Don't you think something's off that only Chinese people are frantically buying up lobsters?

The "lobster" named OpenClaw has been a hot topic for a while now. During that time, many people have asked me to talk about the "lobster," but I always felt that there were many things I couldn't see clearly amidst its explosive popularity, so I didn't respond. These past few days, the public's attention on the "lobster" has finally cooled down a bit, but I still can't resist and have decided to talk about it at a less busy time.

The Origin and Development of "Lobster"

If we only consider the various accounts on social media, then "Lobster" is portrayed as an omnipotent AI agent capable of flying and disappearing. It seems that once deployed, it can write code, run businesses, and manage companies for you, while you, as its owner, can simply sit back, sip coffee, and watch it make money for you. So, how did such a near-omnipotent device come about?

A closer look at the origins of this "lobster" reveals that its creation was actually quite haphazard. In November 2025, Austrian developer Peter Steinberger was traveling in Morocco. Due to the unstable internet connection during his travels, it was difficult to remotely control his computer. So, he came up with an idea: to create an AI assistant that could connect to the instant messaging software WhatsApp and deploy it on his computer. This way, when he was in the wild, he could use WhatsApp to give commands to the assistant to perform corresponding operations on his computer.

With this idea in mind, he directly used "glue code" to connect WhatsApp to Claude Code, creating a simple AI agent. When he sent a message to this agent, it would call the command-line interface to process it and then send the message back to WhatsApp. The entire process reportedly took only an hour. After completing the project, Steinberg named it Clawbot and casually open-sourced it on GitHub. Unexpectedly, this simple project became a popular project on GitHub as soon as it was released. In the following weeks, Clawbot's number of stars on GitHub exceeded 100,000. Including Steinberg, many developers not only spontaneously improved its functionality but also created a lobster meme for it—this is the origin of the "lobster" meme.

However, the explosive popularity of "Lobster" quickly drew the ire of Anthropic. Anthropic argued that Clawbot's various functions were essentially just amplifiers of Claude's capabilities, and its name was easily confused with Claude's, demanding that Steinberg immediately cease infringement. While Anthropic's attitude seemed overbearing, its accusations weren't entirely unfounded. At the time, Clawbot's capabilities did indeed primarily derive from calling Claude models, and its automated operation was so resource-intensive that it was even jokingly referred to as a "Claude addict." Its capabilities relied almost entirely on the proxy generation and code repair functions of Claude Code, making it fair to say it was merely a Claude shell. Under pressure from Anthropic, Steinberg had to abandon the deep integration with Claude, expanding its support from solely Claude models to include multiple AI models, and changing Clawbot's name to Moltbot.

Because the name Moltbot was too difficult to pronounce, Steinberg soon changed it to OpenClaw. Unexpectedly, this seemingly forced adjustment inadvertently set the stage for OpenClaw's future. As OpenClaw began supporting multiple models, this intelligent agent quickly attracted the interest of numerous AI companies (especially Chinese AI companies), and its subsequent popularity largely benefited from their promotion. The "Open" in its name led to an offer from OpenAI for Steinberg. Shortly after, he accepted an invitation from Sam Altman and joined OpenAI.

What is a lobster? And what isn't it?

Through the brief review of OpenClaw's history above, it's clear that OpenClaw is essentially a proxy layer between the user's computer and the large AI model; it doesn't actually have any capabilities of its own. The various capabilities seen in social media ultimately come from these large models; OpenClaw's role is simply to help users manipulate them. Therefore, the effectiveness of OpenClaw depends on the capabilities of the models it uses. In fact, after being forced to part ways with Claude, many users reported that OpenClaw's capabilities when calling other models were significantly diminished. This is because the models it subsequently used had considerably lower capabilities compared to Claude.

Of course, it's not objective to think, as some people do, that "OpenClaw is just a shell and has no meaning in itself." As a proxy, it can still help people achieve many functions that are difficult to implement when directly calling large models.

First, because Open Claw can be deployed locally, it can directly operate on the local computer and manage hardware, given the necessary permissions. This is akin to giving the AI agent hands and feet. In contrast, large AI models and cloud-based agents (like the early Manus), no matter how intelligent, have limited control over local hardware. It's important to note that this convenience of operating local hardware comes at a cost, meaning it can potentially carry higher risks. Conversely, accessing large models through the cloud or deploying agents in the cloud effectively creates a natural "sandbox," offering greater security.

Secondly, OpenClaw possesses long-term memory that large models lack. When using large models, we often encounter a situation where we made many requests in a previous conversation, and the model complied. However, in a different conversation, it completely forgets these requests, requiring us to provide them again. OpenClaw, on the other hand, constructs an externalized long-term memory mechanism. It writes key information such as user preferences, task results, and lessons learned from errors into external storage and injects it into the context of subsequent tasks. In this way, OpenClaw possesses long-term memory, eliminating the need for repeated prompting.

Furthermore, OpenClaw can continuously expand its capabilities through Skill files. A Skill file can be understood as a standardized task description, in which developers can write the workflow for the agent to directly call later. Since these Skills can be shared and reused, a new skill learned by one "lobster" can be quickly copied to other "lobsters." Based on this, OpenClaw can also continuously plan and execute tasks through task loops, thus working until the task is completed. In contrast, large AI models can only provide suggestions and rarely perform tasks directly.

Based on the above analysis, if we compare a large AI model to an advisor, then OpenClaw is an agent, more like a subordinate to whom tasks can be delegated. You no longer need to operate step-by-step; instead, you simply say, "Help me do a competitive analysis," and it starts researching, organizing data, and generating reports on its own. This experience is indeed novel and effective in some scenarios. However, the problem lies precisely here: misinterpreting "can do work" as "can do everything" is the most common misjudgment regarding OpenClaw.

First, it hasn't become smarter. OpenClaw itself doesn't incorporate new cognitive abilities; its capabilities still depend on the underlying model. If the model itself has limitations in reasoning, understanding, or generation, then no matter how complex the outer agent structure is, its performance will ultimately not exceed this limit. In other words, what "Lobster" can do largely depends on which model you connect it to, rather than on what new intelligence it has "evolved."

Secondly, its ability to act doesn't equate to true understanding. While it can indeed break down tasks and execute steps, this breakdown is often based on probabilistic language generation rather than stable, rigorous planning capabilities. Therefore, in complex tasks, it's prone to going astray, such as repeating operations, misunderstanding objectives, or even getting stuck in loops. If such situations arise without timely intervention, the user may end up with nothing more than a hefty lexicon fee bill.

Furthermore, its execution capability does not equate to reliability. Being able to invoke tools does not guarantee correct invocation; continuous operation does not imply that every step is reasonable. This uncertainty can sometimes pose real risks when file operations, system commands, or even external interfaces are involved. In this sense, it's more like an intern requiring supervision than an employee who can be completely delegated.

Who is lobster suitable for? Who is it not suitable for?

After OpenClaw suddenly became a hit, many people developed a "FOMO mentality" (Fear of missing out), worrying that they would fall behind if they didn't quickly adopt or learn "Lobster," and hoping to double their efficiency or even master the so-called wealth code by learning it. To this end, they were willing to spend thousands of yuan to install "Lobster" and tens of thousands of yuan to attend related courses. From a behavioral economics perspective, this FOMO mentality is not incomprehensible. After all, existing research shows that the pain of missing out on opportunities that could have been seized is enormous, and incurring certain costs to avoid such losses cannot be simply regarded as wasteful.

However, if we set aside these anxieties and conduct a more objective analysis, we'll find that "lobster" (OpenClaw) may not be suitable for everyone. Objectively speaking, OpenClaw is a highly selective tool. Unlike search engines or chatbots, it's not "somewhat useful" for almost everyone; rather, it's more like a specialized tool—used correctly, it can exponentially improve efficiency; used incorrectly, it not only fails to help but may also increase the burden or even trigger risks.

So, who is more suitable for "raising lobsters"? In my opinion, it mainly includes the following categories.

First, there are "process-oriented workers" whose work is inherently procedural. Their daily work isn't completed within a single interface, but rather requires constantly switching between multiple tools to connect a series of steps. Take e-commerce operations personnel as an example: their daily work includes reviewing competitor data, compiling keywords, writing product descriptions, uploading products, monitoring conversion rates, and then adjusting strategies based on the data. While each step can be assisted by large models, the real time-consuming part is the connection between steps and the repetitive operations. The value of OpenClaw lies in integrating these "fragmented operations" into a continuous execution process. With OpenClaw, they no longer need to click through each step individually; instead, they can define a goal and let the system automatically complete the entire process, thus significantly improving efficiency.

Secondly, there are the "One Person Company" (OPC) or micro-teams that have emerged in recent years. The core problem faced by these individuals is never a lack of ideas, but rather a lack of manpower. In reality, they often need to simultaneously handle multiple roles such as product, content, marketing, and customer service, quickly exhausting their time. The significance of OpenClaw lies in transforming the "scaling problem" into a "scheduling problem." With OpenClaw, they no longer need to expand by hiring more people, but rather by designing processes and utilizing multiple intelligent agents to extend their capabilities. For example, content creators can have "Lobster" handle information gathering, material organization, and draft generation, while they themselves perform the final review; independent developers can have it automate testing, fix simple bugs, and even complete part of the deployment process. In these scenarios, OpenClaw can effectively compress tasks that originally required multi-person collaboration into the scope that one person can manage, with the user simply acting as a manager.

Secondly, there are those who already possess certain technical skills or an automation mindset. For programmers, data engineers, and even users familiar with scripting tools, OpenClaw acts as a "capability amplifier." They could already automate tasks using scripts, but it required time to write logic and handle exceptions. Now, these tasks can be partially delegated to intelligent agents. For example, previously, they needed to write scripts to scrape web pages, parse data, and generate reports; now, they can run a version of OpenClaw first and only need to make adjustments at key stages. The advantage of these users is that they understand which steps can be automated and which require manual intervention, thus enabling them to better leverage the value of OpenClaw.

If we abstract these groups of people, we'll find a common thread: their work can be described in a "process-oriented way." They know what they need to do and how to complete it step by step. As long as these two questions are clear, OpenClaw's capabilities can be fully utilized.

Following this line of thought, we can also understand which groups of people OpenClaw is not suitable for.

First, there are those whose work revolves around "problem definition," such as academics, strategic consultants, and creative professionals. The difficulty for this type of work lies not in execution, but in constantly reconstructing the problem itself. Their research direction today may be overturned tomorrow; a creative solution often requires repeated trial and error and intuitive judgment. In this situation, it is difficult for them to design a stable process in advance, and without a process, Open-Claw cannot function effectively. Although theoretically it can help with tasks such as researching and writing initial drafts, the added value is not significant, and may even be negative—because these "odd jobs" are often important sources of inspiration, and automating them may actually inhibit thinking.

Secondly, there are users whose tasks are relatively simple or infrequent. For example, writing emails, making simple reports, and researching information are tasks for which a standard large-scale model is sufficient, and is more direct and controllable. Introducing OpenClaw would actually require additional configuration, debugging, and monitoring costs.

Secondly, there are users who lack process-oriented thinking skills. Many people believe that as long as they tell AI "do something for me," it will automatically understand and execute it. But in reality, if the goal is vague and the constraints are unclear, OpenClaw often tries repeatedly in a seemingly diligent manner but never quite gets to the point. For example, if we simply tell OpenClaw to "do a market analysis for me" without specifying the scope, data sources, and output format, it may consume a large number of keywords, grab a lot of irrelevant information, and ultimately generate a lengthy but insightful report.

A popular claim on many social media platforms is that "Lobster" is a tool for equalizing capabilities, allowing everyone to equally master powerful AI abilities. However, as the analysis above shows, this claim is inaccurate. In reality, "Lobster" acts more like a capability filter, dividing people into two categories: those who can break down complex tasks into processes, and those who rely on intuition and experience but struggle to externalize those processes. The former will find "Lobster very useful," while the latter may find it "nothing special." This isn't about the tool being fair or unfair to different people, but rather its inherent reliance on process-oriented thinking. Understanding this allows us to decide whether to "raise Lobster" based on our own needs, avoiding unnecessary costs driven by FOMO (Fear of Missing Out).

"Lobster Craze": Metaverse Carnival 2.0?

If we observe recent public opinion regarding "lobster," we'll find a phenomenon: significant differences exist in attitudes towards "lobster" around the world. According to Google Trends data from March 22nd, the search index for OpenClaw in mainland China reached 100, the highest globally. Hong Kong, Singapore, South Korea, and India followed closely behind, with significantly lower search indices. Even the United States, a long-standing leader in AI innovation, had a search index of only 9 for OpenClaw. European countries showed a more lukewarm attitude towards "lobster," with relatively low search activity in the UK, France, and Germany. Among European countries, Austria was the only one showing significant interest in "lobster," but this interest is primarily attributed to being the birthplace of Peter Steinberg, the "father of lobster."

The content of public opinion also shows significant differences across regions. In China, discussions about OpenClaw mainly focus on how to transform it from a "geek tool" into a "national application" and how to profit from it; in the United States, related discussions focus more on relatively specialized issues such as using OpenClaw to replace traditional SaaS tools; as for Europe, the focus on OpenClaw is more on compliance issues and how to deal with the resulting privacy and data breach risks.

It's worth noting that this phenomenon is not accidental. In the past few years alone, similar differences in public opinion narratives between China and other countries have occurred repeatedly. The most representative example is the metaverse narrative. While the metaverse concept was also popular abroad, only a few giants like Meta were truly fully committed; the rest were mostly participants in the cryptocurrency world fueling the hype. In China, however, a nationwide frenzy for the metaverse emerged. Not only did major internet companies jump in, but local governments also rushed to introduce related plans. Comparing the "lobster craze" with the metaverse craze reveals a striking parallel: the present moment is remarkably similar to that past one.

A further comparison of the drivers behind these two major narratives reveals similarities in both the driving forces and their motivations. The main driving forces in these two surges can be broadly categorized as follows:

First, there are the internet giants. A few years ago, China's internet giants were active promoters of the metaverse narrative, and today, they are similarly key drivers of the "lobster" craze. The reasons are quite similar: back then, their promotion of the metaverse was primarily aimed at driving the development of cloud services. After all, building a metaverse requires massive computing power, which small and medium-sized enterprises (SMEs) cannot obtain on their own and can only rely on the cloud services of the giants. Currently, the operation of "lobster" also relies on large models, which again provides opportunities for the giants to promote cloud services and sell meta-elements.

Of course, compared to the metaverse era, the "big companies" now have an additional purpose in promoting the "lobster" concept: to compete for entry points and user data. In the digital economy era, user entry points and behavioral data are core resources, and in the existing internet ecosystem, these entry points are concentrated in the hands of a few super applications. The emergence of OpenClaw provides "big companies" with a new path: if users directly deploy it on their terminal devices, and the company provides model services in the background, then it is possible to regain control of the entry points through this "lobster." In this sense, the "lobster" competition is not only a competition for the meta-market, but also a competition for entry points.

Secondly, local governments played a crucial role in the recent "lobster" craze. In fact, when "lobster" was still largely confined to the geek community, it was the local governments of two regions that released policy documents supporting related industries, quickly bringing it into the public eye. A similar situation occurred during the metaverse craze of that era. The reasons why local governments are willing to play this role are quite clear: in the current economic environment, finding new growth points and stabilizing employment are important tasks. Against this backdrop, things like metaverse and OpenClaw, which combine technological imagination with investment-driving effects, naturally become the focus of policy attention. In particular, OpenClaw and related concepts such as OPC can, to some extent, fulfill the goals of job creation and expanding the tax base, thus making them more likely to receive support.

Secondly, there are hardware manufacturers, most notably Jensen Huang. He has played a crucial role in both the metaverse craze and the current "lobster" hype. The reason is actually quite simple: Nvidia, as a provider of computing infrastructure, is a typical "shovel seller"—it benefits from any increase in computing power demand. Therefore, it has an incentive to promote any technological narrative that consumes high computing power. In this sense, both the metaverse and the "lobster" are not just technological trends for Nvidia, but also real business opportunities.

Based on the above analysis, it can be seen that, similar to the metaverse narrative of that time, the main drivers of the "lobster" craze each had clear goals. Therefore, the narratives they constructed inevitably contained a certain degree of marketing and amplification.

But does this mean we shouldn't pay attention to "lobster"? In my opinion, the answer is no. From a technological development perspective, moderate market enthusiasm is often a significant catalyst for innovation. Without attention and financial support, technological research and development often struggles to progress sustainably. Even if a bubble eventually forms, from an overall cost-benefit perspective, it's often an acceptable suboptimal path. Just like the legacy of the metaverse boom—computing infrastructure and AI talent reserves—which ultimately became crucial support for the development of large-scale models. From this perspective, even if the current "lobster" craze has some speculative elements, it still constitutes a market opportunity worth paying attention to.

Of particular note is the fact that the fierce competition among major model suppliers during the previous "battle of a hundred models" led to a significant drop in word unit prices. Currently, the word unit prices for large Chinese models are significantly lower than those of foreign competitors. While this price competition benefits users in the short term, it compresses profit margins and limits R&D investment in the long run. The emergence of "lobster" (a type of Chinese model) opens up a new demand scenario for companies, helping to alleviate this competitive situation.

The above discussion shows that while the "lobster" craze contains elements of hype, it still represents a potentially valuable opportunity overall. A more sensible approach to this phenomenon is to neither excessively praise nor excessively criticize it, but rather to maintain a rational judgment. Specifically:

For the average user, if the opportunity to "raise shrimp" arises, it's certainly worth a try; however, if not, there's no need to develop excessive FOMO (Fear of Missing Out) as if you'll miss out on the next technological wave. In fact, at this stage, most people don't need to rely on intelligent agent systems for their daily work, and OpenClaw isn't a necessity. If you're considering using OpenClaw to improve efficiency, ask yourself: Can your work be broken down into stable processes? Are there many repetitive operations across different tools? If the answer is no, then instead of investing a lot of time in learning and debugging, it's better to continue using more direct, large-scale model tools, or simply wait for the technology to mature further.

For AI companies, this wave of enthusiasm has actually provided a rare window of opportunity. OpenClaw has attracted attention not because it's already mature enough, but because it demonstrates a clear direction: users are beginning to need "AI capable of performing tasks." However, its current form also exposes many problems, such as complex deployment, insufficient stability, rudimentary access control, and high security risks. These are precisely the areas where companies can focus their efforts. Instead of simply following the trend and launching "Claw X," companies should focus on improving underlying capabilities, such as lowering the barrier to entry, enhancing execution reliability, and refining access control and sandbox mechanisms. The real value lies not in creating more "lobsters" (AI clones), but in creating more usable and controllable products.

From the government's perspective, restraint is even more crucial. Supporting new technologies and cultivating growth points is important, but over-betting on technologies in their early stages could amplify uncertainty and lead to path dependence. A more sensible approach is to strengthen fundamental capabilities, such as improving computing infrastructure, facilitating the flow of data elements, and promoting the establishment of standards and security norms to provide a level playing field for different technological paths. Simultaneously, it's necessary to proactively consider the new challenges brought about by intelligent agents, such as defining responsibilities in automated execution, data access boundaries, and potential security externalities. Within this framework, the policy focus should not be simply "supporting lobsters," but rather on building an institutional environment conducive to the evolution of diverse technological forms.

In conclusion, individuals, businesses, and governments alike should remain rational when facing the "lobster craze": neither blindly following the trend due to its popularity nor easily dismissing it because of uncertainty. Only in this way can the "lobster craze" become a genuine opportunity, rather than just another bubble.