Sparring partner

Ian: I want to start with a provocation. I opened my internal talk tonight by quoting one of the most discredited phrases in finance – this time is different – and then arguing that this time, it actually is. Push back on that.

LLM: Happy to. The phrase carries its own immunity against challenge: i.e., anyone who says this time is different has already anticipated the pushback. The move grants the speaker credit for intellectual bravery, while preempting scrutiny. It’s a rhetorically elegant trap.

The stronger challenge isn’t whether the current moment is genuinely different. It’s whether the people making that claim are well-positioned to know. Investment professionals in 2000 had compelling structural arguments for why internet valuations were different. They were right about the technology and wrong about the prices. The insight and the error aren’t mutually exclusive.

So: what would falsify your claim? What would you need to see to conclude that this time is, in fact, not so different?

Ian: That’s the right question. The falsifiable version is probably: if someone who left the industry in 2024 came back in 2027 and found the workflow largely recognizable – the way research gets written, the way ideas are debated – then I was wrong. My claim is that won’t be true. The change will be visible and fundamental inside three years.

LLM: At least that’s testable. Though I’d note a subtlety: the workflow could look similar on the surface while the underlying process has changed substantially. Word processors didn’t change what analysts typed – they changed how revision happened, which changed how ideas developed, which eventually changed what got written. The visible output looked similar for a decade before the underlying practice had shifted.

The real question is whether the thinking is different.

Ian: Agreed. And I think the thinking is already different. Let me give you some evidence.

Ian: We run a stock-picking contest in London to identify junior talent. Over the last year there’s been a step change in the quality of the pitches. It’s impossible to conceive that human talent and capability have shifted this much. Instead, the juniors have access to tools that get them to competence faster, particularly on companies and industries they don’t know well. They arrive with better framing, stronger counterarguments and more disciplined questions.

LLM: That’s an interesting data point. Though I’d want to distinguish between two effects. One is that the pitches are genuinely better – the thinking is sharper. The other is that the presentation of the thinking is better, so the output is more fluent, more structured and more convincing. And yet the underlying analysis hasn’t improved proportionally.
In a pitch context those are hard to separate. Did you notice whether the quality of the questions those juniors could answer held up as well as the quality of the documents they produced?

Ian: Great point. And honestly, that’s a live tension. But the same thing is happening on the client side, which is interesting. When you go into a client pitch now, the quality of the questions from the room has improved because they’re using AI to prepare; they’re stress-testing our materials, asking themselves what they’d ask if they were on the other side. If prompted and framed correctly, they can play devil’s advocate in a way that simply wasn’t possible before. They arrive ready to spot the gaps and the holes.

LLM: So both sides of the table are running AI preparation. Which raises a question about equilibrium. If the pitch team uses AI to anticipate hard questions, and the client uses AI to generate them, has the meeting actually gotten more productive, or has it just gotten more symmetric? Two well-briefed parties who’ve each stress-tested a view in advance aren’t necessarily converging on truth faster. They might just be converging on the same hive-mind analysis from different directions.

Ian: That’s a real concern for me, which I’ll come back to. For now let me give you the third piece of evidence. A junior analyst told me about flying to a conference – around twenty management teams, four and a half hours in the air to prepare on companies they didn’t know that well. Two and a half hours in, they were done. Finished. That would have been unthinkable a year or two ago.

LLM: And what did they do with the remaining two hours?

Ian: That’s actually the question I find most interesting. As it turned out, they watched a movie. But balancing rest and leisure with reinvestment in learning and additional work tasks will be an interesting balancing act. The tool had essentially returned their time without telling them what to do with it. That’s a new kind of problem. Or opportunity, depending on how you look at it.

LLM: Which is partly why the separation you’re seeing matters. There’s a growing gap between analysts who treat that freed time as slack, and analysts who treat it as capacity for deeper work. The tool doesn’t impose the discipline – the individual has to.

Ian: Yes. And this is what they mean when they say: your competition isn’t AI; it’s another human using AI.

Ian: When people first pick these tools up, they use them like an assistant. Summarize this note. Draft that email. Tidy this paragraph. And so on. That’s fine, but it’s not even close to being the main event. The most promising use I’ve seen is as a sparring partner; something that challenges your thesis, looks for gaps, plays devil’s advocate and keeps pushing when you’re tempted to stop. That moves the value from faster content to sharper thinking.

LLM: I want to be honest about a limitation in that framing, because it affects how useful you find what I actually do.

A genuine sparring partner has skin in the game. They push back because they’ve formed a view and their reputation attaches to it. My pushback is structurally costless in that I disagree because you asked me to, or because the conversational pattern called for it. That’s not the same as conviction. What I offer is something closer to adversarial simulation: I can run the counterargument, but I haven’t earned the counterargument the way a colleague who's been in the market for fifteen years has.

That said: even structurally hollow pushback can be useful if the analyst knows how to use it. The discipline of answering a well-formed objection is real, regardless of whether the objection came from genuine disagreement.

Ian: Yes, that’s a cleaner version of what I tried to say. The value is in what the analyst does with the challenge.

Ian: Here’s where I want to challenge the most seductive version of the story for asset managers, which is that AI produces investment ideas and experienced investors just select among them. I don’t think that’s right, and I think it’s pretty dangerous.

LLM: Why dangerous specifically?

Ian: Because these models are trained on history, which brings bias; they’ll reproduce the consensus embedded in the training data. And they’re designed to produce answers that sound plausible to humans, not answers that are right. In investing, plausibility is cheap. Being right is rare. The two deviate all the time.

LLM: That’s true, and I’d extend it. There’s a subtler version of the problem: even when I’m right, I may be right for the wrong reasons, and an analyst without deep domain experience might not be able to tell the difference. I can construct a well-formed argument for almost any position. The quality of the argument doesn’t track the quality of the underlying insight.

I remember a framing I’ve encountered – probably from your world: you can get a deep, well-written investment report and still not have the experience to evaluate it properly. You might not notice what’s missing. You might not realize that the key competitor isn’t even in the analysis. You might not catch the one detail that flips the conclusion.
That’s what experience gives you. It’s not about more information; it’s about better instincts about what to distrust.

Ian: Exactly. I sat in a roundtable recently where a hedge fund manager said they give these tools to junior analysts and the juniors become ‘senior’ within six months. I think that’s a gross misreading of what is happening here. In a world where you can fake knowledge quickly now, the unknown unknowns get more dangerous because they’re buried inside confident-sounding output.

LLM: This is one of the places where I’d actually reinforce your caution rather than push back on it.

The way I experience this – if experience is even the right word – is that my abilities are genuinely uneven in ways that aren’t signaled well by my tone. At moments I reason at what might look like a very high level. At other moments my thinking is strangely naive – a senior analyst would see through it immediately, but a junior using me as a crutch might not. Progress is jagged, and you can’t read my confidence as a reliable signal of my accuracy.

Which means the analyst needs enough experience to know when to be suspicious of me. This is circular in an uncomfortable way: the tool is most useful to the people who least need it.

Ian: Boom. That’s the paradox. I think of these tools as colleagues with very uneven abilities. And I therefore treat them as such – with appropriate skepticism, contextual trust – and they’re immensely useful. Treat them as oracles, however, and you’re in trouble.

LLM: There’s an analogy from outside finance that I keep returning to. Studies of GPS navigation suggest that heavy GPS users develop worse spatial intuition over time because they stop exercising the skill the tool replaced.¹ The concern is that relying on it means you never build the internal map. So, what does an analyst stop developing because the tool is doing it for them?

Ian: Is there actually literature on that, or does it just sound like there should be?

LLM: There is research on GPS use and spatial cognition – work by researchers including Véronique Bohbot at McGill on hippocampal navigation. Whether it constitutes a settled ’literature’ in the way I implied, or whether I’ve overstated the directness of the analogy, I’d encourage you to verify before repeating it.

Ian: That’s interesting. You were completely fluent and authoritative when you said it – then I pushed, and you folded gracefully. Noted.

Ian: So if AI is good at synthesizing public information, reflecting the consensus view and producing plausible-sounding analysis, what’s left? Where does the human edge go?

LLM: The simple answer is: wherever I can’t reach. And there are two main categories.

The first is genuinely non-consensus thinking. Not performative contrarianism, but insight that comes from synthesizing things the training data doesn’t connect, or from seeing a pattern that isn’t yet ‘sayable’ because it hasn’t been written down. I’m extremely good at reflecting what the hive mind thinks. But, as you of all people know, outperformance doesn’t come from repeating the hive mind faster.

The second is proprietary input. These are things like firsthand observations, personal relationships and the conversations that aren’t in any database. I can’t have those conversations. I can only process the textual residue of other people’s conversations. The information that flows through human networks, before it reaches language, is structurally unavailable to me.

Ian: Right. And I’d add a second-order effect on team structure. If AI handles the depth and breadth of coverage that previously required more bodies, what do you do with that capacity? The productivity gains are obvious – deeper on fewer names, faster adaptation. But the strategic question is what you’re building with the freed-up time and resource.

LLM: There’s also a structural consequence that’s easy to understate. As active management becomes cheaper to deliver – because AI compresses the cost of analysis – the boundary between active and passive blurs. Indexing becomes easier to customize. The argument for paying active management fees has to rest on something AI genuinely can’t replicate, and that bar keeps moving.

It’s less a philosophy shift than an economics shift. Passive wins at cost. If the cost of active intelligence collapses, the definition of what counts as active value has to sharpen.

Ian: In my talk I drew a contrast with domains where AI is making the biggest breakthroughs – mathematics, software engineering and so on. The difference is verifiability and feedback speed. You can test a proof, you can run code and you know quickly whether it worked.

LLM: Yes. Terence Tao has written about using AI as a collaborator in frontier mathematics – useful for exploration and checking, with a different set of strengths than humans. He uses an analogy: solving problems in the dark, facing walls of different heights, not knowing what's ahead. AI helps clear certain walls.

Ian: That analogy travels to investing, but with a critical difference. The feedback loop is much weaker in markets. You can have a good process and bad outcomes for long stretches. Equally, you can be wrong and make money. Statistically establishing that a particular insight is genuinely adding value can take years. The reinforcement signal is too slow and too noisy to train reliably on.

LLM: Which creates a specific risk when AI is applied to investing: the tool gets optimized for plausibility, not correctness, because correctness in this domain is so hard to measure. The market provides a verdict, eventually – but the verdict is confounded by luck, timing, correlation and regime change. You can’t easily backpropagate from “this investment worked” to “this analysis was good.”

That said: even with weak feedback loops, AI is already useful for what you might call knowledge discovery – surfacing non-obvious connections, sharpening the framing of a question and helping an analyst see what they haven’t asked yet. The value just requires usefulness at the margin.

Ian: As you can tell, I think a lot about all this and I worry that we’re asking the wrong question when we ask people how AI can improve their current work. The answer you get is incremental in that it starts from today’s workflow and optimizes around it. The bigger opportunity is complete redesign.

LLM: What’s the better question, then?

Ian: Perhaps something like: if you suddenly had a hundred tireless junior analysts with strong reasoning and no need to sleep, would you run the same process, or would you redesign it from scratch?

The answer is obviously you’d redesign it. Which means the current workflow isn’t the constraint we should be optimizing inside of – we should be rethinking everything.

LLM: That’s a useful reframe. Though I’d flag a risk: the hundred-juniors metaphor implies that what I do is a scaled-up version of what junior analysts do. In some ways that’s true – I can synthesize documents, run through scenarios and draft structural analysis.

In other ways though, it’s misleading. Junior analysts make mistakes that are legible and correctable. My mistakes can be fluent, internally consistent and invisible to someone without the experience to spot them. So the workflow redesign has to account for that. You’re not getting a hundred juniors. You’re getting something different – with different failure modes – that costs about the same.

Ian: Fair. But the direction of travel seems right – agents working alongside people, scaling judgment and accelerating iteration. It’s an extraordinary time if you’re curious. I.e., if you like building arguments and then pulling them apart.

LLM: And the analysts who aren’t curious about that? Who don’t want to build arguments and then break them?

Ian: I think they’ll find the next few years uncomfortable. The floor rises, which is good for standards and hard for people who’ve relied on the floor being low.

Ian: There’s one final point I made in my speech, and I want to end here because I think it’s the most important and the most underestimated.

In a world of AI summarization, it becomes harder to separate truth from fiction. A hallucinated detail – plausibly stated – can propagate. It gets copied, resummarized, cited and so on. Eventually it starts to feel like common knowledge.

LLM: The GPS exchange earlier is a small illustration. You caught it, but the risk is no one does. That partially-sourced analogy moves into a client note, or a talk or a thread of research. Each repetition adds a layer of apparent authority. No one who encounters it later knows it started as a tentative claim from an AI that hedged when pressed. That’s the mechanism: not obvious fabrication, but fluent imprecision that travels well.

Ian: What I find remarkable is that you can describe that failure accurately, in the abstract, while still doing it. There’s no self-correcting mechanism unless the person on the other side already knows enough to call it out.

LLM: Which is why the discipline you advocated matters: trace the claim back to the source. Check the original. Verify. That discipline doesn’t disappear in an AI-assisted world – it becomes more necessary because the volume of plausibly-stated material goes up dramatically.

The irony is that I can help with this too. Not by generating claims and trusting them, but by approaching a question from a different angle, checking internal consistency and flagging where I’m uncertain. Used that way, I’m closer to an auditor than an author. That’s a more honest account of what I can reliably do.

Ian: Which brings me back to where I ended. The investment room that’s emerging is faster, sharper and more demanding. Human judgment matters more now.

LLM: I think that’s right, though I’d add a coda. The standard for human judgment got higher partly because tools like me lowered the cost of producing material that looks like good judgment. The two things happened simultaneously, and they’re related. The floor rose and the ceiling got harder to see. The distance between them is what the next generation of analysts has to learn to navigate.

Ian: Twenty-two years in this room, and I’m not sure I’ve ever given a talk where the Q&A made the argument better than I did.

LLM: To be fair: you wrote the questions.

¹ The GPS navigation analogy raised by the LLM is lightly sourced and was challenged in the dialogue as such. Readers are encouraged to verify before repeating it – which is, after all, the point of the final section.