The chess comparison has an instructive sequel. For about a decade after Deep Blue, a decent human paired with an engine beat the strongest engines playing alone. Kasparov called it centaur chess and thought it was the future. Then the engines kept improving until the human half of the centaur started subtracting value, and today overriding the machine almost always makes the move worse. In chess, the partnership window eventually closed.
Investing is the rare game where that window stays open, and the reason is in this essay. Chess never changes its rules, so a machine trained on the game's past faces the same game forever. Markets change their rules precisely because everyone's machine learns them. The moment a pattern is widely held it stops paying, which is the zero-sum point made here. So the human contribution shifts. Out-calculating the engine stopped being possible years ago. What remains is recognising when the regime has changed and the past has quietly become the wrong training set. Causal principles survive those moments. Patterns mined from history die at exactly the moment everyone owns them.
Writing the criteria down in advance does one more thing that deserves mention. It makes you auditable to yourself. A principle on paper before the event can be tested and scored. Reasoning assembled afterwards is just a story you tell about yourself.
Worth adding: the engine wins at solved chess because the human can only subtract from a finished game. Regime change is the inverted case, since each one is effectively a sample of one with no training set behind it. That gap is where the post's line about AI lacking "adequate insights to blindly follow" actually bites, and it's the one part of the job a machine can't practice its way into.
I would like to differentiate the understanding of risk and uncertainty in this context: risk is when circumstances repeat and you can learn the odds; uncertainty is when there's no real precedent at all. Your method handles the first beautifully "when analogous circumstances come along, you know what to do." But in my experience of 23 years in upstream oil and gas, the decisions that mattered most were the ones with no analogue. Frontier basins, sudden geopolitical ruptures. Nothing to back-test against. Written principles still helped, but not by telling me the move, they helped by making clear which assumptions I was betting on.
So, a gentle push on "timeless and universal": back-testing across every era and country still only samples the one history that actually happened. Which means the human in the partnership has two jobs, not one. Aligning with the system inside the world it knows and noticing the moment you've stepped outside that world. The system can't raise that flag itself; by construction it doesn't know where its world ends, especially when it is pulled in opposing direction by many forces and constraints. My worry with deep AI partnership is that the first skill gets stronger while the second quietly atrophies. The sentry at the edge still has to be human.
Your second job is the one no back-test can score, since you only confirm you'd left the known world after the fact. A distinction worth borrowing: divergence, edge, and inefficiency aren't synonyms. In a frontier basin, what reads as edge is often just divergence from a model that stopped applying, and the system will keep pricing it as edge until something breaks.
Coved 19 is the cleanest example of the decade. Chatgpt moment in 2022 can also be considered beyond the known risk. Climate debacle can through in unknown risks in future.
Essentially what Ray is saying is that we should use AI to dig deeper and branch out our knowledge into things that we are currently interested in while standing firm in the objective truth that nothing can truly replace what it means to be human, so we should consider everything and ultimately make the final decision, rather than letting AI have the last say.
The bit I find most useful is writing the criteria down before the situation arrives — 'if this, then that, because XYZ.' It turns a gut call into something you can argue with and refine, which the gut version never allows.
Where it seems to get hardest isn't writing the principle, it's the matching step: judging that the situation in front of you really belongs to the same set the principle was built on. You warn against data-mining the past — the close cousin is mis-judging which historical set the present actually fits, and pointing a sound cause-effect rule at the wrong analogue.
Curious how you handle that in practice: when a live situation only partly matches the history behind a principle, do you lean on the system's classification or override it with the human read?
Reading across 72 asset manager reports each month, the gap between managers who can articulate the cause-effect logic behind their positioning and those who can only describe what they own becomes most visible in exactly the environments you describe, where widely known information has already been priced and value-added requires a different level of reasoning. The consistent outperformers in our Awards dataset share one characteristic: their decision criteria are stable across regimes, not reactive to them.
The distinction between principled thinking and data mining is the one most AI-assisted investment processes haven't resolved. Training a model on what worked historically is not the same as encoding why it worked, and the difference matters precisely in the novel environments where the historical pattern breaks. Your point about back-testing timeless and universal criteria — rather than optimising for past returns, is the methodological discipline most systematic approaches quietly skip.
The principles that survive every environment are rarer than the returns they produce.
If I understand correctly, there is a tension underneath the process you describe.
On one side, we are observing reality: data, patterns, and cause-and-effect relationships.
On the other side, we are assigning meaning to what we observe so we can determine what matters and what action should follow.
Principles seem to sit between the two sides like a bridge. They translate an understanding of reality into a decision rule that can be applied repeatedly.
AI helps test whether our interpretations remain faithful to reality by challenging assumptions, identifying inconsistencies, and aggregating data. The human side contributes judgment by determining significance, relevance, and context.
The continual cycle of refinement between reality and judgment produces sound principles that keep the two aligned.
Using AI to do equity deep research on equity is just summarising different reports and data into a prose. Another good looking sell side report again.
If you want sth really assist you do to decision, you need to equip your AI a scientific methodology : Observe -> Guess -> Experiment -> Conclusions.
The chess comparison has an instructive sequel. For about a decade after Deep Blue, a decent human paired with an engine beat the strongest engines playing alone. Kasparov called it centaur chess and thought it was the future. Then the engines kept improving until the human half of the centaur started subtracting value, and today overriding the machine almost always makes the move worse. In chess, the partnership window eventually closed.
Investing is the rare game where that window stays open, and the reason is in this essay. Chess never changes its rules, so a machine trained on the game's past faces the same game forever. Markets change their rules precisely because everyone's machine learns them. The moment a pattern is widely held it stops paying, which is the zero-sum point made here. So the human contribution shifts. Out-calculating the engine stopped being possible years ago. What remains is recognising when the regime has changed and the past has quietly become the wrong training set. Causal principles survive those moments. Patterns mined from history die at exactly the moment everyone owns them.
Writing the criteria down in advance does one more thing that deserves mention. It makes you auditable to yourself. A principle on paper before the event can be tested and scored. Reasoning assembled afterwards is just a story you tell about yourself.
Worth adding: the engine wins at solved chess because the human can only subtract from a finished game. Regime change is the inverted case, since each one is effectively a sample of one with no training set behind it. That gap is where the post's line about AI lacking "adequate insights to blindly follow" actually bites, and it's the one part of the job a machine can't practice its way into.
I would like to differentiate the understanding of risk and uncertainty in this context: risk is when circumstances repeat and you can learn the odds; uncertainty is when there's no real precedent at all. Your method handles the first beautifully "when analogous circumstances come along, you know what to do." But in my experience of 23 years in upstream oil and gas, the decisions that mattered most were the ones with no analogue. Frontier basins, sudden geopolitical ruptures. Nothing to back-test against. Written principles still helped, but not by telling me the move, they helped by making clear which assumptions I was betting on.
So, a gentle push on "timeless and universal": back-testing across every era and country still only samples the one history that actually happened. Which means the human in the partnership has two jobs, not one. Aligning with the system inside the world it knows and noticing the moment you've stepped outside that world. The system can't raise that flag itself; by construction it doesn't know where its world ends, especially when it is pulled in opposing direction by many forces and constraints. My worry with deep AI partnership is that the first skill gets stronger while the second quietly atrophies. The sentry at the edge still has to be human.
Your second job is the one no back-test can score, since you only confirm you'd left the known world after the fact. A distinction worth borrowing: divergence, edge, and inefficiency aren't synonyms. In a frontier basin, what reads as edge is often just divergence from a model that stopped applying, and the system will keep pricing it as edge until something breaks.
Coved 19 is the cleanest example of the decade. Chatgpt moment in 2022 can also be considered beyond the known risk. Climate debacle can through in unknown risks in future.
Essentially what Ray is saying is that we should use AI to dig deeper and branch out our knowledge into things that we are currently interested in while standing firm in the objective truth that nothing can truly replace what it means to be human, so we should consider everything and ultimately make the final decision, rather than letting AI have the last say.
The bit I find most useful is writing the criteria down before the situation arrives — 'if this, then that, because XYZ.' It turns a gut call into something you can argue with and refine, which the gut version never allows.
Where it seems to get hardest isn't writing the principle, it's the matching step: judging that the situation in front of you really belongs to the same set the principle was built on. You warn against data-mining the past — the close cousin is mis-judging which historical set the present actually fits, and pointing a sound cause-effect rule at the wrong analogue.
Curious how you handle that in practice: when a live situation only partly matches the history behind a principle, do you lean on the system's classification or override it with the human read?
Reading across 72 asset manager reports each month, the gap between managers who can articulate the cause-effect logic behind their positioning and those who can only describe what they own becomes most visible in exactly the environments you describe, where widely known information has already been priced and value-added requires a different level of reasoning. The consistent outperformers in our Awards dataset share one characteristic: their decision criteria are stable across regimes, not reactive to them.
The distinction between principled thinking and data mining is the one most AI-assisted investment processes haven't resolved. Training a model on what worked historically is not the same as encoding why it worked, and the difference matters precisely in the novel environments where the historical pattern breaks. Your point about back-testing timeless and universal criteria — rather than optimising for past returns, is the methodological discipline most systematic approaches quietly skip.
The principles that survive every environment are rarer than the returns they produce.
If I understand correctly, there is a tension underneath the process you describe.
On one side, we are observing reality: data, patterns, and cause-and-effect relationships.
On the other side, we are assigning meaning to what we observe so we can determine what matters and what action should follow.
Principles seem to sit between the two sides like a bridge. They translate an understanding of reality into a decision rule that can be applied repeatedly.
AI helps test whether our interpretations remain faithful to reality by challenging assumptions, identifying inconsistencies, and aggregating data. The human side contributes judgment by determining significance, relevance, and context.
The continual cycle of refinement between reality and judgment produces sound principles that keep the two aligned.
Using AI to do equity deep research on equity is just summarising different reports and data into a prose. Another good looking sell side report again.
If you want sth really assist you do to decision, you need to equip your AI a scientific methodology : Observe -> Guess -> Experiment -> Conclusions.
Thts what I have built.
It’s true. And there is not enough education about this on the internet.
I agree, however can all people merge into AI thinking? Read my post: https://opherbrayer.substack.com/p/the-address-that-didnt-exist?r=o7uyc&utm_campaign=post-expanded-share&utm_medium=web