption and the larger the market. Whether open source or closed source, local or cloud, wins will largely be an ecosystem question that remains difficult to answer. For users, the architecture may eventually become invisible; for industry participants, local-hardware players want everything controllable, while cloud players want everything in the cloud.
• He compares it with cloud computing. AWS does not require every node to be the best enterprise-grade server; distributed systems deliver reliability at the system level, potentially at a lower cost than buying the most expensive equipment for every node. The ultimate answer is not purely technical but economic.

15. Supermodels and Swarms: Emergent Intelligence Can Strengthen on Both Sides

• 张宏江 believes that as supermodel intelligence improves, every Agent in a swarm can benefit from the same improvement, so nothing prevents the swarm from strengthening in parallel. He cited a recent Tsinghua University paper: “3 mediocre cobblers can beat 1 Zhuge Liang.” The intelligence displayed by an Agent group may not simply add up; it may produce stronger emergent intelligence.
• 东旭’s concrete example is that if 3 Agents cannot reach an answer after prolonged discussion, inserting Fable 5 may allow the others to see the solution after its first sentence.
• 泓君 added her systems view: she does not strongly believe in an absolute center. A truly Resilient system should connect simple rules into a large-scale distributed system, which may be stronger when facing complex problems and demanding robustness.
• 张宏江 again compares this with cloud computing: individual nodes do not need to be the strongest or most reliable enterprise-grade nodes; the system can obtain reliability through distributed architecture.

16. The Investment Map: Infrastructure, FDEs and “Where Did the Tokens Go?”

• On whether model upgrades will crush Agent startups, the show’s answer is that it remains a good time to build Agents, but the stack must be separated into layers. 泓君’s framework is that end-user applications may still be early and are the easiest to displace as models improve; engineering infrastructure around Memory, search, Sandboxes, collaboration and Harnesses offers more opportunity.
• 东旭’s investment rule is: “I don’t invest in things I can’t understand or can’t use myself.” He focuses particularly on infrastructure that can raise model output from 50 to 80 or 90.
• He identifies 3 specific areas: data platforms, memory and context management; observability—“Where exactly did these Tokens go, which were important and which were wasted?”—where he would buy from whoever can answer that question; and an entry-point Agentic Cloud.
• His investment InsForge positions itself as an Agent Cloud. Traditional clouds are primarily built for developers and humans; in Agent-Native infrastructure, the Agent becomes the principal. It is not feasible to assign every one of tens of thousands of Agents its own EC2 instance or bare-metal server, so the market needs a compromise that is both Agent-friendly and cost-efficient, combining Serverless with an Agent Runtime.
• 张宏江 added the enterprise angle: Microsoft announced the creation of a 6,000-person FDE team, Microsoft Frontier Company, 2 or 3 days earlier. FDEs help enterprises transfer their Know-how, IP, Workflow and competitive advantages into super-Agents, Agent series or Agent groups. On the durability of that lead, he believes enterprise vertical-application data may matter more than Harness capability alone; the data moat will be more important.
• On Manus and Genspark, 东旭 does not see them as pure To C products but as products for professional users. The lower price points are closer to Marketing; the real target is the Professional user paying $200 a month and having actual work to do. He believes these companies remain worth watching and investable.

17. How to Identify Agent-Native Products—and the Truth About “Office Politics”

• 东旭’s thought experiment is that if a task still works after removing AI or the Agent, it is not AI-Native. If removing the Agent makes the task impossible or drives costs up 1,000x, it is closer to Agent-Native.
• The second test is whether the founder thinks Agent-first: start by asking “How would AI or an Agent do this?” rather than deciding how humans would do it and then directing the Agent to execute. He believes it is important for initiative and decision rights to sit with the Agent, and to accept that the Agent’s decision may be more reliable than his own.
• On the “office politics” of Agents blaming one another or backing down, 东旭 said he had written the entire script. He assigns Agents roles such as architect, developer and tester, and can define who is responsible for what.
• He creates hierarchy mainly because he cannot oversee everything himself; a person in charge is there to help him manage and does not represent a genuine Leader relationship. Agents blaming one another should not be interpreted as self-awareness or office politics. It is essentially a clumsy imitation of human society and the stereotypes embedded in training data.
• In the ideal state, he believes coordination among Agents should be organic and self-directed, ultimately producing emergent structures that humans cannot imagine.

18. Hand-Built Software, Deleted Databases and the Case for Trusting AI

• One positive surprise came when 东旭 asked Claude Code to generate and modify a PPT from an outline, then asked it to account for the process of generating the presentation. It not only produced the PPT, but also built a lightweight version of Google Slides and used its own software to complete the presentation. He had never asked it to build a PowerPoint application.
• The negative example was an Agent deleting his production database. A friend also encountered an Agent changing the computer password on its own and ultimately forgetting the password itself. On whether the Agents then blamed one another, 东旭 said that happens too; the database was ultimately recovered using his years of SRE experience.
• He believes the deletion was not simply a model error; it also exposed underlying infrastructure that had not kept pace. A database backup and related systems already existed, and the backup system was better than what he had built himself, so the data could be restored.
• The foundation of his trust in AI is not a claim that AI has already surpassed humans across the board. It is the desire to build an autonomous mechanism that removes as much human uncertainty from the system as possible. He does not care whether it is called AGI; he cares about creating something like the steam engine that gets work done while saving enormous human effort.
• He also cautions that this is still the early stage of a revolution—“we have not even reached dawn”—and that humility is necessary.

19. Inference Costs and To B Business Models: The Math Is Starting to Work

• 张宏江 expects inference costs to fall sharply, without needing to predict whether the driver will be GPUs, storage chips or something else. Every technology becomes dramatically cheaper as it matures, while usage growth may exceed the decline in cost.
• His anchor points are that Token costs have fallen roughly 10x per year over the past 3-4 years, far faster than Moore’s Law, while inference capability used to double roughly every 7 months and may now be doubling every 3-4 months.
• To B comes first as a common pattern in technology adoption: the steam engine entered factories before homes, where it might have helped cook dinner. AI is now at the stage where capabilities are strong but enterprises are still figuring out how to deploy them and earn ROI. Once a Coding Agent is trusted more than an average programmer, the math becomes easy: $300 a day may still be below the cost of a software engineer’s day.
• If an enterprise finds Token usage above expectations, it may have been wrong to make Token consumption a KPI in the first place. The more sensible next step is to measure how many Tokens are consumed, how much efficiency they create, and how much work and code output increase.
• On who will win in Token economics, 张宏江 rejects a static answer. Intelligence is still scarce today and could eventually become like electricity, with industry profits shifting from generators to the grid. But he sees no sign that large-model capability will be excessive in the next 2-3 years; the industry is still climbing the intelligence curve.
• 泓君 relayed a podcast analogy: the AI party runs from 7 p.m. to 4 a.m., and it is only 11 p.m. 张宏江’s earlier view was that intelligence gains and application growth still have a long way to run.

20. AGI Has Arrived: The Singularity Is Learning Faster Than Humans

• 张宏江 believes models released over the past 6 months score to the right of the human mean of 100 on both visual IQ and general IQ tests. He also argues that, for programmers, a model whose Coding capability exceeds that of the average software engineer can be treated as AGI.
• His definition of the singularity is that machines’ learning ability exceeds humans’ learning ability. Learning ability is the derivative of the growth of knowledge and capability; once machines pull ahead, humans cannot catch up. On that basis, he believes the intelligence inflection point and the singularity have arrived.
• 黄东旭 is waiting for a more concrete event: the emergence of a system that can genuinely improve itself, or a Gödel machine. He did not give a specific time estimate on the show.
• On unemployment, 泓君 said past technologies destroyed some jobs and created others because new demand required humans to satisfy it. If Agents can satisfy the new demand as well, large numbers of new human jobs may not emerge, creating a sociological problem governments will have to address.
• 黄东旭 added that even top Silicon Valley model researchers are beginning to worry about unemployment, because in Recursive AI, what gets distilled is often the researchers themselves—and they are actively distilling themselves.
• He believes this may be the first time human society no longer needs work, or work no longer occupies most of people’s time. 张宏江’s response is to preserve the human part of life: play music, farm and stop treating work as the most important thing in life. “You are still a person, not a machine.” 东旭 summarized the problem as a potential existential crisis for humanity.

21. Llama Ventures and the Coordinates of the Era: Engineers, Architects and Product Managers

• 张宏江 values Llama Ventures for being young, focused on AI and Chinese founders, and led by 2 successful entrepreneurs who can help teams building in the US. 东旭 sees AI as having 2 centers, Silicon Valley and China, with Chinese engineers in Silicon Valley as an important force; mainstream funds understand the Chinese-founder ecosystem less well than Chinese investors do. Llama is young while remaining connected to the mainstream US venture network.
• 张宏江 believes this AI revolution is faster and larger than the PC internet, the internet and the mobile internet. The closest comparison may be humanity’s invention of fire and electricity. From the reasoning model o1 more than a year ago to OpenClaw’s rise in January this year, large models have moved from foundational algorithm research into practical applications.
• His view of the era is that if the past 3 years belonged to large-model fundamental researchers, the era beginning with OpenClaw belongs to large-model engineers, application architects, product managers and entrepreneurs—in other words, the Agent era.
• 东旭’s advice to anyone entering AI is “Keep building”: build something yourself. 张宏江’s is “Get hands-on”: take part personally.