I personally tested "lobster" stock picking, thinking I could "sit back and win," but...

“Can I entrust ‘Lobster’ to trade stocks?” With this question, a Chinese securities reporter opened the “Lobster” software. Soon, the reporter found that: ideals are beautiful, but reality is not so easy.

Questions without modular professional toolkits (Skills, which includes directories with SKILL.md files that provide instructions and tool definitions for LLMs) only yield piled-up data. When browsing through Skills guides and requesting various specialized strategy-building needs, the responses often result in “timeout” errors. Perhaps for ordinary individual investors, the human, material, and financial costs of using “Lobster” to trade stocks do not match the accuracy and practicality of the results obtained.

A fund manager told the reporter that their team has not yet integrated applications like “Lobster.” First, from a compliance perspective, such software carries significant risks. Second, existing quantitative models can already quickly handle investment needs such as stock pool screening and strategy backtesting.

From opening “Lobster” to shutting down the computer

Local deployment is an important way to install “Lobster.” However, the reporter’s testing revealed that this method requires very high permissions—it needs to obtain full administrator rights on the computer, exposing personal account passwords and data. If hacked or directed to “go astray,” it could potentially expose one’s funds to high risk.

Next, the reporter tried various “Lobster” applications in the cloud, logging into products from Kimi Claw, Art Claw, JVS Claw, and other major internet companies and AI model providers, and purchased basic memberships for further testing.

The reporter learned that to obtain more reliable and authentic data, installing the modular professional toolkit (Skills) is necessary. For example, with Art Claw, the reporter issued an instruction to install “stock-market-pro,” but was unable to complete the installation.

[Image: Art Claw platform screenshot]

Later, the reporter tried to use a strategy based on “PB-ROE” to get Art Claw to recommend stocks.

[Image: Art Claw platform screenshot]

Although Art Claw outlined a strategy-building approach and provided stock recommendations (see below), the reporter found multiple data errors during the deduction process. For example, the stock price and net profit attributable to shareholders of Kweichow Moutai did not match actual figures.

[Image: Art Claw platform screenshot]

After several hours and multiple attempts to install skills, “Lobster” finally installed the skill and claimed it could fetch the latest real-time stock prices from the Eastmoney API. However, the data still showed significant discrepancies from actual data.

The instructions on the Kimi Claw platform also repeatedly hit walls—if the commands were too complicated, responses would stop. In the first attempt to instruct Kimi Claw to search and install skills for analyzing A-share market data, the system displayed “IM runtime dispatch timed out after 300000ms,” indicating a timeout in computing resources, and the task failed.

[Image: Kimi Claw platform screenshot]

Subsequently, following the example scripts in the “Kimi Claw User Manual,” the reporter tried again. Kimi Claw responded that four professional skills for analyzing A-shares had been created, and it analyzed the financial reports of three stocks for Q3 2025. The results showed the financial data matched the annual reports, with risk warnings and investment suggestions, along with a comprehensive rating.

[Image: Kimi Claw platform screenshot]

The reporter further attempted to install real-time internet news search skills, which succeeded and provided related public opinion information on listed companies. However, when trying to enable the “market monitoring” feature and follow its advice, the system again timed out. The reporter then sought help from the paid K2.5 Agent cluster model at 199 yuan, but the results were unsatisfactory.

[Image: Kimi Claw platform screenshot]

For many ordinary investors, cultivating a smart, responsive “Lobster” requires persistent effort and a relatively rich set of professional skills. Additionally, some investors noted that complex stock screening consumes a large number of tokens, making it costly.

“I can now have ‘Lobster’ send me daily stock market reports, but I need Skills to stay constantly updated to catch the latest iterations. Using some intelligent programming software to assist would improve efficiency,” said an investor using “Lobster” for trading. “The training process has been full of ‘obstacles.’ If I want to add some strategy factors later, I might need further debugging.”

The road to intelligent investing is long and arduous

A fund manager told the reporter that their team has not yet adopted “Lobster”-type applications. Mainly for two reasons: first, from a compliance standpoint, such software carries high risks; second, their existing quantitative models already efficiently meet needs like stock pool screening and strategy backtesting.

“I tried using ‘Lobster’ on my own computer. It can help with some coding tasks, but overall, it hasn’t significantly improved my work efficiency,” said a quantitative fund manager. “Currently, the team has no plans to adopt ‘Lobster.’”

Regarding “Lobster” deployment, Song Weiwei, a fund manager at China Europe Fund, said that unified memory hardware is better suited for OpenClaw deployment. OpenClaw, as a “private AI brain,” has three core requirements: large memory, efficient computation, and persistent operation. In traditional PCs, the CPU uses system memory, while the GPU uses VRAM—these are separate. Data transfer between them involves copying, which is inefficient and wasteful.

Song Weiwei explained that a unified memory architecture, where CPU, GPU, and NPU (neural processing unit) share a physical memory pool, allows seamless access to the same data without copying back and forth. Running large language models is bottlenecked by VRAM, as all model parameters must be loaded into VRAM to operate. On a PC, running a 70-billion-parameter model requires a top-tier graphics card with over 32GB of VRAM, costing tens of thousands of yuan and consuming significant power.

Furthermore, the risks associated with “Lobster” are a concern among industry insiders. Song Weiwei pointed out that relying solely on natural language prompts as safety barriers is extremely fragile. Once AI gains Full Disk Access, any security vulnerability could lead to systemic data leaks. The third-party plugin ecosystem (ClawHub) for OpenClaw may also pose security risks. Additionally, when AI shifts from being a tool to an autonomous executor, traditional responsibility frameworks break down.

If OpenClaw inadvertently leaks business secrets, sends defamatory emails, or even participates in cyberattacks during command execution, who is responsible? The user issuing the commands? The developer writing the code? The underlying model provider? Or the AI itself with “autonomous decision-making” capabilities? Currently, there is almost no legal framework addressing these issues worldwide.

[End of translation]