The Rise of Agentic AI

​Something I've been wondering - if the markets evolve to a point where most trading is done by computers of similar efficiency, will the market flatline, going sideways, making humans trading for profit more or less impossible? It's apparently an ongoing unsettled debate at this point among the smart people who design and build AI agents.

 

Eric Schmidt has been publicly talking about Agentic AI a lot lately, here's one example

https://www.youtube.com/watch?v=_gBxYL2ihc0

 

I asked Gemini what it thinks:

--------------------How will agentic AI affect trading markets for profit? Does it "level the playing field" to the point the markets are effectively flatlined?

Agentic AI, which can act autonomously to make decisions without human intervention, is poised to have a profound and transformative impact on trading markets. This evolution from traditional algorithmic trading, which follows pre-defined rules, to independent, self-adapting AI agents will bring significant changes to the financial landscape.

How Agentic AI Will Affect Trading Markets for Profit

1. Increased Efficiency and Speed: Agentic AI can process vast amounts of data—from earnings reports and news feeds to geopolitical events and social media sentiment—in real time. It can then execute trades in milliseconds, far surpassing human capabilities. This will lead to an acceleration of market activity, with strategies being developed and deployed at an unprecedented pace.

2. Enhanced Strategy and Risk Management: Unlike static algorithms, agentic AI can dynamically adjust its strategies based on new information. This includes not only recognizing patterns but also responding to unexpected market shifts. The AI can manage portfolios, rebalance assets, and even modify hedging positions autonomously to maximize returns and mitigate risk.

3. Democratization of Sophisticated Strategies: Traditionally, high-frequency trading and complex quantitative strategies have been the domain of large financial institutions and hedge funds with significant resources. Agentic AI has the potential to make these advanced capabilities more accessible to a wider range of participants, including smaller firms and even retail investors, who could use personalized AI agents to manage their finances.

4. New Revenue Streams: For financial institutions, agentic AI will not only amplify existing revenue streams by optimizing trading, but it could also create new ones. This could include offering AI-driven investment products or "robo-advisors" that provide hyper-personalized strategies tailored to an individual's specific goals and risk tolerance.

The "Leveling of the Playing Field" and the Risk of Flatlined Markets

The idea that agentic AI will "level the playing field" to the point of flatlining the markets is a complex and highly debated topic.

• Argument for a "Flatlined" or Highly Efficient Market: If a large number of market participants, from large institutions to individual investors, are all using highly effective AI agents that are making decisions based on the same information and with similar efficiency, it could lead to a situation where all profitable opportunities are exploited almost instantly. This would reduce arbitrage and other short-term profit-making opportunities, theoretically making markets incredibly efficient and potentially leading to a "flat" or low-volatility environment. In this scenario, the market would become a perfect representation of all available information, and there would be little room for individual profit from superior analysis.
• Argument Against a "Flatlined" Market: The more likely outcome is not a flatlined market, but a more volatile one. While AI agents might lead to greater market efficiency in some respects, they could also introduce new risks and complexities.
• Herding Behavior: If many AI agents are trained on similar data and use similar logic, they might react to the same market signals simultaneously. This "herding" behavior could amplify market swings and lead to "flash crashes" or other periods of extreme volatility.
• New Forms of Competition: Instead of leveling the playing field, agentic AI could simply shift the competitive advantage. The new differentiator would be who has the best AI models, the most unique and proprietary data to train those models on, and the fastest infrastructure to deploy them. The "game" would move from human skill to machine-to-machine competition.
• Unforeseen Outcomes: The "black box" nature of some AI models, coupled with the potential for errors or malicious manipulation, could introduce new and unpredictable risks. An AI algorithm might act on a wrong prediction, causing rapid market fluctuations, or an adversarial attack could be used to manipulate market behavior.

In conclusion, while agentic AI has the potential to make markets more efficient and accessible, it is unlikely to "level the playing field" to the point of a flatline. The competitive landscape will likely evolve, with the new advantage going to those with the best AI, data, and infrastructure. The most significant risk isn't a lack of profit opportunities but an increase in market volatility and systemic risk due to the autonomous, interconnected, and potentially synchronized actions of these AI agents.

 

 

------------------What happens when state of the art AI's compete at playing chess, or other complex games?

When state-of-the-art AIs compete at complex games like chess or Go, it leads to a new kind of "superhuman" competition that goes beyond human comprehension and reveals new insights into the games themselves. The most famous example is the rivalry between AlphaZero and Stockfish, two of the top chess engines in the world.

The Rise of AI vs. AI

For decades, chess engines operated on a "brute-force" approach. Engines like Stockfish would calculate millions of possible moves per second, using a highly optimized search algorithm and a complex, hand-coded evaluation function to determine the value of a board position. This was a direct extension of human expertise, encoded into a computer.

The landscape changed with the arrival of AlphaZero from Google's DeepMind. Instead of relying on human knowledge, AlphaZero was given only the rules of chess and a framework for self-play. It then played millions of games against itself, learning from scratch through a process called reinforcement learning, using a deep neural network to evaluate positions and guide its search.

The Results of the Competition

When AlphaZero was pitted against Stockfish, the results were a landmark moment in AI history. AlphaZero, despite analyzing far fewer positions per second than Stockfish, consistently beat it. The reason was a fundamental difference in their approach:

• Stockfish: Relied on a deep, narrow search. It was a tactical powerhouse, calculating precise sequences of moves to find the best immediate outcome.
• AlphaZero: Utilized a more intuitive, human-like approach. Its neural network allowed it to "see" the board in a holistic way, identifying complex, strategic patterns and making moves that a human might describe as "beautiful" or "creative." It would often sacrifice material for long-term positional advantages that were unfathomable to traditional engines.

This competition demonstrated that AI could not only surpass human skill but also develop entirely new, innovative strategies that were previously undiscovered. The games played by AlphaZero are now studied by human grandmasters to gain new insights into chess theory.

Broader Implications in Complex Games

The lessons from the AlphaZero vs. Stockfish match extend to other complex games and AI competitions:

• Emergence of New Strategies: In games with vast state spaces, such as Go or Starcraft II, AI-vs-AI competition has led to the emergence of novel strategies that break from traditional human play. AlphaGo's victory over Lee Sedol in Go, for instance, involved moves that were initially considered mistakes by human experts but were revealed to be genius, long-term strategic plays.
• A Shift from Brute Force to Intuition: The most successful modern game-playing AIs are not just faster calculators. They use machine learning to develop a kind of "intuition" for the game, allowing them to make smart, strategic decisions without exhaustively exploring every single possible future state.
• A New Benchmark for AI: The ability to beat other AIs has become a new benchmark for evaluating AI systems. Platforms like Kaggle's Game Arena now pit different large language models and AI agents against each other in various strategic games to test their reasoning, long-term planning, and adaptability.
• Accelerating Discovery: When AIs compete, they are essentially exploring the game's possibility space at an incredibly accelerated rate. This process can lead to the "solving" of certain games or, more broadly, to a deeper understanding of game theory and optimal decision-making in complex environments.

Ultimately, the competition between state-of-the-art AIs in complex games is not just a digital spectacle. It's a fundamental part of AI research, demonstrating the power of autonomous learning and revealing insights that were previously beyond human reach. It shows that AI can not only solve problems but also discover new ways of thinking and playing.

 

 

------------------------But my basic question is what happens when AlphaZero plays chess against it's self? It's an important question because at some point AI's will reach similar capabilities as they self evolve efficiently and quickly. There is an earlier corollary with high speed trading and co-located servers, but there was not enough volume there to effect broad markets very much, but Agentic AI will presumably go well beyond that limited participation "problem".

 

This is an excellent and insightful question that gets to the heart of the "endgame" for many competitive systems, from games to financial markets.

AlphaZero Playing Itself: The Training Process

First, it's important to understand that AlphaZero's primary mode of learning is to play against itself. This is the entire premise of its "tabula rasa" (blank slate) approach. It's not a competition between two fixed, independent entities. Rather, it's a constant, iterative process of self-improvement:

1. Initial State: The AI begins with only the rules of chess and a randomly initialized neural network. Its initial play is effectively random.
2. Self-Play and Learning: The AI plays millions of games against itself. For each move, it uses a Monte-Carlo Tree Search (MCTS) algorithm, guided by its current neural network, to explore the most promising variations. The MCTS builds a "game tree" and assesses the probability of winning from different positions.
3. Reinforcement Learning: After each game, the AI analyzes the outcome (win, loss, or draw). It then uses this information to update its neural network. A win strengthens the connections in the network that led to that outcome, and a loss weakens them.
4. Iteration: This process is repeated millions of times. The AI is constantly playing a new, slightly more refined version of itself. Each game is a tiny increment of learning, building on the knowledge gained from all previous games.

So, when AlphaZero plays itself, it's not a fight between two equal opponents. It's a continuous, one-sided "battle" against its own past self, where the goal is to find better moves and strategies to improve its win rate against its previous iterations.

The "Flatlining" of the Game

This leads to a fascinating outcome. As the AI becomes more and more perfect, it approaches what is called a perfect game. A perfect game is one where every move is optimal, and the outcome is predetermined.

• Chess Theory: In the case of chess, grandmasters have long theorized that with perfect play from both sides, the game would likely end in a draw. The vastness of the game tree makes it impossible for humans to prove this, but for an infinitely trained AI, this would be the theoretical outcome.
• The "Flatline": An AlphaZero trained to perfection would play against itself and every single game would end in a draw. No matter what move is made, the other side would always find the optimal response to ensure a draw. The game would become "flat" in the sense that there would be no winning or losing opportunities.

This is a key distinction from human play. For a human, every game is a struggle to find a winning opportunity against an imperfect opponent. For a perfect AI playing itself, the "struggle" is already won—it's about finding the single perfect path to a predetermined outcome.

The Corollary to Financial Markets

This concept has direct and profound implications for financial markets and the high-speed trading analogy you bring up.

• HFT and Co-location: High-Frequency Trading (HFT) and co-located servers were an earlier, more limited version of this phenomenon. HFT firms competed on speed, seeking to exploit tiny, fleeting inefficiencies (like arbitrage opportunities) milliseconds before anyone else. This was a "race to the bottom" on latency. The result was that these specific opportunities, while still profitable for the fastest players, became incredibly narrow and effectively "flatlined" for everyone else. The volume was limited because the strategies were dependent on speed, not on a fundamental re-evaluation of the market itself.
• Agentic AI: The Next Evolution: Agentic AI is an entirely different beast. It's not just about speed. It's about a fundamental re-evaluation of market dynamics. An agentic AI would be constantly learning and adapting. If all major players in a market (hedge funds, institutional investors, and even retail investors with sophisticated tools) deploy equally advanced agentic AIs, the "competition" would resemble AlphaZero's self-play.
• The "Flatlining" of Alpha: The AI agents would process all available information (news, social media, economic data, etc.) and discover profitable opportunities almost instantly. If all agents are equally capable and using similar reinforcement learning models, they would all arrive at the same conclusions simultaneously. The instant a profitable opportunity appears, it would be exploited by every agent, effectively eliminating the profit margin (the "alpha").
• The New Competition: The competition wouldn't be on finding a strategy, but on having the most subtle, nuanced, and adaptive AI. Who has a model that can predict the herd behavior of other AIs? Who can find a new, non-obvious data source (e.g., satellite imagery, weather patterns) that other AIs haven't incorporated yet? The battle would shift from an information race to a "meta-information" race—predicting the actions of the other agents.
• The Risk of Systemic Volatility: The biggest risk is not a flatline of the entire market, but the introduction of new forms of volatility. If all the AIs are trained on similar data and suddenly encounter an unforeseen event (a major political shock, a "fat finger" error), they might all sell at the same time. This could trigger a "flash crash" far more severe and rapid than anything seen before. The market would be a hyper-efficient system that could also be hyper-fragile.

In essence, an agentic AI ecosystem in financial markets is not a leveling of the playing field to an idyllic, perfectly efficient market. It's more likely a leveling of the playing field for all who can afford the best AI, leading to a new, more intense, and potentially more dangerous form of competition that will make the markets less predictable and more prone to extreme events.