I've seen many people disapprove of AI Trading, even narrowing it down to "using machine learning to optimize moving average strategies," and I want to share my understanding of this term.

First, we cannot define the future based on the past or present. AI Trading is a process in progress, and it is also a future-oriented concept. Just like when the Agent concept first emerged, humans didn't expect it to evolve so rapidly. From the global popularity of ChatGPT, to the widespread adoption of Transformer architectures, and now to large language models (LLMs), the ceiling of AI Trading is no longer limited by human capability but by models, technical architectures, data feedback loops, and computational power (energy).
AI Trading is not just a simple upgrade of tools but a paradigm-level transformation from "human + tools" to "system autonomy."
1. Not "backtest optimization," but "inventing new market mechanisms"
Just like the invention of electricity, what AI Trading truly needs is invention and creation, not just inductive reasoning or strategy backtesting based on current historical data. Most quantitative trading in the past was essentially statistical arbitrage; if it keeps repeating past patterns, it cannot be considered a paradigm-level advancement.
The real potential of AI Trading lies in its generative and exploratory capabilities. It can invent trading patterns, liquidity acquisition methods, and even new financial primitives that humans have never seen before. Future Agents will no longer be merely tools for executing instructions but will be autonomous entities discovering Alpha on their own. They can build world models in real-time using multimodal data (news, satellite images, on-chain behavior, social sentiment), then generate and validate new hypotheses.
This is similar to the early days of the internet, where initially people thought "email + web pages" were enough, but later new entities like platform economies, algorithmic recommendations, and Web3 emerged.
Major corporations are undoubtedly investing heavily in AI Trading, though these investments are rarely disclosed. The truly valuable assets are never the published papers but the closed-source real-time systems. DeepSeek, in its early days, was a prototype of Liang Wenfeng’s Fantom Fund’s quantitative bot. Many top quantitative teams (including Fantom, Jiukun, Mingxi, etc.) initially used ideas similar to LLMs for signal extraction—just without publicly calling it "AI Trading."
2. Capital, fault tolerance, and accumulated history are the real moats
All this talk is just to say that we should maintain a positive outlook on "the future of the future," because more transformative changes may happen to us.
The big players with capital, wealth, and power, holding vast fault tolerance, are all-in on AI for a reason. When AI makes knowledge borderless and information transparent, the real determinants of victory are money and power. Because the substitutability of knowledge and talent is increasing. If I hire someone today, it might be better to train a 100% instruction-following Agent. The cost of Agents will decrease as computing power becomes more accessible, but human costs are rigidly determined by social economy and living expenses.
Many say, "The wealthy are now also investing in AI," but that’s a misconception. It’s precisely because they’ve made money in other fields, are wealthy and idle, and have extra fault tolerance that they use time to buy future options. Google invested 40 billion USD in Anthropic, Microsoft invested in OpenAI, Amazon invested in various AI infrastructure—these are essentially using surplus cash flow to buy future options. For giants, hundreds of millions of dollars in investment, even if it fails, is just a "department-level experiment"; but if successful, it could reshape the entire capital market (high-frequency trading, market making, asset management, even central bank-level tools).
Returning to AI Trading itself, for ordinary or small-to-medium enterprises (SMEs), competing in this global liquidity race is exhausting and unproductive. The barriers of data, computing power, talent, and regulatory fault tolerance are rising rapidly. Individuals or small teams might focus on niche ecosystems (like specific on-chain protocols’ Agents), but challenging the entire liquidity landscape is difficult. For monopolistic corporations or tech giants, this might just be one of thousands of projects they can implement in their spare time. They have money, power, and historical accumulation, plus fault tolerance. They don’t necessarily need immediate results—just occupying a position and betting on a possibility.
This also means AI could intensify class conflicts and make social stratification more apparent:
• Top tier: Capital and AI form a flywheel, with knowledge workers partially replaced.
• Middle tier: Traditional traders and researchers face intense internal competition from Agents, with skills depreciating rapidly.
• Bottom tier: Information asymmetry might be partially leveled by open-source AI, lowering entry barriers, but overall wealth concentration is likely to increase further.
3. Future landscape: Agent vs. Agent
I remain optimistic about the future but also stay realistic. The future might be Agent versus Agent—AI traders competing against AI traders. In zero-sum or near-zero-sum markets, victory depends on architecture, training paradigms, real-time feedback loops, rather than just model size. The future could be dominated by Agent Swarms.
The most probable evolution paths are:
• 2026-2028: Reinforcement learning combined with LLM Agents, possibly achieving superhuman performance in specific sectors (like crypto, options, cross-border arbitrage), with humans mainly overseeing and intervening in anomalies.
• 2028-2032: Multi-Agent cooperation and adversarial systems become mainstream, with genuine "AI fund managers" emerging, and retail investors directly purchasing Agent portfolios.
• Longer-term: Trading itself might be redefined. When most liquidity is provided by Agents, market microstructure will undergo fundamental changes (lower latency, more complex order flow, more dynamic liquidity pools), rendering traditional backtesting frameworks obsolete.
Of course, risks are equally evident: AI hallucinations, pattern collapses, regulatory crackdowns, and systemic risks (such as multiple super Agents learning the same biases simultaneously, causing market crashes). Therefore, fault tolerance is key to this game.
In summary, expanding imagination and tightening execution
Although the ceiling of AI Trading depends on architecture, data feedback loops, and computational power, humans’ greatest advantage remains the ability to create new games, design new Agent training paradigms, new market rules, and new incentive mechanisms. This core part is still led by humans (or a few top teams).
"Using surplus fault tolerance and time to buy space"—many are anxious about AI stealing jobs, but those who seize the first opportunity are often those who first grow the cake and then share it. Humans must bow to reality, but we can still find our own place.
For AI Trading, the right attitude is "expand imagination + tighten execution." Ordinary people or small teams shouldn’t fight head-on with giants, but can find their own Alpha through edge innovations, niche ecosystems, and open-source collaborations.