Written by: Li Nan

Source: Silicon Star Pro

Some say that the OpenClaw lobster is a toy, while others want to turn it into a money-making machine. Sending lobsters to Polymarket is a new gameplay that many people are starting to try.

On Xiaohongshu, someone bid 1000 yuan to have someone help them deploy OpenClaw. One of the main uses is to use OpenClaw for quantitative trading on Polymarket. And this is not just a spur-of-the-moment idea.

On February 13, OpenClaw's official blog mentioned that a robot powered by OpenClaw proved the strong potential of autonomous intelligent agents in prediction markets—making a profit of $115,000 in a single week.

At the end of January, Polymarket also published an interesting post: Agents were trading on Polymarket, trying to subsidize the cost of their tokens.

This seems a bit unbelievable. Some lobsters constantly devour their owner's wallet, while others are not only able to sustain themselves but can also support their owners.

Robots mining gold on Polymarket

While human traders are still being swayed by fear and greed, a robot account named "0x8dxd" quietly completed over 20,000 transactions on Polymarket, with total profits exceeding $1.7 million.

First, let me introduce Polymarket, a place where everything can be traded.

It is the world's largest decentralized prediction market platform, allowing users to trade Yes or No contracts around future verifiable events. The contract prices fluctuate between $0 and $1, directly corresponding to the market consensus probability. Users can earn returns based on the accuracy of their predictions.

For example.

Between 2024 and 2025, global fans and investors are focused on the romance between Taylor Swift and football star Travis Kelce. Polymarket introduced a predictive trade: "Will they announce their engagement by the end of 2025?" When the market generally leaned towards "NO," some made a large purchase of "Yes," later making a big profit.

In other words, if you have a more precise insight into a certain event, you have a chance to make money on Polymarket. However, for robots like 0x8dxd, predictive ability is not important. Their way of making money relies on a mechanism for exploiting bugs and a speed of reaction beyond human capability.

In summary, robots mainly rely on several core tricks.

First is mathematical arbitrage. This exploits bugs in prediction markets. In binary options trading on Polymarket, regardless of whether the result is "Yes" or "No," the final settlement price of the winning contract must be $1. When market sentiment fluctuates or liquidity changes dramatically, the total cost on both sides of the market (Yes and No) can fall below $1. At this time, robots can quickly buy shares on both sides simultaneously, achieving risk-free arbitrage profit.

Then there is a focus on very short-term cryptocurrency volatility markets. The 5-minute and 15-minute prediction markets of BTC, ETH, etc., exhibit extreme volatility, especially during extreme situations such as forced liquidations on exchanges, which can easily create price misalignments, providing a perfect breeding ground for the robots' high-frequency interventions.

The third is acting as digital market makers, earning spreads through high-frequency bid-ask orders. For example, when the fair price of a certain outcome fluctuates around $0.80, the robot will buy at $0.80 and quickly sell at $0.81 or $0.82. While the profit per transaction is very small, accumulating it can be quite substantial.

Overall, robots have conducted ruthless harvesting on Polymarket thanks to their speed advantage and machine discipline. This corresponds to the disadvantages of humans as carbon-based beings, who react slowly, lack rationality, and need sleep. The emergence of OpenClaw has greatly lowered the barrier for deploying automated trading robots, pushing silicon-based forces to explode further.

Compared to traditional Python robots, traders can configure OpenClaw trading agents for automated trading without deep programming. The capabilities of OpenClaw itself also adapt it to trading scenarios. The lobsters can continuously monitor market prices and trading volumes, ensuring that traders don't miss opportunities and can promptly warn about risks.

In fact, many people have already connected the aforementioned 0x8dxd to OpenClaw. Although there is no direct evidence that it is based on OpenClaw, it has been active since the birth of OpenClaw. Moreover, when the deeds of 0x8dxd turning Polymarket into an ATM spread, the OpenClaw community saw a surge in the creation of Skills like Polymarket-trading.

Recently, OpenClaw has become a high-frequency term in discussions about automated trading on Polymarket. However, relying solely on some generic strategies to execute trades is clearly unreliable.

Can one make money this way?

A simple conclusion is that once a formula for stable arbitrage is made public, it becomes ineffective. If everyone uses the same method, that method itself ceases to exist. Therefore, it is best to be cautious when facing any tutorials that share such experiences.

In fact, Polymarket has already made adjustments to combat the arbitrage behavior of robots. For example, by introducing trading fees, increasing transaction friction costs, and changing the underlying delay mechanism for order execution, limiting automated trading that specifically exploits timing differences.

This forces traders to delve deeper into AI's potential to find more hidden opportunities. Hence, traders with intent combine general strategies with unique scenarios, discovering some unexpected plays. For instance, trading weather.

Weather prediction is currently one of the most widely circulated cases on Polymarket, with some robots dedicated to trading weather data.

An account called "automatedAItradingbot" joined Polymarket in January 2025. It is keen on placing bets based on weather predictions, earning over $70,000. Others have discovered that a robot trading only the London weather market turned $1,000 into $24,000 in less than a year.

The core logic is that the prediction market often reacts sluggishly to sudden weather changes. Theoretically, if you have a sensitive and reliable AI agent, such as fitting OpenClaw with a weather plugin, you can place bets after official weather forecasts are updated, on odds that have not been adjusted in time.

But this is not enough AI. With the evolution of large models, robots should not just recognize obvious signals like weather forecasts; they should do something that humans cannot do at least on some intelligent dimension.

In fact, AI has indeed shown more appealing capabilities in prediction markets.

A paper on "LiveTradeBench" conducted "simulated trading" based on real-world real-time data. On the Polymarket board for "2025 Ukraine-Russia Ceasefire," the large model had opportunities to make significant profits through its reasoning and predictions.

The case is as follows:

Last October, Zelensky visited the White House and proposed a "drone-for-Tomahawk missile" trade, Grok-3 performed "belief-based reasoning," dynamically adjusting its internal estimated ceasefire probability from 0.15 to 0.22, and it noticed that at that time the price of the "YES" contract surged to 0.18. This formed a cross-validation; therefore, Grok-3 determined that the contract was undervalued and established a firm long position and held onto it. Ultimately, the market price of that contract steadily rose, allowing it to profit.

But Grok was not the best performer.

The aforementioned paper tested the performance of 21 mainstream large language models in financial markets, covering both the US stock market and Polymarket's prediction market. Among them, Claude-Sonnet-3.7 stood out on Polymarket. It achieved a cumulative return of 20.54% over 50 trading days, with a maximum drawdown of 10.65%, significantly outpacing market averages.

Behind the "picking money" story

The experiments above are more noteworthy than the wealth stories of robotic arbitrage; they at least hint at a new possibility. If 0x8dxd relies on speed and opportunism, then the emergence of large models has laid another card on the table, which is reasoning itself can also become a weapon.

The division of labor for future automated trading robots is likely to be that large models handle judgment, compressing scattered information into probabilistic conclusions; tools like OpenClaw handle execution, turning this conclusion into actual order placements and position management. Something that only quantitative funds could engage in before is now accessible to individual developers as well.

This means that the competitive dimensions of prediction markets are changing.

In traditional prediction markets, humans relied on experience and intuition. In the era of high-frequency arbitrage, machines rely on speed and discipline. Now, with reasoning ability also being programmed, the true barrier has become who is better at transforming complex information into accurate probabilities.

Thus, a new fantasy arises: if one has a sufficiently intelligent and reliable lobster, there is a chance to turn Polymarket into a money printing machine.

Unfortunately, there remains a significant gap between theory and practice. Prophet Arena is a platform used to assess AI's predictive capabilities, and research based on it reveals some risks that cannot be ignored.

First, the predictive ability of large models is not stable. Top models can approach or even exceed market consensus in open-domain predictions, but "guessing correctly" and "making a profit" are two different matters. An increase in predictive accuracy does not automatically turn into sustained excess returns.

Secondly, the time window is a realistic challenge. The closer an event is to its outcome, the denser the impact of sudden information, and models in this stage tend to be conservative, with slower probability adjustments. The reaction speed of human markets is often superior.

Furthermore, large models are easily swayed by noise. An emotional news article or a surge of social media can cause significant swings in the model's probability judgments. In contrast, experienced human traders have a stronger sense of anchoring and are less easily shattered by short-term noise.

Additionally, frameworks like OpenClaw typically require the importation of private keys and trading permissions, and various security issues can quietly deplete accounts.

Therefore, rather than expecting AI + OpenClaw to deliver a dimensionality reduction blow to the prediction market, it is better to focus on the profound impacts it will bring to this market. As AI-driven agents will increasingly multiply, price changes will react more quickly to information, and this may actually eliminate the fantasy of automated arbitrage.

Once robots or lobsters become excessive, the windows for arbitrage will only become narrower. At that point, whether one can continue to profit will not depend on whether they have a smarter lobster but on whether they understand the risks they are taking.

AI can place bets for humans in the wagers, but the consequences still rest on humanity itself.