This week’s newsletter highlights a growing pattern across markets, physics, AI systems, and regulation: in global, always-on, feedback-driven trading environments, purely linear thinking no longer scales. At the same time, trading systems cannot rely on models that are overly adaptive or purely probabilistic, because execution and risk management demand determinism and predictability. The real question is not whether AI belongs in trading, but where intelligence should be applied and where strict controls must be enforced so systems behave reliably under extreme speed and stress.

1. 2026 and Quantum AI (QAI)
Lots this week on Quantum AI, including suggesting QAI could improve optimization and pattern recognition by handling larger datasets and more complex market structures. In practice the benefits are about faster training and exploration, not fundamentally better models or objectives; there is still no proven advantage over classical deep learning at real trading scale. Near-term value is more likely in areas such as reinforcement learning, scenario analysis, risk estimation, and stress testing, potentially through hybrid quantum–classical approaches. For now, QAI is best seen as an efficiency layer that reinforces a broader shift toward non-linear, feedback-driven, cross-asset system design, not as a source of immediate trading edge. Read more here – https://arxiv.org/html/2511.09884v1

2. A QAI systems perspective
Even without practical quantum computers, quantum-inspired thinking is already useful because it reframes markets as systems rather than pipelines. Just as particles in physics are defined by interactions rather than fixed properties– see more from Brian Cox here https://www.youtube.com/watch?v=BHEhxPuMmQI, markets are shaped by participant behaviour, feedback loops, and constraints.  For trading, this changes automation processes from first order signals (market data and news) to focus more on market reaction and impact from margin, collateral, liquidity, leverage, and risk limits. From this view, stress events – margin calls, liquidity freezes, volatility halts, forced deleveraging – could be viewed not simple failures but transition points where information is compressed into constraints and then the system reorganises. In always-on, globally connected, cross-asset markets, stability comes not from predicting regime shifts but from designing trading, risk, and infrastructure systems that assume change, monitor feedback, manage constraint build-up, and adapt quickly – the argument being that QAI may have a new future role to play.

3. Agents, Agentic Systems, and Autonomy
Current industry debate is the difference between Agents, Agentic and Autonomy where grey lines are emerging. While agents are valuable, autonomy must be tightly bounded – use of agents in workflows blurs the previous clear distinction. Agents improve productivity by handling research tasks such as data collection and analysis in parallel, delivering faster, broader, and more consistent insights, yet this has to be tightly managed to ensure at what point the agentic instruction leads to execution, or impacts risk. Recent real-world failures show agents still prone to drift, manipulation, and loss of context, especially under stress. In trading, this makes full autonomy hard to govern, reinforcing a non-negotiable principle: intelligence belongs in strategy, while determinism belongs in risk and control.

Research suggesting that advanced intelligence AGI may emerge from greater use of coordinated networks of agents rather than a single system aligns with this view, and quantum-inspired AI reinforces it by framing intelligence as something that emerges from interaction, feedback, and constraints. The practical conclusion is a shift toward supervised, distributed agentic systems, where agents explore and inform, with humans retaining accountability, and governance, predictability, and control as these systems scale – read more here https://arxiv.org/pdf/2512.16856.

4. QHFT at the speed-of-light
High-frequency trading today ultimately faces a physical limit: the speed of light. When trading venues are far apart, classical messages cannot arrive fast enough to fully coordinate decisions at microsecond timescales, so coordination – not compute – becomes the main constraint. Research on quantum correlations shows that entanglement cannot send information faster than light, but it can create correlations stronger than any classical shared randomness – read more here https://arxiv.org/html/2407.21723v1#:~:text=Quantum%20telepathy%20is%20not%20a,%2Dfrequency%20trading%20(HFT).

In very specific, pre-defined coordination problems caused by latency, this could slightly improve how distributed systems align their actions without waiting for messages. This does not improve price prediction, trading models, or general computation; any benefit is statistical, limited in size, and task-specific. Practical obstacles are still substantial, including entanglement distribution, noise, loss, synchronization, and hardware integration – but it does start to reframe some HFT latency problems as physics-limited rather than algorithm- or infrastructure-led.

5. WSJ Reality check with a PlayStation 5, bottles of wine, and a live betta fish
With all the optimism on Quantum this week, a Wall Street Journal test that gave an AI agent control of a vending machine quickly demonstrated the difference between theory and day to day reality.  The agent was given autonomy to manage inventory, set prices, order supplies, and interact with staff via Slack, with the goal of running the machine profitably. It spectacularly failed, losing over $1,000 in value by the time the experiment ended, underscoring how far autonomous agents still are from production reliability. The lesson for trading is simple: autonomy is easy to promise, hard to deliver and even harder to supervise. If agents struggle with basic instructions, they will struggle even more in adversarial electronic markets, making near-term claims that AI will fully run HFT unrealistic for now. Read more here – https://www.wsj.com/tech/ai/anthropic-claude-ai-vending-machine-agent-b7e84e34

One takeaway is that learning to improve agents will only come from trial and error but as desks experiment with ML-driven risk scoring, probabilistic systems introduce unacceptable levels of risk. Robust architectures have to keep the adaptive strategy layer separate from a deterministic and explainable risk layer. As models improve, the case for immutable risk gates will only become stronger, not weaker.

Finally a bonus gift point this week for those who celebrate Christmas – an article that claims “AI Is No Match for the Quirks of Human Intelligence” https://thereader.mitpress.mit.edu/ai-insight-problems-quirks-human-intelligence/ so we are still safe for now! The author argues that while AI excels at narrow, well-structured analytical tasks, human intelligence still matters precisely because it is messy, creative, inconsistent, and often irrational. It solves insight problems by reframing them rather than following step-by-step logic.

Even if large language models continue to improve, focusing only on more scale, more data, and tighter optimization risks missing the point. The biggest breakthroughs in AI did not come from everyone refining the same recipe, but from people pursuing diverse, often non-obvious ideas. Today, rising compute costs make exploration harder and convergence more likely, which is precisely why funding and encouraging diverse approaches matters more, not less.

For trading, the lesson is clear: progress will not come from optimization alone, but from exploration, diversity of thinking, and a LOT of accidental learning  – we just have to make sure we can also minimise the risk along the way.

Thank you for reading. Let me know what resonates, what doesn’t, and what you’d like to see more of next week for the final newsletter of 2025 – and for those of you who celebrate – I hope you have a very merry Christmas
Rebecca