139. A Conversation with 苏煜: Agent History and the OpenClaw Moment
Summary
苏煜 distills the 2026 Agent thesis into a single chain: continual learning builds a broad world model, ultimately specializing general-purpose Agents into reliable, fast, low-cost expert agents. On tasks that are not highly specialized, current systems get roughly 60-70% right, but fail inexplicably on the remaining 30-40%; neither RL post-training nor the non-parametric learning represented by writing SOUL.md and SKILL.md has yet replicated how humans go from intern to expert. “All the way continual learning, all the way world modeling.”
OpenClaw’s breakthrough lies less in underlying technical innovation than in putting already-ready Agent capabilities into the hands of the mass market in a personal, always-on, near-full-permission form. Like ChatGPT, it changes interaction and perception: an instant-messaging entry point, an independent runtime environment, and open permissions let users see for the first time how much Agents can already do. 苏煜 believes that, looking back in 2 years, its impact could approach the scale of the ChatGPT moment, marking the start of a highly automated personal-agent paradigm.
Coding is dissolving the boundaries between Web Agents, Computer Use Agents, and Coding Agents, with the endpoint a universal digital agent. Programming languages are languages too, and code is the fundamental fabric of the digital world; GUI, CLI, API, browser, desktop, and mobile are all means to an end that can be rendered into equivalent forms through code. GUI will not disappear: humans are visual animals, and existing GUIs encode decades of accumulated business logic, constraints, and long-tail workflows.
Once general intelligence becomes cheap, the real differentiation in the Agent market will come from specialization across millions of “small worlds.” Model companies are better positioned to compete for a unified gateway to the digital world, but not necessarily to go deep into every profession, enterprise, software product, and workflow; the total entropy of these micro worlds is nearly infinite, leaving broad room for non-model companies and ordinary people to build expert agents. “This world is not one world; it is made up of perhaps millions of small worlds.”
SaaS will not necessarily be rewritten wholesale by Agents, but its valuation logic and margins are already under pressure as competition shifts from selling tools to delivering labor and outcomes. Traditional software still has substantial value, and many existing solutions are already good enough; the unresolved question is whether AI-native companies, frontier model companies, or incumbent SaaS vendors can first turn software platforms into labor markets, and what equilibrium ultimately emerges.
Neo Cognition’s $40M seed bet on specialized intelligence also reflects the powerful Matthew effect in AI financing. The company was founded in July-August 2025 and raised the round in March 2026, roughly six months later; 苏煜 estimates that OpenAI and Anthropic may account for 30-50% of total market funding, while mid-sized funds are under pressure. A startup’s defense is not that big tech can never replicate it, but choosing a problem with nearly unlimited upside, extreme technical uncertainty, and enough room for multiple approaches.
In the foreseeable future, 苏煜 sees the more realistic risk not as AI autonomously eliminating humanity, but as job displacement outpacing the creation of new work and redistribution mechanisms. He has not seen a viable path to giving AI innate goals, survival pressure, and native intention, but Agents replacing knowledge workers at scale while gains accrue to a handful of leading companies or pools of capital would create a direct social shock. Researchers therefore have a responsibility to ensure frontier Agent capabilities are not limited to giants, so individuals can turn unique insights into Agents and “make money out of it.”
Deep dive
1. Agents Were Never a New Concept; What Changed Was the Technology Needed to Realize Them
苏煜’s minimum definition has three conditions: an Agent is an entity with a boundary, situated in an external environment and engaged in goal-directed activities rather than wandering randomly. Under this definition, animals are Agents, and humans are currently the most advanced human agents.
From AI’s formative era in the 1940s-1960s, researchers wanted to build artificial agents that could imitate humans; at the time, however, pursuing a complete Agent was unrealistic and, in some ways, counterproductive, so subfields such as vision, language, and logical reasoning gradually split apart.
Today’s shift is more like “reunification after a long separation”: capabilities once split apart by inadequate technology are converging again. That is why 苏煜 repeatedly stresses that Agent is not a suddenly invented new direction, but the root problem running through all of AI.
2. Memory and Autonomy Form a Unified Coordinate System for Evaluating Agents
苏煜’s definition of memory goes far beyond conversation history: it covers knowledge representation and the acquisition, updating, and forgetting of knowledge; the content includes semantic knowledge, episodic memory, and procedural memory such as riding a bicycle.
Autonomy is a complete loop: the Agent perceives its environment, reasons with existing knowledge, makes a decision, and finally takes action to change the external world. It is not a synonym for “can answer questions,” but a continuous capability from perception to action.
The two are “two sides of one whole”: memory is the foundation of autonomy, and an Agent’s ability to act reliably depends on the world knowledge it has accumulated.
This coordinate system applies not only to software Agents but also to robots and humans. Later discussions of self-learning, world models, specialization, and reliability can all be reduced to these two dimensions.
3. Logical Agents Never Delivered on the Expert-System Promise Because Knowledge Acquisition Was the Bottleneck
The dominant paradigm from the 1950s through the 1990s was the logical Agent: engineers interviewed domain experts, encoded their knowledge in first-order predicate logic and similar formalisms, then had an inference engine reason through new problems. It was the earliest Agent form that genuinely worked.
The limitations hit both memory and autonomy: most of the world’s knowledge cannot be expressed in a finite set of logical statements, while higher-order, fuzzy, and probabilistic logic filled only a small part of the gap; the system could do little more than “receive a question → perform logical inference → output an answer.”
The fatal flaw was the knowledge-acquisition bottleneck. Manually interviewing experts, translating their knowledge into formal rules, and maintaining it was “extremely painful and extremely inefficient,” and the results remained limited; expert systems failed to deliver on their promise, directly helping set off the AI winter of the 1980s and 1990s.
The first edition of Stuart Russell and Peter Norvig’s Artificial Intelligence: A Modern Approach was published around 1995. Russell stressed to 苏煜 that although it looks like an AI textbook, it is “essentially a book about agents,” beginning in Chapter 1 with the intelligent agent.
4. Deep RL Made Agents Go Deep, but They Remained Trapped in Single Games and Fixed Compute
After 2000, and especially after 2010, neural Agents rose through deep reinforcement learning; AlphaGo, Atari, Dota, and StarCraft became representative environments, showing how neural networks could train deeply on one task inside a constrained world.
These Agents typically consisted of networks with tens of millions of parameters, at most roughly 100M; their inputs were game screens and their outputs a finite action set, while memory only needed to hold that game’s rules and perception/action knowledge. By today’s standards, they were small and highly specialized.
Their reasoning was hidden inside a single forward pass, so the available compute was basically fixed no matter how complex the position; humans, by contrast, adjust the amount of thought to the difficulty. Their “generality” mostly meant that one architecture could play multiple games, not that it could be dropped into an unfamiliar environment and learn quickly.
Asked by the host why researchers favored games, 苏煜 said he had not personally built this kind of Agent and could only speculate: beyond Demis’s personal preference, highly repeatable virtual environments offered data abundance, easing the sample inefficiency of potentially needing millions of games even for a simple title.
5. Semantic Parsing Expanded the Action Boundary
In parallel with deep RL’s drive to go deep, semantic parsing in NLP tried to go broad: converting what people say into machine-readable meaning representations that could connect to databases, knowledge graphs, websites, or other digital environments.
It addressed communication, language understanding, and action at once. Once language could be mapped to a formal expression in the environment, people could use natural-language commands to make machines complete more tasks, expanding the Agent’s action space.
苏煜’s PhD research came out of this tradition. He mentioned Percy Liang, Luke Zettlemoyer, 于涛, and later LLM/Agent researchers including himself and 孙欢; many came from semantic parsing because the two problem families are deeply connected.
6. Language as a Scaffold Gave Agents Scalable Reasoning for the First Time
苏煜 calls the generation after ChatGPT language Agents because language is not merely an input/output channel but the scaffold running through perception, reasoning, and action. Natural language makes human-machine interaction more flexible, while formal and machine languages can directly manipulate the digital world.
Chain of thought changed the neural Agent’s fixed compute budget: simple tasks generate fewer tokens and complex tasks more; because each token corresponds to a forward pass, models gained adaptive computing and adaptive reasoning instead of running the same single feed-forward pass for every problem.
From a memory perspective, LLM training is also a process of “shaping memory”: it compresses massive language surface forms into internal representations of the world. 苏煜 says “stochastic parrot” is not enough to describe this process, because those representations directly drive an Agent’s perception and action.
Language is therefore this generation’s defining feature. Even if the underlying system eventually becomes a broader world model rather than an LLM alone, the ability to understand the world, reason, and interact through language “cannot disappear.”
7. Language Was an Exponential Accelerator in Human and AI Evolution
苏煜 draws an analogy to biological evolution: complex life and mammals took hundreds of millions to more than a billion years to emerge; the Homo genus dates back more than 2M years; systematic symbolic expression emerged roughly 100K years ago; and mature written language is only 5-6K years old.
Human civilization exploded over the most recent 5-6K years. Language lets knowledge accumulate across individuals, space, the present moment, and generations; it is not an auxiliary communication tool but infrastructure for expanding intelligence.
AI is now also “living in compressed timelines”: progress that once took 10 years may be compressed into 1 year or even 1 month. 苏煜’s 2023 blog, Language Agents: A Critical Evolutionary Step for Artificial Intelligence, treated language agents as a landmark in AI evolution.
8. CoT and ReAct Built the Minimal Loop of Modern Agents in 2022
苏煜 regards Chain of Thought in early 2022 as one starting point: it initially served reasoning environments such as math problems, but for the first time let models increase computation dynamically with task complexity.
Around October 2022, ReAct extended CoT into the external environment: the Agent perceives the current state, performs CoT reasoning, chooses an action, and lets that action change the environment, forming a loop. The idea sounds simple, but having the right insight at the right time is not easy.
Around the same time, 苏煜’s team produced LLM-Planner, one of the earlier efforts to use LLMs for robot planning and embodied-agent planning; Google’s SayCan was, in his view, another widely recognized early example. The architecture of the modern language Agent was beginning to take shape.
9. Toolformer and AutoGPT Took Tool Use and Autonomous Execution Mainstream for the First Time
In February 2023, Meta’s Toolformer demonstrated LLM tool use; 苏煜 calls it the first representative work of its kind. It was circulated internally at Microsoft by CEO Satya Nadella because a model’s ability to call existing tools would have entirely different implications for enterprise businesses.
AutoGPT in March 2023 wrapped an LLM in an Agent shell that appeared able to “do anything.” It quickly passed 100K GitHub stars and later reached roughly 180K, at one point the fastest-growing repository by stars in GitHub history; 苏煜 says bluntly that the number of things it could actually accomplish was “very small.”
Another project billed as a fully autonomous AI engineer later evolved into Lovable, now one of the representative companies in Web coding.
10. Multimodality Moved Web Agents from Text Structure to Real Visual Interfaces
苏煜’s team began Mind2Web around October 2022, before or around ChatGPT’s release, and published it in early 2023; by his recollection it was among the earliest LLM-based Web Agents or Computer Use Agents.
As 苏煜 tells it, GPT-4o arrived in late 2023 as “the first multimodal model with genuinely practical capability.” While the team was building the MMMU multimodal benchmark, it was also building the visual Web Agent SeeAct; GPT-4o had no API at the time, so they wrapped the webpage interface as an API to run the experiment.
WebArena appeared around July 2023. It cloned several complete websites to create a reproducible space closer to an RL environment, avoiding the constant change of live websites, legal risk, and poor experimental reproducibility.
After that, most Computer Use Agents moved toward vision-based or hybrid approaches. Text representations remained useful, but Agents increasingly relied on visual perception and direct interaction with the environment.
11. Seeing Screens and Clicking Pixels Like Humans Became Mainstream Embodiment for Computer Use
In 2024, Agents moved beyond the Web into desktop and mobile; OSWorld from 于涛’s team became a representative desktop environment, while SWE-bench helped coding Agents develop as a distinct track. The Agent’s operating range was now clearly broader than the browser.
In UGround, released by 苏煜’s team in the second half of 2024, they proposed that “agents should use computers like humans do.” Human embodiment means looking at the current screen, reasoning, and then clicking specific pixels or entering text; it differs from reading HTML or another text-based representation.
Claude Computer Use and OpenAI Operator subsequently adopted visual perception plus pixel-level actions; Claude Code began supporting desktop computer use, and later forms such as Claude in Chrome continued the approach. Computer Use was thus gradually becoming more than GUI automation and approaching a general digital agent.
Operator, ChatGPT Agent, and Claude Code appeared in 2025; around the arrival of Opus 4.5, 苏煜 felt coding practice in Silicon Valley had been turned upside down within “1-2 months,” with many people essentially no longer writing code themselves. OpenClaw appeared around November 2025 and truly broke out around February 2026.
12. Coding Is Dissolving the Product Boundaries Between Web, Desktop, GUI, and APIs
Browser, desktop, and mobile; GUI, CLI, and API; and even coding versus tool use were useful early research categories, but 苏煜 has always seen them as temporary boundaries. “At the end of the day, everyone wants a universal digital agent.”
He agrees with Anthropic CEO Dario’s view of coding’s foundational role: code is the digital world’s most fundamental fabric and building layer, and almost anything can be expressed through it. A GUI is itself rendered code, so a program can turn it into a CLI, API, or another equivalent interface.
“Coding agent” does not supersede the definition “language agent,” because a programming language is language in the first place. Natural language, programming languages, diagrams, and gestures are all symbolic tools for describing, communicating, and manipulating the world; the differences are not fundamental enough to define separate Agent categories.
13. 苏煜 Entered Agent Research Because He Refused to Let Humans Become Slaves to Computers
At the time, semantic parsing was also a niche in NLP: few researchers worked on it, its papers were harder to publish, and citations could be lower; some professors even told students, “Choose any topic you want, just don’t do semantic parsing.” 苏煜 chose it because the problem mattered more than its status as a field.
He saw software and websites becoming ever more complex: ordinary people needed months of Excel training, and becoming an AWS expert could take years. As the digital world specialized, people were being forced to specialize like machines—“as if everyone were becoming a slave to computers.”
The declaration he left in his PhD defense was: “Let machines understand human thinking, don’t let humans think like machines.”
That is why the move from semantic parsing to language Agents was not a sudden change of field, but a technological upgrade of the same goal after LLMs emerged: the method shifted from building a parser for one environment to making an Agent adapt to many environments.
14. The OpenClaw Moment Reset Interaction, Permissions, and Product Imagination
苏煜 compares the OpenClaw moment with the ChatGPT moment. Before ChatGPT, BERT, ELMo, GPT-1, GPT-2, and GPT-3 had already made the underlying technology increasingly ready; OpenAI merely fine-tuned the models into a chatbot and opened it to the public, but the interaction shift ignited an entire paradigm.
OpenClaw similarly makes Agent researchers feel that “nothing is new here”: coding, tool use, and Agent design all had years of prior work. The real change is that users can interact through entry points such as WhatsApp, while the Agent has its own environment and runs 24/7, always on.
The other change is permissions. Academic labs and large companies previously kept scope, security, and access tightly constrained; OpenClaw largely opens permissions up, letting the Agent do what it wants. Its open-source form makes the risks relatively manageable and lets users see for the first time how much a full-permission Agent can do.
苏煜 believes that, looking back in 2 years, OpenClaw’s impact could approach ChatGPT’s scale: the ChatGPT moment marked a change in the LLM paradigm, while the OpenClaw moment marks the rise of a highly automated personal-agent paradigm.
15. The US Is Going Deep in Developer Circles; China Is Turning OpenClaw into a Mass Narrative
After OpenClaw took off, Anthropic accelerated the inclusion of similar features in Claude Code, while OpenAI wound down experimental projects and shifted toward Agents, productivity, and coding; 黄仁勋 also said every enterprise now needs a cloud strategy. 苏煜 sees OpenClaw more as a catalyst than the sole source of these capabilities.
The US is equally excited, but the heat is concentrated among developers and people close to tech, with the discussion focused on “how to take it deeper.” China quickly turned it into a government, industry, and street-level narrative—both a tool for personal upward mobility and a fear that anyone who does not learn it will be left behind by the era.
张小珺 noted that children too busy working had no time to install it, so elderly family members took their computers to an event and found someone to install OpenClaw. 苏煜 sees both a difference in media-spread patterns and a familiar Chinese speed in front-end applications and technology access.
16. “Good Enough” Foundation Models Are Pushing a Swath of Edge Demand Across the Commercial Threshold
苏煜 believes model intelligence has crossed a threshold: for many useful tasks, “it’s good enough.” Tasks were not ignored because demand did not exist, but because execution friction was too high and the economics did not work; once AI lowers that friction, they move from “not worth doing” to “worth doing.”
What is truly scarce, then, is insight and execution: identifying which activities have crossed the threshold and capturing the value. China’s speed at the application layer may be an advantage here, even if it produces waste such as “paying to install it, discovering it is useless, then paying to uninstall it.”
苏煜 also makes no secret that value realization remains far away: OpenClaw has a high barrier, and most ordinary people may not find a use for it. The research goal should be to let anyone with a distinctive idea turn it into an Agent that creates real value.
17. Model Companies Own the General-Purpose Gateway; They May Not Own Millions of Specialized Small Worlds
Asked whether products must still be built by model companies, 苏煜 answers, “not necessarily.” If the goal is to become the digital world’s single gateway or build a super-general Agent, model companies have an advantage because the product is tightly coupled to the underlying intelligence.
But “this world is not one world”: every profession, domain, company, software product, website, and specific environment has its own micro world. The total entropy of these worlds is nearly infinite, and no single model or Agent can fully capture them.
Value comes from specializing and becoming an expert. Model companies’ organizational structures and platform businesses naturally favor uniformity, so they are often unwilling—or lack a comparative advantage—to go deep on every niche. Application companies, non-model vendors, and individuals therefore still have abundant room.
18. SaaS’s Threat Is Not Immediate Disappearance, but a Forced Shift from High-Margin Tools to Outcome Delivery
On the “SaaS apocalypse,” 苏煜’s answer is “yes and no.” Software valuation logic has changed and margins are being compressed quickly, explaining the sharp selloffs in many software stocks; but he does not believe Agents will replace or rewrite all software.
Much software still has substantial value, and existing solutions are already good enough in many places. Software retaining value and software retaining the high business margins of the past are two different questions.
Software is fundamentally a tool, while customers now want to buy results, labor, and even an AI employee. Traditional SaaS is therefore trying to turn itself from a tool into a labor market: use existing software as the platform, complete the work directly, and deliver the result.
苏煜 has no definitive answer on whether AI-native companies, frontier model companies, or traditional SaaS will win. The shift “only started happening this year,” and the outcome may be a new equilibrium rather than a one-sided takeover.
19. Neo Cognition Is Betting Not on a Single Industry Agent, but on the Method of Learning to Specialize
Neo Cognition is positioned as an Agent research lab with a long-term scope covering every problem relevant to the Agent end state; its medium- and long-term focus is specialized intelligence, rather than continuing to stack general intelligence that is already powerful and gradually becoming standard.
The company is horizontal, with no prior commitment to a handful of verticals such as HR, finance, or legal. Enterprise will be a natural focus because its demands on Agent capability are deeper.
The goal is not to hand-build a batch of specialized Agents, but to develop a continual-learning method: given any domain, profession, or environment, a general Agent should be able to learn quickly and become an expert Agent for that setting.
The company was founded around July-August 2025 and completed a $40M seed round in March 2026, roughly six months later. 苏煜 acknowledges that this was a large seed round and partly attributes the financing to the team’s long track record of work on Agents.
20. AI Funding Is Splitting in Two; Startup Defensibility Comes from Difficulty and Upside, Not “Cannot Be Copied”
苏煜 sees a clear Matthew effect in US fundraising: the scale and valuations of top projects keep getting larger, while ordinary companies find it increasingly difficult to raise money; OpenAI and Anthropic may account for 30-50% of total market funding.
VCs are also splitting: mega-funds such as a16z and Lightspeed can participate through the growth stage, while highly vertical boutique firms build an edge through relationships and deep knowledge. Funds caught in the middle are more likely to struggle.
Investors look beyond résumés to whether the thesis is differentiated, credible, and feasible; whether the team is “one of the best to do that”; whether success would create enough commercial value; and whether OpenAI or Anthropic could easily copy it.
苏煜 does not define defensibility as something big tech can never replicate. He asks whether the problem is difficult enough and its upside high enough: if solving it could fundamentally change society while no one knows the right technical path, it can accommodate multiple players and “crazy bets,” as robotics has.
21. Good Benchmarks and Good Companies Must Both Preserve Construct Validity Against Real-World Value
When building benchmarks, 苏煜 puts the greatest weight on construct validity or ecological validity: the capability being evaluated must be highly correlated with the capability that will ultimately create real value for an AI system, or the benchmark will not have a long life.
He says GPT-5.4 may have used more than 20 benchmarks, 3 of them possibly from students on his team. This was not about chasing rankings; it showed that once the structure of a real task is captured correctly, an evaluation can have a long lifecycle.
Choosing a startup field follows the same logic: do not look for a narrow problem that is easy to win; look for one that would fundamentally change human society regardless of who solves it. Specialized intelligence’s high ceiling and high uncertainty are precisely what allow multiple routes.
22. A World Model Is More Than Video Prediction; It Is a Person’s Entire Understanding of the Work World
苏煜 agrees that visual world models matter: next-frame prediction, video prediction, 3D reconstruction, and learning latent variables for planning as in JEPA are all capabilities that LLMs clearly lack and worth betting on.
But his definition is broader. On a new hire’s first day, the person learns more than the surface org chart; they discover the “actual org chart”—who really has authority, whom to approach for approval, and the software, workflows, and theory of mind among people.
Going from intern to expert is the process of continuously building this micro-world model. It includes vision, but also organizational rules and conceptual relationships that are inherently symbolic, as well as experiential structures that may be neither symbolic nor visual.
An Agent today is “very strong at first use,” but succeeds only around 60-70% of the time; an expert can approach 100% because they know all the ins and outs while working faster and more cost-effectively. Today’s Agents are “not reliable, not efficient, and particularly expensive” because they lack the world model that specialization creates.
23. RL and Markdown Skills Have Not Recreated the Human Path from Intern to Veteran
Model companies currently rely mainly on RL post-training, synthetic environments, and RL gyms. 苏煜 says these still differ markedly from human continual learning in speed, accuracy, what is learned, and scope.
Another route is non-parametric learning, including the SOUL.md and SKILL.md files in OpenClaw and Claude Code, along with the harnesses that are gradually evolving around them. Meta’s Harness and AutoHarness work, for example, attempts to automatically rewrite or optimize these external structures.
苏煜 has long been bullish on the room for non-parametric learning, but believes its ceiling is limited if used alone.
The core problem is not to predefine HR, finance, or legal Agents, but to invent a general learning mechanism for “learning to specialize by learning the world model of work.”
24. Continual Learning’s Key Is Not How Long It Runs, but What It Actually Learns
Classical continual learning means having a model learn new tasks without forgetting old ones; today the term also covers personalization, recursive self-improvement, RL post-training, and OpenClaw-style learning.
苏煜 says “continual learning” by itself is too broad; the question is what the model is actually learning. The biggest gap in the literature, in his view, is that continual learning should build a broad world model rather than merely append rules, task labels, or scattered memories.
Several popular concepts are therefore different parts of the same thing: self-learning and continual learning are the process; a world model is the content learned; specialization and expert Agents are the result; and reliability, speed, and cost-effectiveness are the observable gains.
25. The Neocortex Offers a Biological Hypothesis for How a Unified Learning Machine Develops Expertise
Neo Cognition’s “Neo” comes from neocortex. 苏煜 says the neocortex accounts for roughly 70% of the human brain, appeared relatively late only in mammals, and has more than 200M years of history; a short evolutionary timeline supporting many functions suggests it is unlikely to have evolved a highly specialized structure for each form of intelligence.
A more economical route would be to build a general-enough learning machine and replicate it at scale. The neocortex has functional regions for vision, hearing, language, and so on, but its anatomy is highly similar across regions; its basic unit, the cortical column, is repeated roughly 150K times.
Jeff Hawkins argues in A Thousand Brains that each cortical column learns a world model; a single column may contain hundreds of small models, while the same concept is distributed across multiple columns to drive perception, reasoning, and decision-making together.
苏煜 acknowledges that the theory remains new, with early evidence, and is not a solved account of the human brain. But even without accepting the full theory, the observation that human continual learning produces cognitive maps and conceptual frameworks is relatively mature; AI remains far from reproducing it.
26. Language and Thought Can Be Separated, but Civilization Cannot Escape Symbolic Systems
苏煜 does not deny the importance of vision: the visual cortex is one of the largest regions of the neocortex. But the civilizational gap between humans and apes does not come from sharper sight, hearing, or motor skills; it comes from humans’ unique language and capacity for symbolization.
Terrence Deacon’s The Symbolic Species offers a framework of co-evolution between symbols and the human brain: an early human could tell the tribe about a tiger or food far away, allowing information to cross space and generations; descendants better adapted to a symbolic environment gained a survival advantage, opening a self-reinforcing evolutionary track.
Neuroscience has proposed a dissociation between language and thought: during complex reasoning, language regions such as Wernicke’s and Broca’s areas may not show significant activation. 苏煜 accepts that an individual need not think in language continuously, but argues this reflects knowledge that has already been internalized; before internal connections form, language remains a critical scaffold.
His conclusion is: “Individual thought doesn’t need language, but civilization needs language.” Modern industries, industrial systems, diagrams, and programming languages all rest on symbolic expression; the term “language agent” may eventually disappear because language will have become a default Agent capability, not because the capability itself has disappeared.
27. GUI Will Not Be Eliminated by CLI; Agents Can Leverage Interfaces Human Society Already Built
GUI exists first for humans, and humans are visual animals; visual interfaces can cut comprehension and response times to “zero-point-something seconds,” while also supporting validation, trust, and auditing. GUI will not disappear.
For Agents, most corners of the digital world are already accessible through a GUI, and those interfaces encode accumulated knowledge, constraints, and business logic. An Agent that can use GUIs can piggyback on that accumulation to reach the long tail without waiting for every system to be rewritten as an API.
Text is a one-dimensional representation, while vision can express two-dimensional and even higher-dimensional relationships; complex states are not always described more efficiently by a CLI. GUI is therefore not merely historical baggage and may be the better information representation for some tasks.
Tim Berners-Lee spent more than 20 years advancing the Semantic Web, yet adoption remained low despite a complete technical system—a sign that society will not rewrite everything simply because a new standard is globally superior. MCP and CLI face the same incentives, migration costs, and local-optimum constraints; even critical banking infrastructure still runs on COBOL systems built decades ago, so change will not come quickly.
28. 2026 Competition Will Center on Learning Loops, Not Another Chatbot Shell
苏煜’s 2026 summary is “all roads lead to Rome”: memory, self-learning, continual learning, world models, and specialized intelligence all point to the same bottleneck. Continual learning will be the main theme, but whether it must take the world-model route he favors remains “one of the bets.”
OpenAI, Anthropic, and others currently use a pilot model and hire many forward-deployed engineers to sit with customers and build Agents, which itself shows that reliability, cost, and ease of use remain unsolved.
Security and safety need to be separated: security focuses on worst-case scenarios and still requires dedicated methods, while many safety failures are fundamentally capability failures. 苏煜’s analogy is that an intern is more likely to make a dangerous mistake, whereas a veteran who knows the process and risk boundaries is more stable.
The speed of adoption will depend on how quickly these problems are solved. Once continual learning breaks through, Agents can move beyond a handful of demos and high-end customers into more occupations, organizations, and infrastructure across society.
29. Big-Tech Strategies Are Converging, but a Few Key Contrarian Bets Remain
苏煜 believes the major bets were less clear in the past, but Anthropic has now stood out and shown the way, pushing companies toward convergence: Anthropic is covering productivity broadly, while OpenAI is also narrowing toward Agents, coding, and productivity.
Google leaves him “a little hard to read”: its model capability and ecosystem position are both strong, but adoption and momentum always seem to be missing something. He does not make a definitive attribution, allowing only that “there may be a deeper problem we have not yet seen.”
xAI’s Macrohard is betting on Computer Use and knowledge work, perhaps borrowing from Tesla FSD: a smaller model, video-first, end-to-end modeling. It is at least a different bet; whether it works remains unknown.
Jeff Bezos is participating in Project Prometheus as co-CEO, which 苏煜 understands may have raised $6B-$7B and may build a substantial portion of its work around computer-use agents, extending into manufacturing, logistics, infrastructure, and factories. In China, ByteDance has UI-TARS and the Doubao phone, while 智谱 was early with AutoGLM and AgentBench.
30. Agent Research Must Enter Deployment to Obtain Next-Generation Learning Signals
苏煜 left academia to start a company not because he disliked academia. He once gave up Microsoft compensation worth 3-4x a university salary to teach, because a university is best for exploring 10 “weird ideas” at once; but from 2025, Agent research had moved beyond the low-cost proof-of-concept stage.
The genuinely interesting questions now require GPUs, huge numbers of APIs, a strong team, and rapid iteration; more importantly, they require “continual learning from deployment”: the biggest learning signal comes from real environments, which requires deployment first and is hard for universities to establish.
He did not agonize over whether to become an entrepreneur, only over when the field was ready and what to build. By early to mid-2025, tool use, coding, and multimodality had matured and his view of the Agent bottleneck had sharpened, so he concluded research and production could no longer be separated.
His answer to failure is: “Then I’ll find somewhere else to do research.” That choice reflects the devil-may-care streak he attributes to himself: he does not overthink the fallback, but once he decides something is right, he commits his energy to it.
31. The Imminent Risk Is a Breakdown in Production Relations, Not AI Suddenly Developing a Survival Drive
苏煜 expects continual learning to be solved in the next few years, allowing Agents to spread broadly and fundamentally change the production relations of many industries. Whether that change remains manageable depends on the responsibility of everyone, especially AI researchers.
On existential risk, he sees no foreseeable path by which AI would proactively eliminate humanity simply as intelligence increases. What is missing is not only intelligence but innate goals, intention, and survival pressure; at least for now, he sees no viable way to inject those native objectives into AI.
The more realistic danger is job displacement: if Agents replace knowledge workers quickly, new jobs fail to keep up, gains concentrate among a few leading companies or pools of capital, and effective redistribution or social safety nets are absent, the result will be “a massive impact.”
He places researchers’ responsibility in democratizing the technology: lower the barrier to expert-level Agents so individuals can build Agents from their own insights and create income, rather than letting frontier capabilities be monopolized by a handful of giants because their resource demands are too high.
32. 苏煜’s Intellectual Lineage Ultimately Returns to Evolution, Neuroscience, and World Modeling
His top recommendation is A Brief History of Intelligence, which he says combines AI, evolution, and neuroscience with unusual depth and accessibility. Its chapters “Mice in the Imaginarium” and “Learning by Trial and Error” directly inspired the team’s paper LLMs in the Imaginarium and its work on learning tools through simulated trial and error.
Second is Jeff Hawkins’s A Thousand Brains. 苏煜 acknowledges that the evidence remains preliminary, but considers it one of the boldest unified theories—and one of the ones that makes the most sense—of how the human brain works.
The key paper chain he cites runs from 1940s computational models of single neurons and backpropagation through AlexNet in 2012, Word2Vec in 2013, attention in 2014, Transformer in 2017, BERT in 2018, and then GPT, ChatGPT, Chain of Thought, and Toolformer.
The final key bet is unchanged: “All the way continual learning, all the way world modeling.” In 苏煜’s framework, this is both a technical path and the necessary leap from general-purpose demos to specialized Agent productivity.