I think you’re identifying something important here, but I’d make a distinction between the convergence of information and the convergence of institutional constraints.
The former doesn’t necessarily produce identical investment decisions. Two investors can analyse the same information and reach very different conclusions about intrinsic value.
The latter is potentially much more dangerous.
If thousands of managers have similar leverage, drawdown limits, investment horizons and risk-management systems, then even independently formed investment views can translate into highly correlated trading behaviour.
And there’s an interesting paradox here: what constitutes rational risk management at the individual portfolio level can create systemic instability when everyone responds to the same constraints simultaneously.
But I’d challenge the conclusion that fundamental research becomes irrelevant.
If a company loses 30% of its market value because the marginal shareholder is a forced seller, without a corresponding deterioration in its future cash flows, that doesn’t invalidate fundamental research. It potentially creates a much larger fundamental mispricing.
The distinction, I think, is between being right about intrinsic value and having the capital structure necessary to monetise that insight.
And that’s why I’d argue duration alone isn’t an edge. What matters is independence: independence of thought, independence of positioning, and perhaps most importantly, independence from the constraints forcing everyone else to transact.
A five-year investment thesis supported by capital that cannot tolerate a three-month drawdown isn’t really a five-year strategy.
I also wonder whether the proliferation of AI-powered research makes this problem more interesting. If information becomes increasingly commoditised while institutional constraints remain homogeneous, perhaps the real opportunity isn’t discovering what everyone else has missed, but understanding why market participants are being compelled to act irrespective of what they actually believe.
In that world, fundamental research doesn’t become obsolete. The ability to distinguish genuine changes in economic value from mechanically induced changes in market prices potentially becomes even more valuable.
The scarce resource might not be information or even analytical sophistication, but the freedom to act independently of everyone else’s constraints.
The former doesn’t necessarily produce identical investment decisions. Two investors can analyse the same information and reach very different conclusions about intrinsic value.
The latter is potentially much more dangerous.
If thousands of managers have similar leverage, drawdown limits, investment horizons and risk-management systems, then even independently formed investment views can translate into highly correlated trading behaviour.
And there’s an interesting paradox here: what constitutes rational risk management at the individual portfolio level can create systemic instability when everyone responds to the same constraints simultaneously.
But I’d challenge the conclusion that fundamental research becomes irrelevant.
If a company loses 30% of its market value because the marginal shareholder is a forced seller, without a corresponding deterioration in its future cash flows, that doesn’t invalidate fundamental research. It potentially creates a much larger fundamental mispricing.
The distinction, I think, is between being right about intrinsic value and having the capital structure necessary to monetise that insight.
And that’s why I’d argue duration alone isn’t an edge. What matters is independence: independence of thought, independence of positioning, and perhaps most importantly, independence from the constraints forcing everyone else to transact.
A five-year investment thesis supported by capital that cannot tolerate a three-month drawdown isn’t really a five-year strategy.
I also wonder whether the proliferation of AI-powered research makes this problem more interesting. If information becomes increasingly commoditised while institutional constraints remain homogeneous, perhaps the real opportunity isn’t discovering what everyone else has missed, but understanding why market participants are being compelled to act irrespective of what they actually believe.
In that world, fundamental research doesn’t become obsolete. The ability to distinguish genuine changes in economic value from mechanically induced changes in market prices potentially becomes even more valuable.
The scarce resource might not be information or even analytical sophistication, but the freedom to act independently of everyone else’s constraints.
A working theory I have is that crowding and contagion in public markets, particularly with MM funds, is basically unavoidable, right?
You have a cohort of senior analysts / PMs who came through the same training programs, leverage the same alt data, have similar tight risk constraints, and are trading similar ~quarterly time horizons. Oh, and now folks are likely using the same LLMs as part of their research. If you get fired, another pod picks you up, you carry your process there, rinse, repeat.
Same process, parameters, participants, and risk constraints across many levered shops. Correlation seems like it has to approach 1.
And if you’re effectively playing quarters, you’re forced into momentum trades to make money, because it fits your time horizon.
With tight risk constaints, if your pod blows up, you’re a forced seller, causing a cascade of other forced sellers who were likely positioned similarly, etc.
The simple answer is “duration is my edge” but if your “duration” still has you holding the same positions as the pods who are blowing out, you still suffer a 30% drawdown (see Lone Pine).
Fundamental research doesn’t matter if your position is hit by a tsunami of forced sellers when narrative (or a 2nd derivative data point) shifts. You still get smoked, LPs could pull capital, PMs lose bonuses, etc.
If anything, the value of “duration” is that you can sell or avoid overheated stuff that pods are forced to trade and hold bc they have a short time horizon and “that’s what’s working.” You can sell into euphoria / momentum and capitalize on blowouts when they happen. We’ve seen this cycle multiple times this year.
If market structures cause whiplash, even “fundamental” investors have to be aware of these high correlation moves and trade accordingly. “We can hold through this blowout and trust the long-term value of the stock” is cope.
0
1
5
1,613
I think David raises an excellent question.
At some point, the persistence of outperformance should meaningfully change our assessment of whether a manager possesses genuine skill.
The challenge is that almost any successful track record can be explained retrospectively through factor exposures, favourable regimes or randomness.
But that explanation needs to be falsifiable too. Otherwise, we’re effectively setting an impossible standard for demonstrating skill.
I suspect the answer isn’t a fixed number of bets or years, but how much the observed evidence should update our prior beliefs about the manager’s ability.
The independence of the bets, the economic rationale behind the strategy, the returns unexplained by factor exposures and, importantly, the performance outside the original sample all matter.
Ultimately, the question should be symmetrical: what evidence would convince us a manager has skill, and what evidence would convince us they don’t?
At some point, the persistence of outperformance should meaningfully change our assessment of whether a manager possesses genuine skill.
The challenge is that almost any successful track record can be explained retrospectively through factor exposures, favourable regimes or randomness.
But that explanation needs to be falsifiable too. Otherwise, we’re effectively setting an impossible standard for demonstrating skill.
I suspect the answer isn’t a fixed number of bets or years, but how much the observed evidence should update our prior beliefs about the manager’s ability.
The independence of the bets, the economic rationale behind the strategy, the returns unexplained by factor exposures and, importantly, the performance outside the original sample all matter.
Ultimately, the question should be symmetrical: what evidence would convince us a manager has skill, and what evidence would convince us they don’t?
"He's just been lucky on factor bets."
How many individual bets, and after how many years, does it become unreasonable to keep thinking this?
1
0
7
1,259
If AlphaSense is becoming common research infrastructure across the investment industry, the interesting question isn’t simply whether everyone will have access to the same information. It’s whether we’re confusing the ability to retrieve information with the ability to reason independently about it.
Those are fundamentally different things.
Consider two investors analysing exactly the same company. Both have access to identical earnings transcripts, financial statements, consensus estimates and expert interviews. Yet one might see a structurally deteriorating business, while the other sees a temporary earnings dislocation and a compelling investment opportunity.
The information is identical. The underlying models of how the business operates are not.
And this is where I think AlphaSense could become particularly interesting.
I’ve been thinking about the potential for what I’d describe as an interconnected investment brain. Not another AI research assistant that retrieves documents, summarises management commentary or produces financial models, but a system that builds and continuously updates an interconnected understanding of businesses, industries, competitive dynamics and market expectations.
Imagine a system where every piece of evidence is connected to the hypotheses it supports or contradicts. Those hypotheses are connected to the underlying economic drivers of a business. Those drivers feed into probabilistic forecasts, which in turn influence valuations and investment decisions.
Now extend that architecture across thousands of companies.
A change in consumer behaviour doesn’t simply trigger an alert about potentially affected businesses. The system can trace the implications across categories, competitors, suppliers, operating margins, capital allocation and ultimately valuations.
Importantly, it should recognise that these relationships are uncertain, test competing causal explanations and learn from previous forecasting errors.
Perhaps most importantly, it should constantly evaluate its conclusions against what the market is already discounting.
Because generating a more accurate forecast isn’t necessarily sufficient to generate alpha. You need to understand how that forecast differs from expectations embedded in the share price.
This is where the distinction between research automation and investment intelligence becomes meaningful.
The former helps everyone reach an answer faster.
The latter could help investors discover relationships, challenge assumptions, identify contradictions and evaluate possible outcomes they might otherwise have overlooked.
And here’s the paradox: AlphaSense’s enormous content library, financial datasets and increasingly interconnected research infrastructure could make it exceptionally well positioned to facilitate precisely this transition.
Of course, simply connecting more information doesn’t guarantee differentiated conclusions. If every investor uses the same reasoning architecture, the same assumptions and the same optimisation objectives, we may simply recreate intellectual conformity at a much higher level of sophistication.
The solution, therefore, isn’t necessarily preserving human judgment as the final source of differentiation. AI systems themselves could develop increasingly sophisticated, independent approaches to hypothesis generation, causal inference and probabilistic forecasting.
What matters is whether the architecture encourages genuine exploration and falsification rather than simply reproducing the dominant interpretation.
Ultimately, I suspect we’re moving towards a world where information retrieval becomes increasingly commoditised, while the ability to construct, test and continuously refine an accurate model of the world becomes the real competitive advantage.
AlphaSense could be much more than common research infrastructure.
Those are fundamentally different things.
Consider two investors analysing exactly the same company. Both have access to identical earnings transcripts, financial statements, consensus estimates and expert interviews. Yet one might see a structurally deteriorating business, while the other sees a temporary earnings dislocation and a compelling investment opportunity.
The information is identical. The underlying models of how the business operates are not.
And this is where I think AlphaSense could become particularly interesting.
I’ve been thinking about the potential for what I’d describe as an interconnected investment brain. Not another AI research assistant that retrieves documents, summarises management commentary or produces financial models, but a system that builds and continuously updates an interconnected understanding of businesses, industries, competitive dynamics and market expectations.
Imagine a system where every piece of evidence is connected to the hypotheses it supports or contradicts. Those hypotheses are connected to the underlying economic drivers of a business. Those drivers feed into probabilistic forecasts, which in turn influence valuations and investment decisions.
Now extend that architecture across thousands of companies.
A change in consumer behaviour doesn’t simply trigger an alert about potentially affected businesses. The system can trace the implications across categories, competitors, suppliers, operating margins, capital allocation and ultimately valuations.
Importantly, it should recognise that these relationships are uncertain, test competing causal explanations and learn from previous forecasting errors.
Perhaps most importantly, it should constantly evaluate its conclusions against what the market is already discounting.
Because generating a more accurate forecast isn’t necessarily sufficient to generate alpha. You need to understand how that forecast differs from expectations embedded in the share price.
This is where the distinction between research automation and investment intelligence becomes meaningful.
The former helps everyone reach an answer faster.
The latter could help investors discover relationships, challenge assumptions, identify contradictions and evaluate possible outcomes they might otherwise have overlooked.
And here’s the paradox: AlphaSense’s enormous content library, financial datasets and increasingly interconnected research infrastructure could make it exceptionally well positioned to facilitate precisely this transition.
Of course, simply connecting more information doesn’t guarantee differentiated conclusions. If every investor uses the same reasoning architecture, the same assumptions and the same optimisation objectives, we may simply recreate intellectual conformity at a much higher level of sophistication.
The solution, therefore, isn’t necessarily preserving human judgment as the final source of differentiation. AI systems themselves could develop increasingly sophisticated, independent approaches to hypothesis generation, causal inference and probabilistic forecasting.
What matters is whether the architecture encourages genuine exploration and falsification rather than simply reproducing the dominant interpretation.
Ultimately, I suspect we’re moving towards a world where information retrieval becomes increasingly commoditised, while the ability to construct, test and continuously refine an accurate model of the world becomes the real competitive advantage.
AlphaSense could be much more than common research infrastructure.
I think AlphaSense could be an important part of the problem. But I’d make a distinction: AlphaSense isn’t necessarily producing bad research. It may be helping the entire investment industry arrive at the same research conclusions.
That’s a much more interesting criticism.
Think about what happens when a research platform becomes standard equipment across hedge funds, asset managers, investment banks and corporate research teams.
Everyone searches the same transcripts.
Everyone reads the same broker research.
Everyone searches for the same management comments.
Everyone asks AI to identify the same risks and catalysts.
Everyone gets increasingly similar summaries and comparisons.
The platform becomes an information equalizer.
But an information equalizer can also become a source of intellectual conformity.
6. Here’s why AlphaSense is particularly interesting
The scale is substantial.
According to an April 2026 Sacra company profile:
Approximately 75% of the top hedge funds use AlphaSense.
Approximately 80% of top asset managers use it.
7,000 enterprise customers.
It serves major financial institutions and corporate research teams. sacra-pdfs.s3.us-east-2.amazonaws.com
AlphaSense itself advertises a library of more than 500 million business documents and broker research from over 1,700 providers. AlphaSense
That creates an interesting paradox.
The same platform that helps investors find differentiated insights is becoming common infrastructure across the investment industry.
There are three potential problems:
Consensus risk: Everyone can quickly identify what the Street thinks.
Crowding risk: Everyone can discover the same catalysts and risks.
Reaction risk: When the thesis changes, many investors may react to the same information simultaneously.
One qualification: shared access does not mean identical portfolios. Firms have proprietary research, different mandates, different models and different risk limits. AlphaSense also maintains permission boundaries around internal customer information.
But the common research layer is worth examining.
The question I would put to the hedge fund industry
If 75% of major hedge funds are using the same research infrastructure, where exactly is their information advantage?
And if everyone has access to the same research, perhaps the scarce resource is no longer information.
It is independent judgment.
My proposed thesis:
The investment industry may be approaching a point where better research technology is producing diminishing informational advantages and increasing the risk of crowded decisions.
That is not an indictment of AlphaSense as a company. It is a question about what happens when an entire industry standardizes its intelligence infrastructure.
And I think that’s a much more important discussion than whether AI can replace a junior analyst.
2
7
122
11,937
I think there’s a fundamental assumption here worth challenging: why should machine judgment resemble human judgment in the first place?
Human intuition is, in many respects, an evolutionary workaround for severe cognitive constraints. Limited working memory, finite attention, emotional attachment, confirmation bias and an inability to consistently reason probabilistically.
We seem to assume that the pinnacle of investment decision-making is the judgment of an exceptional human investor. But why should that be the ceiling?
A sufficiently capable AI could simultaneously evaluate thousands of competing hypotheses, continuously update probabilistic beliefs, identify causal relationships and dynamically allocate attention toward information with the highest expected value.
Crucially, it need not inherit the same cognitive limitations that make human judgment so fallible. Of course, machines can develop their own biases, but those need not resemble ours.
And today’s hedge funds are hardly a definitive test of machine intelligence. Most still use sophisticated technology within fundamentally human-designed investment frameworks.
Perhaps we’re asking the wrong question.
Instead of asking whether AI can replicate the intuition of a brilliant portfolio manager, we should ask whether that intuition is simply a computational shortcut developed by an intelligence with severely limited processing capacity.
The real opportunity isn’t to automate human judgment. It’s to discover forms of judgment that human cognition could never achieve.
Human intuition is, in many respects, an evolutionary workaround for severe cognitive constraints. Limited working memory, finite attention, emotional attachment, confirmation bias and an inability to consistently reason probabilistically.
We seem to assume that the pinnacle of investment decision-making is the judgment of an exceptional human investor. But why should that be the ceiling?
A sufficiently capable AI could simultaneously evaluate thousands of competing hypotheses, continuously update probabilistic beliefs, identify causal relationships and dynamically allocate attention toward information with the highest expected value.
Crucially, it need not inherit the same cognitive limitations that make human judgment so fallible. Of course, machines can develop their own biases, but those need not resemble ours.
And today’s hedge funds are hardly a definitive test of machine intelligence. Most still use sophisticated technology within fundamentally human-designed investment frameworks.
Perhaps we’re asking the wrong question.
Instead of asking whether AI can replicate the intuition of a brilliant portfolio manager, we should ask whether that intuition is simply a computational shortcut developed by an intelligence with severely limited processing capacity.
The real opportunity isn’t to automate human judgment. It’s to discover forms of judgment that human cognition could never achieve.
Think about how sophisticated today’s hedge funds have become.
Billions of dollars under management.
PhDs in mathematics, physics and computer science.
Massive data science teams.
Thousands of data feeds.
Alternative data.
Satellite imagery.
Machine learning.
Artificial intelligence.
High-frequency trading.
Thousands of GPUs.
Complex risk-management systems.
They can monitor markets, companies, supply chains and investor behavior at a level that was almost unimaginable 20 years ago.
And yet, some of the most sophisticated investment organizations in the world just suffered enormous losses.
That should make everyone stop and think.
If this level of technology, data and intellectual firepower isn’t enough to consistently produce good investment decisions, what exactly are we building?
The next generation will have even more data, faster computers, more powerful AI agents and increasingly complex models.
But will they have better judgment?
Or will we simply create bigger, faster and more sophisticated versions of the same mistakes?
The future of investing may not belong to whoever has the most data or the most computing power. It may belong to whoever knows what information to ignore.
Technology can process information.
It cannot guarantee wisdom.
3
1
20
1,640
I think you’re missing a fairly fundamental point about portfolio construction.
The S&P 500 is not the universal benchmark against which every investment strategy should be measured.
If you’re evaluating a long-only US large-cap equity manager, absolutely. If they’re charging substantial fees to deliver something resembling the index, they deserve to be challenged.
But hedge funds encompass an entirely different range of strategies. Market-neutral equity, relative value, merger arbitrage and global macro are not necessarily designed to outperform the S&P 500 in absolute terms. Their objective may instead be to generate returns that are relatively independent of broader equity markets.
And that’s an important distinction.
Consider a strategy generating 12% annual returns with substantial equity beta, significant volatility and large drawdowns. Now compare that with one generating 8%, but with considerably lower volatility and almost no correlation to equities.
The latter might have underperformed the S&P 500 in absolute terms, but that doesn’t necessarily make it an inferior investment. Incorporated into a broader portfolio, it could improve risk-adjusted returns and meaningfully reduce overall drawdowns.
That’s the point of diversification. You’re not simply trying to maximise the return of every individual investment. You’re trying to optimise the risk and return characteristics of the portfolio as a whole.
There’s also an element of recency bias in assuming the S&P 500 will continue delivering the returns investors have become accustomed to.
US equities have enjoyed an extraordinary period of performance since the global financial crisis, supported by expanding valuations, strong earnings growth and increasing market concentration. But investors seem remarkably quick to forget that the S&P 500 delivered negative annualised total returns between 2000 and 2009.
Extrapolating the experience of the past 15–20 years into the next 20 is hardly a sound investment thesis.
The more interesting questions are whether a manager can generate genuine alpha after fees, how much of their performance is simply compensation for systematic risk exposures, and whether incorporating their strategy improves the characteristics of the overall portfolio.
Of course, hedge funds charging excessive fees for repackaged equity beta deserve criticism. And many have failed to deliver the diversification or alpha they promise.
But using the S&P 500 as the universal yardstick for every investment strategy fundamentally misunderstands what institutional portfolio construction is trying to achieve.
The objective isn’t necessarily to beat the S&P 500. It’s to maximise portfolio returns relative to the risks you’re willing to bear.
Those are two very different things.
The S&P 500 is not the universal benchmark against which every investment strategy should be measured.
If you’re evaluating a long-only US large-cap equity manager, absolutely. If they’re charging substantial fees to deliver something resembling the index, they deserve to be challenged.
But hedge funds encompass an entirely different range of strategies. Market-neutral equity, relative value, merger arbitrage and global macro are not necessarily designed to outperform the S&P 500 in absolute terms. Their objective may instead be to generate returns that are relatively independent of broader equity markets.
And that’s an important distinction.
Consider a strategy generating 12% annual returns with substantial equity beta, significant volatility and large drawdowns. Now compare that with one generating 8%, but with considerably lower volatility and almost no correlation to equities.
The latter might have underperformed the S&P 500 in absolute terms, but that doesn’t necessarily make it an inferior investment. Incorporated into a broader portfolio, it could improve risk-adjusted returns and meaningfully reduce overall drawdowns.
That’s the point of diversification. You’re not simply trying to maximise the return of every individual investment. You’re trying to optimise the risk and return characteristics of the portfolio as a whole.
There’s also an element of recency bias in assuming the S&P 500 will continue delivering the returns investors have become accustomed to.
US equities have enjoyed an extraordinary period of performance since the global financial crisis, supported by expanding valuations, strong earnings growth and increasing market concentration. But investors seem remarkably quick to forget that the S&P 500 delivered negative annualised total returns between 2000 and 2009.
Extrapolating the experience of the past 15–20 years into the next 20 is hardly a sound investment thesis.
The more interesting questions are whether a manager can generate genuine alpha after fees, how much of their performance is simply compensation for systematic risk exposures, and whether incorporating their strategy improves the characteristics of the overall portfolio.
Of course, hedge funds charging excessive fees for repackaged equity beta deserve criticism. And many have failed to deliver the diversification or alpha they promise.
But using the S&P 500 as the universal yardstick for every investment strategy fundamentally misunderstands what institutional portfolio construction is trying to achieve.
The objective isn’t necessarily to beat the S&P 500. It’s to maximise portfolio returns relative to the risks you’re willing to bear.
Those are two very different things.
@uncovered_alpha Bro. Who is going to beat the S&P over the next 20 years and if you can’t answer you are paying fees to underperform with more risk.
1
0
11
2,104
This is a fascinating observation @futurewealthP, and I think you’ve identified perhaps the most important challenge in what I’m envisioning.
Connecting thousands of businesses, economic variables, hypotheses and potential causal relationships creates a search space that becomes computationally intractable remarkably quickly.
But I think your point about intuition is particularly interesting.
Perhaps much of what we call human judgement is actually a highly developed mechanism for allocating attention under conditions of extreme uncertainty.
We don’t consciously investigate every possible explanation for an observed outcome. Our intuition directs us towards a small subset of possibilities that seem worth exploring.
In other words, human intelligence doesn’t solve the combinatorial explosion problem through exhaustive reasoning. It solves it, imperfectly, through selective attention.
And perhaps that’s precisely the next frontier for autonomous research systems.
Imagine an interconnected investment knowledge graph where the system doesn’t attempt to investigate every conceivable relationship, but instead continuously estimates where additional research is most likely to improve its understanding or change an investment decision.
It could prioritise hypotheses according to their expected information value, economic materiality, uncertainty, existing evidence and computational cost.
But here’s the part I find particularly exciting.
What if the system could learn not only from the outcomes of its predictions, but from the success or failure of its decisions about where to investigate in the first place?
Then you’re no longer simply building a system that discovers relationships.
You’re building one that learns how to discover relationships more efficiently.
Its research process becomes endogenous and adaptive. An unexpected observation in one industry could trigger an investigation elsewhere, not because someone explicitly programmed that relationship, but because the system has learned that certain patterns of evidence warrant further attention.
Of course, this doesn’t eliminate the search-space problem. It introduces another difficult problem: how to distinguish genuinely promising research directions from attractive but ultimately spurious patterns.
That requires careful treatment of uncertainty, exploration versus exploitation, and the cost of pursuing the wrong hypotheses.
But it does suggest a very different architecture from simply deploying increasingly powerful agents against existing research workflows.
And I think your final philosophical observation is equally important.
If we eventually develop systems that are superior at determining how to achieve a particular objective, then deciding which objectives are worth pursuing becomes increasingly consequential.
In investing, we often treat maximising risk-adjusted returns as an obvious objective. But even that conceals questions about acceptable risk, time horizons, liquidity, externalities and what we’re ultimately trying to achieve with capital.
Optimisation is not the same thing as choosing what is worth optimising.
Perhaps the really interesting transition isn’t simply from human judgement to machine judgement, but from humans being primarily responsible for finding answers towards being increasingly responsible for deciding which questions and outcomes are worth pursuing.
I suspect we’re only beginning to appreciate the implications of that distinction.
Connecting thousands of businesses, economic variables, hypotheses and potential causal relationships creates a search space that becomes computationally intractable remarkably quickly.
But I think your point about intuition is particularly interesting.
Perhaps much of what we call human judgement is actually a highly developed mechanism for allocating attention under conditions of extreme uncertainty.
We don’t consciously investigate every possible explanation for an observed outcome. Our intuition directs us towards a small subset of possibilities that seem worth exploring.
In other words, human intelligence doesn’t solve the combinatorial explosion problem through exhaustive reasoning. It solves it, imperfectly, through selective attention.
And perhaps that’s precisely the next frontier for autonomous research systems.
Imagine an interconnected investment knowledge graph where the system doesn’t attempt to investigate every conceivable relationship, but instead continuously estimates where additional research is most likely to improve its understanding or change an investment decision.
It could prioritise hypotheses according to their expected information value, economic materiality, uncertainty, existing evidence and computational cost.
But here’s the part I find particularly exciting.
What if the system could learn not only from the outcomes of its predictions, but from the success or failure of its decisions about where to investigate in the first place?
Then you’re no longer simply building a system that discovers relationships.
You’re building one that learns how to discover relationships more efficiently.
Its research process becomes endogenous and adaptive. An unexpected observation in one industry could trigger an investigation elsewhere, not because someone explicitly programmed that relationship, but because the system has learned that certain patterns of evidence warrant further attention.
Of course, this doesn’t eliminate the search-space problem. It introduces another difficult problem: how to distinguish genuinely promising research directions from attractive but ultimately spurious patterns.
That requires careful treatment of uncertainty, exploration versus exploitation, and the cost of pursuing the wrong hypotheses.
But it does suggest a very different architecture from simply deploying increasingly powerful agents against existing research workflows.
And I think your final philosophical observation is equally important.
If we eventually develop systems that are superior at determining how to achieve a particular objective, then deciding which objectives are worth pursuing becomes increasingly consequential.
In investing, we often treat maximising risk-adjusted returns as an obvious objective. But even that conceals questions about acceptable risk, time horizons, liquidity, externalities and what we’re ultimately trying to achieve with capital.
Optimisation is not the same thing as choosing what is worth optimising.
Perhaps the really interesting transition isn’t simply from human judgement to machine judgement, but from humans being primarily responsible for finding answers towards being increasingly responsible for deciding which questions and outcomes are worth pursuing.
I suspect we’re only beginning to appreciate the implications of that distinction.
@uncovered_alpha@uncovered_alpha @uncovered_alphaus obstacle I see is combinatorial explosion: even for a super super super computer, the search space will be too large to search without some sort of choosing-where-to-search
Incidentally, this is what I think humans do with our “feelings” or “intuitions”)…these are our indicators of where to jump into the unfathomably large possibility space…This is all - or most - of what I think human judgement is - where to focus attention in the search space and our emotional/valence/hedonic reactions are the pointers or metal detectors if you will of this…what would this look like in AI…I think we are already seeing it…
Im no expert, but once AI agents approach models where they can have a model of where to direct their attention- and learn how to do so based on their vastly superior working memory and analytical resources, its game over for human judgment in terms of making “better”’decisions in arena’s where we aim for the same goal (better risk adjusted investment returns for example)…the most important is that we remain the masters of choosing values - that they are OUR values…
I think society is entering a phase where what used to be rarefied philosophical questions (what is valuable, what is meaning, what is good, etc)…will be much more important for many to work through, for many have built their identity and lives on their differential intellectual or analytical ability…we are entering the age of values and qualities imo (in short: in the trinity of Truth, Beauty, and Goodness, Beauty and Goodness is about to get a much needed positive re-valuation)….
1
3
6
1,019
The industry has a remarkable tendency to attribute success to skill and failure to incompetence, when the reality is often far more nuanced.
That doesn’t excuse poor risk management, excessive leverage or structural flaws in a strategy.
But there’s a difference between critically examining why a fund failed and taking pleasure in watching people’s careers and livelihoods unravel.
Some of the loudest critics might benefit from asking how their own strategies would perform under a genuinely hostile market regime.
A little intellectual humility goes a long way in this business.
That doesn’t excuse poor risk management, excessive leverage or structural flaws in a strategy.
But there’s a difference between critically examining why a fund failed and taking pleasure in watching people’s careers and livelihoods unravel.
Some of the loudest critics might benefit from asking how their own strategies would perform under a genuinely hostile market regime.
A little intellectual humility goes a long way in this business.
Can somebody explain to me why people love to grave dance when hfs or pods blow up?
Anybody that’s done the job knows that you tow the line between skill and luck more often than you’d like
And this could happen to any one of us unless your arrogance has been juiced beyond return
Have some humility…
2
1
20
2,587
Agree. I know quite a few people at Walleye, and Samlyn Capital for that matter. Tough to see what’s happening at both firms. October was certainly brutal for parts of the industry.
But I think there’s a more interesting question here than whether the pod model continues taking share.
What happens to market structure when an increasingly large proportion of active capital operates under similar risk constraints?
Traditional fundamental investing is principally concerned with the relationship between price and intrinsic value. Multi-manager platforms introduce another variable: the relationship between expected returns and the amount of risk capital required to capture them.
And those objectives don’t always align.
As more capital migrates towards platforms with tightly controlled drawdowns, exposure limits and relatively short performance measurement windows, you potentially create a market increasingly sensitive to positioning and flows rather than fundamentals.
The irony is that individually rational risk-management decisions can become collectively destabilising.
One PM reducing exposure after a drawdown is prudent. Hundreds of PMs doing so simultaneously, particularly where their portfolios share common factor exposures, can create price dislocations entirely disconnected from intrinsic value.
Which raises an interesting possibility: could the continued institutionalisation of the pod model actually increase the opportunity set for patient fundamental investors?
If the marginal investor is increasingly constrained by short-term volatility, the investor capable of underwriting longer-duration uncertainty without facing forced liquidation potentially possesses a structural advantage.
Of course, that advantage is only meaningful if you can survive the volatility yourself.
Perhaps the real distinction isn’t pods versus single managers, but investors whose time horizons are dictated by their investment theses versus those whose horizons are dictated by their risk budgets.
And if the latter increasingly dominate price discovery, the former may ultimately have more opportunities to exploit.
A fascinating paradox if the very model taking market share ends up creating the inefficiencies that others monetise.
But I think there’s a more interesting question here than whether the pod model continues taking share.
What happens to market structure when an increasingly large proportion of active capital operates under similar risk constraints?
Traditional fundamental investing is principally concerned with the relationship between price and intrinsic value. Multi-manager platforms introduce another variable: the relationship between expected returns and the amount of risk capital required to capture them.
And those objectives don’t always align.
As more capital migrates towards platforms with tightly controlled drawdowns, exposure limits and relatively short performance measurement windows, you potentially create a market increasingly sensitive to positioning and flows rather than fundamentals.
The irony is that individually rational risk-management decisions can become collectively destabilising.
One PM reducing exposure after a drawdown is prudent. Hundreds of PMs doing so simultaneously, particularly where their portfolios share common factor exposures, can create price dislocations entirely disconnected from intrinsic value.
Which raises an interesting possibility: could the continued institutionalisation of the pod model actually increase the opportunity set for patient fundamental investors?
If the marginal investor is increasingly constrained by short-term volatility, the investor capable of underwriting longer-duration uncertainty without facing forced liquidation potentially possesses a structural advantage.
Of course, that advantage is only meaningful if you can survive the volatility yourself.
Perhaps the real distinction isn’t pods versus single managers, but investors whose time horizons are dictated by their investment theses versus those whose horizons are dictated by their risk budgets.
And if the latter increasingly dominate price discovery, the former may ultimately have more opportunities to exploit.
A fascinating paradox if the very model taking market share ends up creating the inefficiencies that others monetise.
Tough to see what happened at Walleye. I know alot of people there. But I think it highlights my continued belief the pods will continue to take share. Fund is down 7 pct for the year and they are shutting down an entire division. Not great. Hard business. If you are a $30bn university endowment that needs to grow mid singles a year to fund your annual budget, down 7 is a lot better than whatever is happening with the big high profile single managers right now. A down 20-30 year is unacceptable. Citadel up in October and up mid teens for the year is incredible x.com/ThinkingUSD/st…
10
10
173
35,938
I think Al is pointing to something far more profound than most people in investment management appreciate.
Ultimately, an investment system needs to understand three things:
1. What is likely to happen?
2. What does the market currently believe will happen?
3. Where is the difference sufficiently large, and sufficiently mispriced, to generate attractive risk-adjusted returns?
That third question is where predictive intelligence becomes investment intelligence.
And it introduces another challenge: distinguishing genuine signal from relationships that merely look convincing in hindsight.
Any system claiming to learn autonomously needs rigorous point-in-time data, out-of-sample testing, calibrated uncertainty and a mechanism for recognising when previously successful relationships have stopped working.
Otherwise, we’re simply industrialising overfitting.
But imagine getting those foundations right.
This raises an uncomfortable question for traditional fundamental investment managers.
If an autonomous system can investigate more hypotheses, assimilate more information, identify more cross-company relationships and systematically learn from its mistakes, where exactly does the sustainable advantage of the traditional research organisation reside?
Human judgement will remain valuable, particularly in setting objectives, understanding incentives and challenging poorly specified assumptions.
But relying on the supposed uniqueness of human judgement feels increasingly like a dangerous competitive strategy.
I think the biggest mistake would be to view this transition as another technology upgrade.
It may instead require fundamentally rethinking how investment research is organised, how knowledge is accumulated, and ultimately how alpha is discovered.
The future of fundamental investing may not be about building a better analyst.
It may be about building an entirely different kind of intelligence.
Ultimately, an investment system needs to understand three things:
1. What is likely to happen?
2. What does the market currently believe will happen?
3. Where is the difference sufficiently large, and sufficiently mispriced, to generate attractive risk-adjusted returns?
That third question is where predictive intelligence becomes investment intelligence.
And it introduces another challenge: distinguishing genuine signal from relationships that merely look convincing in hindsight.
Any system claiming to learn autonomously needs rigorous point-in-time data, out-of-sample testing, calibrated uncertainty and a mechanism for recognising when previously successful relationships have stopped working.
Otherwise, we’re simply industrialising overfitting.
But imagine getting those foundations right.
This raises an uncomfortable question for traditional fundamental investment managers.
If an autonomous system can investigate more hypotheses, assimilate more information, identify more cross-company relationships and systematically learn from its mistakes, where exactly does the sustainable advantage of the traditional research organisation reside?
Human judgement will remain valuable, particularly in setting objectives, understanding incentives and challenging poorly specified assumptions.
But relying on the supposed uniqueness of human judgement feels increasingly like a dangerous competitive strategy.
I think the biggest mistake would be to view this transition as another technology upgrade.
It may instead require fundamentally rethinking how investment research is organised, how knowledge is accumulated, and ultimately how alpha is discovered.
The future of fundamental investing may not be about building a better analyst.
It may be about building an entirely different kind of intelligence.
People are missing the point here
Calling this quant, systematic, quantamental or whatever else doesn’t explain what’s happening.
The frontier has moved from helping analysts do their existing process faster to the most ambitious firms pushing towards autonomous, end-to-end fundamental research
These are systems that decide what to investigate, gather evidence, test competing explanations, build forecasts and challenge their own conclusions. Then work out what to investigate next.
to give context, we've shown LLMs can successfully identify genuine regime shifts before shares react
they can also find and map alternative data to company KPIs at scale
they can investigate test and demonstrate genuinely additive mosaic theory across thousands of businesses at once.
we’ve seen AI earnings forecasts significantly beating consensus across 2 earnings seasons and even base models call earnings-day share-price reactions
the market thought all these tasks were 'too human' and required too much judgement to do well but they've all been bitter lessoned
now that key facets of the 'existing investment process' can largely be managed by agents, the frontier is thinking deeply about how to use compute discover total novel methods to extract signal from huge swathes of structured and unstructured data
Fundamental equity teams need to think seriously about how they’re set up to compete there
I’ll give you a clue: it’s not signing up for another app with some “agentic features”.
Best-in-class research infrastructure is the first step, build it or buy it, but assess whether you can afford to wait for your dev team to take 12mo to build it
Without that foundation, you can’t reliably automate research end to end, let alone start exploring novel methods...
2
1
24
4,081
uncoveredalpha retweeted
People are missing the point here
Calling this quant, systematic, quantamental or whatever else doesn’t explain what’s happening.
The frontier has moved from helping analysts do their existing process faster to the most ambitious firms pushing towards autonomous, end-to-end fundamental research
These are systems that decide what to investigate, gather evidence, test competing explanations, build forecasts and challenge their own conclusions. Then work out what to investigate next.
to give context, we've shown LLMs can successfully identify genuine regime shifts before shares react
they can also find and map alternative data to company KPIs at scale
they can investigate test and demonstrate genuinely additive mosaic theory across thousands of businesses at once.
we’ve seen AI earnings forecasts significantly beating consensus across 2 earnings seasons and even base models call earnings-day share-price reactions
the market thought all these tasks were 'too human' and required too much judgement to do well but they've all been bitter lessoned
now that key facets of the 'existing investment process' can largely be managed by agents, the frontier is thinking deeply about how to use compute discover total novel methods to extract signal from huge swathes of structured and unstructured data
Fundamental equity teams need to think seriously about how they’re set up to compete there
I’ll give you a clue: it’s not signing up for another app with some “agentic features”.
Best-in-class research infrastructure is the first step, build it or buy it, but assess whether you can afford to wait for your dev team to take 12mo to build it
Without that foundation, you can’t reliably automate research end to end, let alone start exploring novel methods...
Calling this quant, systematic, quantamental or whatever else doesn’t explain what’s happening.
The frontier has moved from helping analysts do their existing process faster to the most ambitious firms pushing towards autonomous, end-to-end fundamental research
These are systems that decide what to investigate, gather evidence, test competing explanations, build forecasts and challenge their own conclusions. Then work out what to investigate next.
to give context, we've shown LLMs can successfully identify genuine regime shifts before shares react
they can also find and map alternative data to company KPIs at scale
they can investigate test and demonstrate genuinely additive mosaic theory across thousands of businesses at once.
we’ve seen AI earnings forecasts significantly beating consensus across 2 earnings seasons and even base models call earnings-day share-price reactions
the market thought all these tasks were 'too human' and required too much judgement to do well but they've all been bitter lessoned
now that key facets of the 'existing investment process' can largely be managed by agents, the frontier is thinking deeply about how to use compute discover total novel methods to extract signal from huge swathes of structured and unstructured data
Fundamental equity teams need to think seriously about how they’re set up to compete there
I’ll give you a clue: it’s not signing up for another app with some “agentic features”.
Best-in-class research infrastructure is the first step, build it or buy it, but assess whether you can afford to wait for your dev team to take 12mo to build it
Without that foundation, you can’t reliably automate research end to end, let alone start exploring novel methods...
THIS IS NOT A DRILL
I don't think you could see a better signal for direction of travel
Do not kid yourselves that this technology is just 'help me do my job faster so I can think more' type of change
Balyasny's head of research and the guy that set up applied AI research has head of autonomous strategies
Clearly they've seen enough from their work that risk could be taken without human involvement with fundamental research foundations
This is who you're going to have to compete with, not larry the long only with his new chatgpt subscription
great flag @HochstatMichael
5
3
65
13,257
This is a fascinating demonstration of where investment research is heading.
What excites me isn’t necessarily the ability to run 10,000 scenarios through a financial model, impressive as that is. It’s what becomes possible when we start questioning the structure of the investment process itself.
Traditional fundamental research is inherently constrained by human bandwidth. We simplify complex businesses into a manageable number of assumptions, construct a handful of scenarios, and ultimately arrive at a point estimate of intrinsic value.
But reality doesn’t operate in neat bull, base and bear cases.
Businesses are complex, interconnected systems. Revenue growth influences margins, competitive responses influence pricing power, capital allocation changes future returns, and the relationships between these variables are neither static nor necessarily independent.
Much of traditional modelling introduces false precision precisely because we have to simplify these relationships to make them analytically manageable.
AI potentially changes that constraint.
But I think there’s an even bigger opportunity beyond probabilistic forecasting.
What if we could construct an interconnected network of specialised research agents, each independently examining different aspects of a business, testing competing hypotheses, challenging assumptions and incorporating new evidence as it emerges?
Rather than simply running thousands of scenarios through a predetermined model, the system could continuously question whether the model itself accurately represents economic reality.
And crucially, connect those findings across companies and industries.
An observation about restaurant throughput at Chipotle might inform assumptions about CAVA. Changes in consumer behaviour at McDonald’s could challenge an existing thesis on Taco Bell. Evidence gathered in one part of the system could dynamically inform hypotheses elsewhere.
The other essential component is expectations.
Predicting a company’s future is only half the investment problem. The other half is understanding what the market has already discounted.
The most valuable system wouldn’t merely tell us what a business might earn in five years. It would identify where its probabilistic assessment of the future differs from the expectations embedded in today’s share price, and explain why.
Of course, greater computational sophistication doesn’t automatically mean better forecasts. Ten thousand simulations built on flawed assumptions can create an illusion of precision just as easily as a traditional DCF.
Which is why the ultimate test has to be empirical: are these systems better calibrated, more accurate and capable of identifying genuine mispricing out of sample?
To me, that’s the frontier.
Not simply automating how analysts conduct research today, but asking whether we can build an entirely better process for discovering, evaluating and updating investment opportunities.
The financial model becomes one evolving hypothesis within a much larger network of evidence, rather than the final destination of the research process.
That’s a much more interesting future for fundamental investing.
What excites me isn’t necessarily the ability to run 10,000 scenarios through a financial model, impressive as that is. It’s what becomes possible when we start questioning the structure of the investment process itself.
Traditional fundamental research is inherently constrained by human bandwidth. We simplify complex businesses into a manageable number of assumptions, construct a handful of scenarios, and ultimately arrive at a point estimate of intrinsic value.
But reality doesn’t operate in neat bull, base and bear cases.
Businesses are complex, interconnected systems. Revenue growth influences margins, competitive responses influence pricing power, capital allocation changes future returns, and the relationships between these variables are neither static nor necessarily independent.
Much of traditional modelling introduces false precision precisely because we have to simplify these relationships to make them analytically manageable.
AI potentially changes that constraint.
But I think there’s an even bigger opportunity beyond probabilistic forecasting.
What if we could construct an interconnected network of specialised research agents, each independently examining different aspects of a business, testing competing hypotheses, challenging assumptions and incorporating new evidence as it emerges?
Rather than simply running thousands of scenarios through a predetermined model, the system could continuously question whether the model itself accurately represents economic reality.
And crucially, connect those findings across companies and industries.
An observation about restaurant throughput at Chipotle might inform assumptions about CAVA. Changes in consumer behaviour at McDonald’s could challenge an existing thesis on Taco Bell. Evidence gathered in one part of the system could dynamically inform hypotheses elsewhere.
The other essential component is expectations.
Predicting a company’s future is only half the investment problem. The other half is understanding what the market has already discounted.
The most valuable system wouldn’t merely tell us what a business might earn in five years. It would identify where its probabilistic assessment of the future differs from the expectations embedded in today’s share price, and explain why.
Of course, greater computational sophistication doesn’t automatically mean better forecasts. Ten thousand simulations built on flawed assumptions can create an illusion of precision just as easily as a traditional DCF.
Which is why the ultimate test has to be empirical: are these systems better calibrated, more accurate and capable of identifying genuine mispricing out of sample?
To me, that’s the frontier.
Not simply automating how analysts conduct research today, but asking whether we can build an entirely better process for discovering, evaluating and updating investment opportunities.
The financial model becomes one evolving hypothesis within a much larger network of evidence, rather than the final destination of the research process.
That’s a much more interesting future for fundamental investing.
We were thinking about how swarms will change the investment process… and I blew my own mind with this one.
I saw CAVA down 10% and ran my automated investment analysis on it – came away thinking it was a good long.
My AlphaSense skill creates its own forecast, reasons on th / penetration vs peers, sensitivities, waterfalls, etc. It does work how I’d do the work.
But then, since SuperAnalyst leverages the fully drivable Canalyst models as a tool, I asked it to run my model scenarios like a weather forecast for me. It ended up running 10,000 !
You have to start from the right base: Canalyst model, leveraging our forecasting skills (dissecting drivers in a HF manner), leveraging my investment analysis skill. But then from that base, we can run this at scale.. And this is the worst it’s ever going to be?
Disclaimer: this is not investment advice. (though I spent 10 years as a hedge fund analyst, and this is better analysis than I could run).
3
2
140
15,540
Couldn’t agree more. It strikes me that most AI adoption in finance has focused on automating or accelerating existing workflows, rather than questioning whether those workflows are optimal in the first place.
As an analyst, I understand the appeal of getting to an answer more interesting question is whether AI can help us arrive at fundamentally better answers?
I’ve always been fascinated by the idea of an interconnected investment brain. Something akin to Obsidian, but purpose-built for investing, where companies, industries, themes, mental models, investment hypotheses and evidence exist within a continuously evolving knowledge graph.
I’ve actually started experimenting with a prototype of this concept, which I’m calling Labyrinth.
Imagine discovering something material about one business and being able to trace its implications across an entire network of companies, assumptions and investment theses. Rather than research existing in isolated notes, PDFs and spreadsheets, knowledge becomes interconnected, cumulative and increasingly useful.
Now combine that with probabilistic modelling.
Instead of a financial model producing a single target price based on point estimates, imagine one that represents a distribution of possible outcomes, understands the dependencies between assumptions and updates those probabilities as new evidence emerges.
The analyst’s role shifts from manually maintaining models and processing information towards challenging assumptions, evaluating evidence and exercising judgement.
I suspect Excel will eventually become less central to the investment process, replaced by something more dynamic, interconnected and probabilistic.
Perhaps the biggest opportunity for AI in investing isn’t helping us do the same work faster.
It’s fundamentally changing how we think, connect information and make decisions under uncertainty.
As an analyst, I understand the appeal of getting to an answer more interesting question is whether AI can help us arrive at fundamentally better answers?
I’ve always been fascinated by the idea of an interconnected investment brain. Something akin to Obsidian, but purpose-built for investing, where companies, industries, themes, mental models, investment hypotheses and evidence exist within a continuously evolving knowledge graph.
I’ve actually started experimenting with a prototype of this concept, which I’m calling Labyrinth.
Imagine discovering something material about one business and being able to trace its implications across an entire network of companies, assumptions and investment theses. Rather than research existing in isolated notes, PDFs and spreadsheets, knowledge becomes interconnected, cumulative and increasingly useful.
Now combine that with probabilistic modelling.
Instead of a financial model producing a single target price based on point estimates, imagine one that represents a distribution of possible outcomes, understands the dependencies between assumptions and updates those probabilities as new evidence emerges.
The analyst’s role shifts from manually maintaining models and processing information towards challenging assumptions, evaluating evidence and exercising judgement.
I suspect Excel will eventually become less central to the investment process, replaced by something more dynamic, interconnected and probabilistic.
Perhaps the biggest opportunity for AI in investing isn’t helping us do the same work faster.
It’s fundamentally changing how we think, connect information and make decisions under uncertainty.
This is so bang on, a lot of investing involves far too much false precision, and it's because it simplifies it for human understanding...this is why I'm excited about AI's use in investing
when you think from first principles, it seems crazy to be putting a precise number on something we know is uncertain...
analysts assign target prices, estimate “fair value” and forecast cash flows five years into the future. The model gives us an answer like $47.83.
but take a step back and think about what went into that number
we don’t know exactly how quickly revenue will grow. we don’t know what margins will be, how much the company will need to reinvest, or what return investors will demand in five years. Every assumption has a range of possible outcomes.
We choose one number for each, run the calculation, and end up with an answer that looks way more certain than anything we put into it.
We all know that investing is about gaining confidence that the probabilistic range of outcomes has a favourable skew
This is why I’m so excited about compute changing how we do this work.
Modelling is part of the problem, it's unfathomable for each cell in excel to have a probabilistic range of outcomes and for you to update those probabilities constantly as information emerges
But AI and newly written software will genuinely make this possible, not to mention the most advanced systems being able to ingest information and update those weights
of course, judgement still exists about the assumptions and their likelihood but I'm really excited to see how this evolves
10
10
132
8,288
Great framing. Two things I'd add.
First, LLMs aren't born overconfident. GPT-4's pre-trained model was well calibrated, and post-training is what broke it. The junior analyst isn't naturally overconfident. They get promoted for sounding certain. We built the same incentives into the models.
Second, next-token prediction is a consensus engine. In maths, consensus converges on the answer. In markets, consensus is the price, and the price is what you're betting against.
The job was never to be right about the company. It's to be right about where the market is wrong about the company. A model trained on the most probable view is built to fade the variant view, which is where the alpha sits.
And unlike a proof, an edge doesn't stay true once everyone knows it. If every fund runs the same model, its view doesn't generate alpha. It becomes the price.
That's why regime changes and "n of 1" events are the killer. The base rates stop working exactly when the money is made.
Where models shine: mapping consensus. You can't find a variant view until you've defined what you're varying from.
Maths rewards the right answer. Markets reward the right odds, relative to everyone else's.
First, LLMs aren't born overconfident. GPT-4's pre-trained model was well calibrated, and post-training is what broke it. The junior analyst isn't naturally overconfident. They get promoted for sounding certain. We built the same incentives into the models.
Second, next-token prediction is a consensus engine. In maths, consensus converges on the answer. In markets, consensus is the price, and the price is what you're betting against.
The job was never to be right about the company. It's to be right about where the market is wrong about the company. A model trained on the most probable view is built to fade the variant view, which is where the alpha sits.
And unlike a proof, an edge doesn't stay true once everyone knows it. If every fund runs the same model, its view doesn't generate alpha. It becomes the price.
That's why regime changes and "n of 1" events are the killer. The base rates stop working exactly when the money is made.
Where models shine: mapping consensus. You can't find a variant view until you've defined what you're varying from.
Maths rewards the right answer. Markets reward the right odds, relative to everyone else's.
This quote from Steve Mandel makes me think about why frontier LLMs are still terrible stock pickers (once you give them a clean sandbox adjusting for lookahead and contamination) even as they conquer frontier mathematics. And are likely to continue to be terrible stock pickers. (I believe this will be particularly true around points of regime change and "n of 1" events in markets...the times when the best human investors drive real separation and retail/consensus/momentum players get washed out).
Frontier LLMs are deterministic optimizers trained on verifiable-reward RL to find one checkable, binary answer. And they are getting incredibly good at it (see: Navier-Stokes). But RL-tuned LLMs fundamentally suffer from premature convergence, like a junior analyst who develops "conviction" after 3 hours of work... (if you have managed over-confident junior analysts, you know this can be quite frustrating)
The fundamental challenge, per Steve Mandel framing below, is that stock picking has no verifiable reward at the decision point. A good thesis can lose money and a bad thesis can make money (though the market gods tend to hand investors the former far more often than the latter...). And that over-confident junior analyst has a way of getting lucky the first time or two (i don't make the rules).
Frontier LLMs have been optimized to collapse uncertainty and generate a definitive, coherent final answer to maximize confidence metrics. This is fundamental to the system, and bigger training runs just make the system better at this job. But fundamental investing is ultimately a game of making predictions where investors must embrace & price uncertainty, think probabilistically, and update those views dispassionately as the evidence changes. This doesn't take frontier-math capabilities, in fact it takes almost no math at all, but it takes the right way of thinking about distributional asymmetry.
Lately have become fascinated with this idea of pricing distributional asymmetry, more to come on that (and love to share notes with any AI-pilled people working on this problem).
0
1
20
1,410
I've spent too long this week staring at the FT chart on US productivity in tech, finance and professional services. If you haven't seen it, the US line has run well above its pre-Covid trend while the UK, Europe and Canada have basically sat on theirs. EPB has the US heavy adopters going from 2.2% a year to 5% since end-2022.
My first reaction was that this is AI. I'm less sure now.
When you split it into output and headcount, US tech output hasn't done anything unusual. It's roughly on trend. Headcount is what moved. Big Tech overhired in Covid, rates went up, and US companies cut much harder than anyone in Europe would. That's what US companies always do. You saw the same thing in 2009.
What I can't explain away is that the line hasn't flattened. If this were only layoffs you'd expect a one-off jump and then back to normal. Instead it's kept steepening into this year, which is right when coding tools started doing real work. My guess is the cuts started for macro reasons and AI is the reason the jobs haven't come back.
The bit I find most interesting is the company data. Meta, Microsoft, Alphabet and Spotify have grown revenue a long way past headcount since 2022. Accenture and Deloitte haven't. That makes sense. A consulting firm still needs more people to take on more clients, however much time AI saves each one.
From an investing point of view I'm not sure the "is it AI" question matters much. Fewer people doing the same output still shows up in margins either way. What I'm watching is which business models let revenue grow without adding people, and whether US output actually breaks above trend over the next year or two. If it does, I'll be a lot more convinced.
My first reaction was that this is AI. I'm less sure now.
When you split it into output and headcount, US tech output hasn't done anything unusual. It's roughly on trend. Headcount is what moved. Big Tech overhired in Covid, rates went up, and US companies cut much harder than anyone in Europe would. That's what US companies always do. You saw the same thing in 2009.
What I can't explain away is that the line hasn't flattened. If this were only layoffs you'd expect a one-off jump and then back to normal. Instead it's kept steepening into this year, which is right when coding tools started doing real work. My guess is the cuts started for macro reasons and AI is the reason the jobs haven't come back.
The bit I find most interesting is the company data. Meta, Microsoft, Alphabet and Spotify have grown revenue a long way past headcount since 2022. Accenture and Deloitte haven't. That makes sense. A consulting firm still needs more people to take on more clients, however much time AI saves each one.
From an investing point of view I'm not sure the "is it AI" question matters much. Fewer people doing the same output still shows up in margins either way. What I'm watching is which business models let revenue grow without adding people, and whether US output actually breaks above trend over the next year or two. If it does, I'll be a lot more convinced.
Hard for me to look at these charts from the FT, especially the first one, and not see some evidence of “ROI on AI.”
Seems like a productivity surge post-Covid era overhiring that began to reverse before ChatGPT came out and then resumed its upward trend after ChatGPT.
a16z data on revenue per employee at their portfolio companies and the @EPBResearch data showing that the 3 industries that have most heavily adopted AI have gone from 2.2% productivity growth to 5% productivity growth.
And makes sense that it is showing up in the US first - America is 72% of Anthropic revenue vs. 50%ish at most application software companies.
Great stuff from @jburnmurdoch even if I disagree with his conclusion.
0
0
8
1,275
The point of my original post was to ask genuine questions and have a spirited debate, not to dunk on Shaun.
A lot of people are using this thread and others to attack him. Very few have done what he did: put a full model ($SPCX) and its assumptions out in public for anyone to pick apart. Disagree with the numbers all you like, but that takes courage, and it's how markets get smarter.
And it's been genuinely useful. Shaun engaged thoughtfully and shared his view that:
2027-28 is a scarcity window with ~1yr paybacks
rev/GW falls towards ~$10bn over the next 5 years
from 2031, growth is driven by orbital volume
steady state is 10-20% net margins alongside enormous capex
I still have open questions on D&A and what steady-state capex per GW requires. But this is what good debate looks like.
Thanks @shaunmmaguire
A lot of people are using this thread and others to attack him. Very few have done what he did: put a full model ($SPCX) and its assumptions out in public for anyone to pick apart. Disagree with the numbers all you like, but that takes courage, and it's how markets get smarter.
And it's been genuinely useful. Shaun engaged thoughtfully and shared his view that:
2027-28 is a scarcity window with ~1yr paybacks
rev/GW falls towards ~$10bn over the next 5 years
from 2031, growth is driven by orbital volume
steady state is 10-20% net margins alongside enormous capex
I still have open questions on D&A and what steady-state capex per GW requires. But this is what good debate looks like.
Thanks @shaunmmaguire
I've seen a lot of debate on Shaun's SpaceX ($SPCX) model,so here's my two cents.
First, credit where due. His 2020 Starlink model, built when Starlink had zero revenue, had it at a ~$23bn run-rate by end-2025. It's at ~$20bn now. That's a remarkable call. And the wood-for-the-trees camp is right that taxes and working capital are noise. Novel businesses need first-principles forecasting, not extrapolation.
But here's the irony. Terrestrial compute is $569bn of his $731bn 2028 revenue (~78%). That's the least futuristic thing SpaceX does: renting GPUs in data centres on Earth. And unlike Starship or orbital compute, it has observable market prices.
So let's engage on his terms. He says compute is a commodity, SpaceX is the low-cost producer, and rev/GW and capex/GW move together.
In a commodity, the low-cost producer wins on cost, not price. And if scarcity reprices the whole market to $40bn/GW, Oracle earns it too on $37bn of capex. That's a boom for everyone, not distress that SpaceX can buy cheaply.
His 2028: $40bn revenue per GW, $30bn capex per GW
Oracle (Investor Day): $16bn revenue, $37bn capex
CoreWeave (reported): ~$7bn annualised revenue per active GW
So SpaceX is modelled as the low-cost producer AND charging 2.5x Oracle's price.
The coupling doesn't fix it. At Oracle's $16bn price, his ratio implies ~$12bn of capex per GW. Nvidia says its own share alone is ~$25bn/GW on Blackwell and ~$40bn on Rubin.
And his own 2026 numbers show capex of $40-55bn per GW, so the cost advantage is still a forecast.
Starship is an extraordinary moat in launch. It doesn't help sell compute in a Texas data centre at 2.5x market.
Genuine question: if compute is oil, why does SpaceX get 2.5x the oil price?
0
0
6
1,367
The irony is that the people who understand the commercial problems best are being restricted from building solutions to them.
The real disruption from agentic coding isn’t making engineers more productive. It’s allowing domain experts to become builders.
AI dramatically reduces the cost of creating software, yet institutions seem determined to preserve the organisational bottlenecks it was supposed to eliminate.
Manage the risks, absolutely. But governing who can build, rather than what can be deployed, seems like fundamentally the wrong abstraction.
The firms that figure this out will have a considerable advantage.
The real disruption from agentic coding isn’t making engineers more productive. It’s allowing domain experts to become builders.
AI dramatically reduces the cost of creating software, yet institutions seem determined to preserve the organisational bottlenecks it was supposed to eliminate.
Manage the risks, absolutely. But governing who can build, rather than what can be deployed, seems like fundamentally the wrong abstraction.
The firms that figure this out will have a considerable advantage.
GS banning traders and bankers from using AI over productivity and risk concerns
Make Tokenmaxxing Great Again
2
0
10
1,541
The full paper is worth reading. A summary, plus a few thoughts on applying it.
The argument GAAP earnings spread costs over time through accruals and depreciation, but value depends on when cash actually moves. Their worked example is a factory project that reports positive net income every year but has a negative NPV on its cash flows. The point isn't that earnings are useless, but that they can mislead on value creation, particularly where pay is tied to EPS.
The evidence They rank stocks monthly on trailing FCF yield, using a 45-day reporting lag, across the Russell 1000, 2000, 3000 and MSCI World. Yields are normalised by industry, and by country for MSCI World, so the test compares companies with their peers rather than sectors with each other.
Results are consistent across all four universes. In MSCI World (1996–2024), the top quintile returned 9.8% a year and the bottom 4.1%, against 7.2% for all stocks. The lowest-yield quintile was also the most volatile in every universe.
Some context Returns are equal-weighted and exclude transaction costs, which the authors note.
The effect is larger in small caps.
A good part of the spread comes from the bottom quintile underperforming, so the measure is useful for screening out names as well as finding them.
The top quintile lagged from 2017 to 2020 before recovering.
Trailing FCF growth also carried information, though less consistently than yield.
Applying it The study uses reported FCF. In practice I'd adjust for a few things:
Share-based compensation, which is a real cost but is added back in cash flow
Lease payments, which IFRS 16 reclassifies as financing and which matter a lot for retail, restaurants and travel
Growth versus maintenance capex
One-off working capital movements
Combined with a view on returns on capital and how management allocates the cash, it's a sensible starting point for valuation work.
The argument GAAP earnings spread costs over time through accruals and depreciation, but value depends on when cash actually moves. Their worked example is a factory project that reports positive net income every year but has a negative NPV on its cash flows. The point isn't that earnings are useless, but that they can mislead on value creation, particularly where pay is tied to EPS.
The evidence They rank stocks monthly on trailing FCF yield, using a 45-day reporting lag, across the Russell 1000, 2000, 3000 and MSCI World. Yields are normalised by industry, and by country for MSCI World, so the test compares companies with their peers rather than sectors with each other.
Results are consistent across all four universes. In MSCI World (1996–2024), the top quintile returned 9.8% a year and the bottom 4.1%, against 7.2% for all stocks. The lowest-yield quintile was also the most volatile in every universe.
Some context Returns are equal-weighted and exclude transaction costs, which the authors note.
The effect is larger in small caps.
A good part of the spread comes from the bottom quintile underperforming, so the measure is useful for screening out names as well as finding them.
The top quintile lagged from 2017 to 2020 before recovering.
Trailing FCF growth also carried information, though less consistently than yield.
Applying it The study uses reported FCF. In practice I'd adjust for a few things:
Share-based compensation, which is a real cost but is added back in cash flow
Lease payments, which IFRS 16 reclassifies as financing and which matter a lot for retail, restaurants and travel
Growth versus maintenance capex
One-off working capital movements
Combined with a view on returns on capital and how management allocates the cash, it's a sensible starting point for valuation work.
Cash flow is the most overlooked metric in stock valuation nowadays.
0
3
14
2,025
I've seen a lot of debate on Shaun's SpaceX ($SPCX) model,so here's my two cents.
First, credit where due. His 2020 Starlink model, built when Starlink had zero revenue, had it at a ~$23bn run-rate by end-2025. It's at ~$20bn now. That's a remarkable call. And the wood-for-the-trees camp is right that taxes and working capital are noise. Novel businesses need first-principles forecasting, not extrapolation.
But here's the irony. Terrestrial compute is $569bn of his $731bn 2028 revenue (~78%). That's the least futuristic thing SpaceX does: renting GPUs in data centres on Earth. And unlike Starship or orbital compute, it has observable market prices.
So let's engage on his terms. He says compute is a commodity, SpaceX is the low-cost producer, and rev/GW and capex/GW move together.
In a commodity, the low-cost producer wins on cost, not price. And if scarcity reprices the whole market to $40bn/GW, Oracle earns it too on $37bn of capex. That's a boom for everyone, not distress that SpaceX can buy cheaply.
His 2028: $40bn revenue per GW, $30bn capex per GW
Oracle (Investor Day): $16bn revenue, $37bn capex
CoreWeave (reported): ~$7bn annualised revenue per active GW
So SpaceX is modelled as the low-cost producer AND charging 2.5x Oracle's price.
The coupling doesn't fix it. At Oracle's $16bn price, his ratio implies ~$12bn of capex per GW. Nvidia says its own share alone is ~$25bn/GW on Blackwell and ~$40bn on Rubin.
And his own 2026 numbers show capex of $40-55bn per GW, so the cost advantage is still a forecast.
Starship is an extraordinary moat in launch. It doesn't help sell compute in a Texas data centre at 2.5x market.
Genuine question: if compute is oil, why does SpaceX get 2.5x the oil price?
First, credit where due. His 2020 Starlink model, built when Starlink had zero revenue, had it at a ~$23bn run-rate by end-2025. It's at ~$20bn now. That's a remarkable call. And the wood-for-the-trees camp is right that taxes and working capital are noise. Novel businesses need first-principles forecasting, not extrapolation.
But here's the irony. Terrestrial compute is $569bn of his $731bn 2028 revenue (~78%). That's the least futuristic thing SpaceX does: renting GPUs in data centres on Earth. And unlike Starship or orbital compute, it has observable market prices.
So let's engage on his terms. He says compute is a commodity, SpaceX is the low-cost producer, and rev/GW and capex/GW move together.
In a commodity, the low-cost producer wins on cost, not price. And if scarcity reprices the whole market to $40bn/GW, Oracle earns it too on $37bn of capex. That's a boom for everyone, not distress that SpaceX can buy cheaply.
His 2028: $40bn revenue per GW, $30bn capex per GW
Oracle (Investor Day): $16bn revenue, $37bn capex
CoreWeave (reported): ~$7bn annualised revenue per active GW
So SpaceX is modelled as the low-cost producer AND charging 2.5x Oracle's price.
The coupling doesn't fix it. At Oracle's $16bn price, his ratio implies ~$12bn of capex per GW. Nvidia says its own share alone is ~$25bn/GW on Blackwell and ~$40bn on Rubin.
And his own 2026 numbers show capex of $40-55bn per GW, so the cost advantage is still a forecast.
Starship is an extraordinary moat in launch. It doesn't help sell compute in a Texas data centre at 2.5x market.
Genuine question: if compute is oil, why does SpaceX get 2.5x the oil price?
Over the last week I built a bottoms up SpaceX financial model through the end of 2028
The numbers that came out are insane
There are wide error bars given how quickly things are changing but I wanted to distill something complicated into something that can at least be debated
nt advice!!!
** Sequoia is a major SpaceX shareholder
*** Past performance is not indicative of future results
32
36
462
136,985
Read the whole thing over coffee, really good piece of work. Hashing the tests before looking at outcomes and publishing the ones that failed is more rigour than most sell-side or buy-side backtests ever get.
The bit that made me laugh was the rules of thumb making the model worse. Half the industry's earnings-morning process is basically a list of those, so that one stung a bit.
A few things I've been chewing on from a UK small/mid cap angle:
A lot of UK names don't put out a classic earnings release. Half the reporting calendar is trading updates with a line on full-year expectations and no proper P&L. Did those end up in the European sample, and did the model handle them as well as full results? I'd guess the expectations packet matters even more there.
On the after-hours point in the limitations, in the UK most of this happens at 7am RNS with nothing trading until 8, so you've got a clean hour where the call could actually be used before any price exists. Feels like a better test bed than the US for the live season.
And the 93% on confident "much bigger" calls is great, but on 57 calls I'd want to see how wide that interval is before anyone builds a process around it. Would be good to see it hold up in the live run.
Looking forward to the next one.
The bit that made me laugh was the rules of thumb making the model worse. Half the industry's earnings-morning process is basically a list of those, so that one stung a bit.
A few things I've been chewing on from a UK small/mid cap angle:
A lot of UK names don't put out a classic earnings release. Half the reporting calendar is trading updates with a line on full-year expectations and no proper P&L. Did those end up in the European sample, and did the model handle them as well as full results? I'd guess the expectations packet matters even more there.
On the after-hours point in the limitations, in the UK most of this happens at 7am RNS with nothing trading until 8, so you've got a clean hour where the call could actually be used before any price exists. Feels like a better test bed than the US for the live season.
And the 93% on confident "much bigger" calls is great, but on 57 calls I'd want to see how wide that interval is before anyone builds a process around it. Would be good to see it hold up in the live run.
Looking forward to the next one.
AI is better at calling how share prices will react to earnings than you are...
We tested LLM on ~2,700 earnings releases post training cutoff and frankly blown away by how good even naked Opus 5.5 was at calling share prices correctly.
Gets the direction correct 64% of the its confident about the call, it's right 74% of the time.
Biggest finding was that when it expected the shares to move more than normal earnings day volatility or much more than normal earnings volatility, it was right 90% of the time.
1
0
40
6,080
I'm starting a weekly series going back through the special situations in famous investor letters and rebuilding the numbers myself. First up is from Buffett's 1960 partnership letter.
The business
Sanborn made incredibly detailed maps of every US city for fire insurers. Water ations, roof materials, building by building. For about 75 years it was effectively a monopoly. Profitable every year, recession-proof, no sales force needed.
Then in the early 1950s insurers started using a cheaper underwriting method called "carding". Map profits went from over $500k a year in the late 1930s to under $100k by 1958.
The business needed no capital, so from the 1930s Sanborn had been putting retained earnings into stocks and bonds. About $2.5m went in. By 1958 it was worth roughly $7m, or $65 a share.
The stock was $45.
Compare that to 1938. Stock at $110, portfolio worth $20 a share, so the market was paying +$90 a share for the map business. Twenty years later, with the Dow up about 5x, it was valuing the same business at -$20. You were being paid to own a business that still made money.
Of 14 directors, nine came from the insurance industry. Between them they owned 46 shares out of 105,000. The insurers were Sanborn's customers, not its owners, and had been net sellers of the stock for a decade.
The portfolio doing well took away any urgency to fix the core business. Dividends were cut five times in eight years. Buffett notes he couldn't find a single suggestion to cut salaries or directors' fees.
So you had a hidden asset, a board with no skin in the game, and a shareholder base that had given up.
The son of a former president quit after being refused the top job. Buffett bought his mother's ~15,000 shares, then built the position to ~24,000 (with associates) in the market, about 23% of the company. Two other unhappy blocks lined up with him, taking the group to 46,000 shares, roughly 44%.
He joined the board and pushed to separate the portfolio from the map business. Rather than a proxy fight, they agreed a deal: any shareholder who wanted out could swap their stock for portfolio securities at fair value. The SEC signed off.
About 72% of the stock took the exchange.
The clever part was the structure
If Sanborn had sold the portfolio and paid out cash, it would have owed corporate capital gains tax on ~$4.3m of gains. At 25% that's about $1.1m, or ~$10 a share. Handing over the securities in kind avoided it entirely. Buffett puts the tax eliminated at over $1m.
The holders who stayed kept the same map profits spread over about 28% of the shares, so EPS went up roughly 3.5x, plus a bond reserve of $1.25m+.
The numbers
I rebuilt it in Excel.
Base case, per share:
Share price: $45.00 Portfolio: +$20.00 to $65.00 Latent corporate tax: -$10.30 Map business at 10x earnings: +$9.52 Value even if sold and taxed: $64.23 Tax saved by the exchange: +$10.30 Intrinsic value: $74.52
The letter doesn't give Buffett's entry price. Assuming around $45 and an exit at $65 through the exchange, that's ~44% over roughly 1.75 years, or ~23% a year, on a position that was 35% of partnership assets. In 1960 the partnership made 22.8% while the Dow lost 6.3%.
And the downside: with the map business at zero, the portfolio could fall ~30% before you lost money.
What I'm taking from it
1. The best asset plays hide in balance sheet lines that P/E screens never see.
2. Check board ownership. 46 shares out of 105,000 tells you most of what you need to know.
3. Cheap isn't enough. Someone has to force the value out, or there needs to be a clear path to it.
4.Structure can be worth as much as the discount. Here the tax saving was about half the size of the discount itself.
5. Liquid assets beat a turnaround story as downside protection.
Where does this show up today? Holdcos, family-controlled companies sitting on cash or listed stakes, businesses where the operating company is valued below zero.
The business
Sanborn made incredibly detailed maps of every US city for fire insurers. Water ations, roof materials, building by building. For about 75 years it was effectively a monopoly. Profitable every year, recession-proof, no sales force needed.
Then in the early 1950s insurers started using a cheaper underwriting method called "carding". Map profits went from over $500k a year in the late 1930s to under $100k by 1958.
The business needed no capital, so from the 1930s Sanborn had been putting retained earnings into stocks and bonds. About $2.5m went in. By 1958 it was worth roughly $7m, or $65 a share.
The stock was $45.
Compare that to 1938. Stock at $110, portfolio worth $20 a share, so the market was paying +$90 a share for the map business. Twenty years later, with the Dow up about 5x, it was valuing the same business at -$20. You were being paid to own a business that still made money.
Of 14 directors, nine came from the insurance industry. Between them they owned 46 shares out of 105,000. The insurers were Sanborn's customers, not its owners, and had been net sellers of the stock for a decade.
The portfolio doing well took away any urgency to fix the core business. Dividends were cut five times in eight years. Buffett notes he couldn't find a single suggestion to cut salaries or directors' fees.
So you had a hidden asset, a board with no skin in the game, and a shareholder base that had given up.
The son of a former president quit after being refused the top job. Buffett bought his mother's ~15,000 shares, then built the position to ~24,000 (with associates) in the market, about 23% of the company. Two other unhappy blocks lined up with him, taking the group to 46,000 shares, roughly 44%.
He joined the board and pushed to separate the portfolio from the map business. Rather than a proxy fight, they agreed a deal: any shareholder who wanted out could swap their stock for portfolio securities at fair value. The SEC signed off.
About 72% of the stock took the exchange.
The clever part was the structure
If Sanborn had sold the portfolio and paid out cash, it would have owed corporate capital gains tax on ~$4.3m of gains. At 25% that's about $1.1m, or ~$10 a share. Handing over the securities in kind avoided it entirely. Buffett puts the tax eliminated at over $1m.
The holders who stayed kept the same map profits spread over about 28% of the shares, so EPS went up roughly 3.5x, plus a bond reserve of $1.25m+.
The numbers
I rebuilt it in Excel.
Base case, per share:
Share price: $45.00 Portfolio: +$20.00 to $65.00 Latent corporate tax: -$10.30 Map business at 10x earnings: +$9.52 Value even if sold and taxed: $64.23 Tax saved by the exchange: +$10.30 Intrinsic value: $74.52
The letter doesn't give Buffett's entry price. Assuming around $45 and an exit at $65 through the exchange, that's ~44% over roughly 1.75 years, or ~23% a year, on a position that was 35% of partnership assets. In 1960 the partnership made 22.8% while the Dow lost 6.3%.
And the downside: with the map business at zero, the portfolio could fall ~30% before you lost money.
What I'm taking from it
1. The best asset plays hide in balance sheet lines that P/E screens never see.
2. Check board ownership. 46 shares out of 105,000 tells you most of what you need to know.
3. Cheap isn't enough. Someone has to force the value out, or there needs to be a clear path to it.
4.Structure can be worth as much as the discount. Here the tax saving was about half the size of the discount itself.
5. Liquid assets beat a turnaround story as downside protection.
Where does this show up today? Holdcos, family-controlled companies sitting on cash or listed stakes, businesses where the operating company is valued below zero.
2
2
63
3,180