My background is in quantitative trading and portfolio construction.
Markets teach you pretty quickly that being right is not enough.
You can have the correct thesis and still lose because you run out of capital before the thesis pays off.
A lot of finance is really about surviving the path: sizing, leverage, drawdowns, liquidity, financing.
That habit has affected how I read AI strategy.
One assumption kept standing out to me. A lot of discussion about frontier AI implicitly assumes that the capital keeps coming.
That may be reasonable. But it is still an assumption.
Frontier AI is extremely capital intensive, so I became interested in what changes during a serious funding contraction.
My main concern is that everything probably does not slow equally.
Capability work, product work, and safety work sit in different economic positions inside a company. Some are easier to connect to near term revenue than others.
During a contraction, that matters.
I would not claim that safety automatically gets cut first. But I do think its funding structure deserves more attention.
There is another asymmetry I find interesting.
Capital can disappear much faster than infrastructure does.
The GPUs remain. The trained models remain. The people and strategic incentives remain.
So a downturn can change who owns or controls frontier AI assets without making those assets disappear.
That could mean acquisitions, consolidation, government support, asset sales, or some combination of them.
The state involvement case is especially interesting to me.
We usually discuss centralized AI development as a deliberate policy choice. Financial distress could create concentration for much more mundane reasons.
You could end up with a more centralized system without having designed the governance around that concentration in advance.
This feels very familiar from finance.
During good times, funding structure looks boring. During stress, it suddenly becomes the whole story.
Institutions that looked independent turn out to share lenders, counterparties, infrastructure, or incentives.
I suspect AI has similar hidden dependencies.
This is probably the part of AI safety where my markets background is most useful.
I am interested in the assumptions underneath a strategy and what happens when one of them breaks.
Capital is one of those assumptions.
I do not think we understand the failure modes around it very well yet.
]]>Expertise becomes easier to capture.
A good analyst working with an AI system is not only producing an answer. She is also showing the system around her how she works.
What does she check first? Which result makes her suspicious? Which objections matter? When is something good enough to send?
Do this every day and parts of that judgment gradually become prompts, evaluation rules, software, and workflow.
The company becomes better at reproducing the work.
I think this is more important than the usual image of AI suddenly replacing a job.
The transition can be much quieter.
A team starts using better tools. One person can cover more work. A junior role is not replaced. Another process gets automated. Two years later, the organization needs fewer people to produce the same output.
Nobody had to do anything obviously wrong.
In fact, the employee doing a great job can help accelerate the process.
That is the uncomfortable part.
I sometimes think of this as organizational distillation. The analogy is imperfect, but the direction is useful: knowledge that used to live mainly inside a person becomes encoded in something cheaper to reproduce.
Attribution then becomes messy.
If an internal AI system improves after two years of interaction with a research team, it is hard to separate what came from the base model, the engineers, the firm’s data, and the experts who repeatedly corrected it.
For the worker, that distinction matters a lot.
For the firm, maybe less so.
This is why I care more about distribution than the headline number of jobs created or destroyed.
AI can make an economy much more productive while making some forms of knowledge work less valuable in the labor market.
Those two things are completely compatible.
And if expertise increasingly becomes an asset owned by institutions rather than something workers rent to institutions through their labor, the economic consequences could be large.
I do not have a confident forecast for how large.
But I think the mechanism is already worth paying attention to.
The change may look less like mass replacement and more like slow compression: smaller teams, fewer entry level roles, and more of the firm’s knowledge living in software.
No dramatic moment.
Just a different labor market a few years later.
]]>One lesson has become pretty consistent: the biggest improvement in output often comes from improving my plan, not changing the model.
When I know what I am testing, break the work into the right pieces, and check intermediate results, models are extremely good. Sometimes much better than me inside a specific step.
When I ask for the whole thing at once, quality drops quickly.
The missing piece is often not intelligence. It is structure.
Most real work is messy. An experiment fails. A result changes the original question. Something that looked promising on Monday gets thrown away on Thursday.
Current models can reason about each of these decisions. They are still less reliable at holding the entire project together across many of them.
I notice this constantly in research and coding.
The model can propose ten experiments. I still decide which one is worth running.
It can find problems in a result. I still decide whether the result is salvageable or whether we should restart.
It can optimize toward an objective. I am usually the one deciding whether that objective still makes sense.
So I am somewhat skeptical when we casually describe current systems as having goals in the same way people do.
In most of my workflows, a lot of the apparent goal directedness comes from the surrounding system: prompts, tools, memory, evaluations, and a human deciding what happens next.
I expect this to change.
Models are getting better at longer tasks, tool use, self correction, and maintaining state. As that improves, more of the structure I currently provide will move into the model or agent system.
That is obviously useful.
It also means we are gradually removing the human who currently notices when the plan stops making sense.
I think that part deserves more attention.
Some of what looks like human friction today is also oversight.
]]>The root problem is representational. When you flatten everything into a token sequence, you impose a linear ordering on data that does not have one. Topology does not care about order. It cares about connectivity, holes, branches — the shape of a space. None of that is naturally expressible in one dimension. A model can learn to approximate it implicitly, or lean on sophisticated positional encodings, but those fixes are brittle and uninterpretable. If your representation cannot encode the topology, you are planning blind: interpolating, not reasoning.
State-space models do not escape this. I find them genuinely interesting, and Mamba’s input-dependent dynamics are a real step forward. But an SSM is still a 1D recurrence. A richer recurrent state is not the same as the right inductive bias for geometric structure; you are still asking a sequence model to learn topology from data, which is a hard inductive leap. Architecturally, I expect better answers to arise natively in the right geometric domain: graph networks, equivariant architectures.
But architecture is not the only way out. Tool use and memory already relieve some of the pressure, because they externalize the topology problem. Instead of compressing geometric structure into weights or a hidden state, the model queries something that understands the structure natively — a graph database, a symbolic reasoner, a search algorithm. The model becomes a coordinator rather than the geometric reasoner itself. That does not close the gap. It is still a very meaningful practical mitigation.
]]>I would put parts of my own field in that bucket. Backtesting, signal generation, execution optimization — much of quantitative research and systematic trading can be automated, and whatever can be automated gets commoditized. So far the damage is concentrated at the entry level: the junior-analyst pipeline is quietly being automated away. I would not assume seniority stays protected for long.
But fairness requires a caveat. Even in quant trading, there is a layer AI has not cracked: generating a truly novel hypothesis, or reading a situation no dataset has seen. History keeps teaching this lesson. Quant strategies melt down, and discretionary, fundamental judgment outperforms every model trained on the past.
What we are watching is not disappearance. It is a compression of the talent pyramid. A team of ten becomes two people with AI. The jobs are being redefined, not erased — fewer rungs, steeper ladder.
So where does human judgment stay irreplaceable? In high-stakes relational work. M&A. Distressed situations. Board-level strategy. A client in the middle of a transformative deal does not just want the right answer; they want someone who will own it and answer for it. Trust and accountability are the last things you can automate.
]]>In finance, Sharpe ratio measures expected excess return relative to volatility.
I started playing with a very simple question:
Even if the underlying strategy is good, how often should it still look bad?
First, the assumptions, because the numbers below depend on them.
I am using a toy model where monthly excess returns are independent and identically distributed, normally distributed, and generated from the same mean and volatility every month. “Underperform” means the monthly excess return is below zero. I assume the true Sharpe is known and constant, and I look at 60 months, or five years.
Obviously markets do not behave exactly like this. Returns can have fat tails, autocorrelation, changing volatility, and regime shifts.
This is a clean baseline, not a model of reality.
Suppose annual Sharpe is 2.
Under the usual square root of time annualization,
\[S_{month}=\frac{2}{\sqrt{12}}\approx0.577\]Why does this imply a 28.2% chance of underperforming in one month?
Let monthly excess return be
\[r\sim N(\mu,\sigma^2)\]Monthly Sharpe is simply
\[\frac{\mu}{\sigma}=0.577\]I call a month bad when \(r<0\).
Standardizing:
\[P(r<0) = P\left( \frac{r-\mu}{\sigma} < -\frac{\mu}{\sigma} \right)\]So:
\[P(r<0)=\Phi(-0.577)\]where \(\Phi\) is the standard normal CDF.
And
\[\Phi(-0.577)\approx0.282\]So even in this very clean Sharpe 2 toy world, about 28% of individual months underperform.
That was already more than I expected.
The next calculation is slightly different.
Because I assumed monthly returns are i.i.d., every month has the same bad month probability
\[p=0.282\]and one month’s outcome does not change the next month’s probability.
For one specific four month window, the probability that all four months are bad is just
\[p^4 = 0.282^4 \approx0.0063\]or about 0.63%.
But over five years, there are many opportunities for a four month streak to occur.
There are 57 possible four month windows inside 60 months.
We cannot simply multiply \(57\times0.63\%\), because those windows overlap. January through April and February through May, for example, share three months.
So I calculate the exact probability with a tiny Markov chain.
At any point, the strategy can be in one of four states:
\[0,\;1,\;2,\;3\]meaning the current run contains zero, one, two, or three consecutive bad months.
Each new month does one of two things.
With probability \(1-p\), the month is good and the state returns to 0.
With probability \(p\), the month is bad and the streak increases by one.
Once state 3 gets another bad month, we have hit four consecutive underperforming months.
Run that process for 60 months and calculate the probability that we ever hit four.
The result is:
\[P(\text{at least one 4 month streak in 60 months}) \approx23.3\%\]So under these assumptions, even a Sharpe 2 strategy has roughly a one in four chance of experiencing at least one four month underperformance streak over five years.
That is the part I actually find interesting.
A few months of bad outcomes can contain information.
They can also just be variance.
The difficult part is knowing which one you are looking at.
Now take annual Sharpe = 1.2.
Monthly Sharpe becomes
\[\frac{1.2}{\sqrt{12}}\approx0.346\]So the probability of a bad month is
\[p=\Phi(-0.346)\approx36.5\%\]Using exactly the same 60 month Markov calculation:
\[P(\text{at least one 4 month streak}) \approx49.3\%\]Basically a coin flip.
Again, that is not a claim that real Sharpe 1.2 strategies have exactly a 49.3% probability of doing this.
It is what falls out of this particular i.i.d. normal model.
That distinction matters.
I also played with combining two strategies.
This calculation needs another set of assumptions.
Assume the two strategies have the same expected excess return \(\mu\), the same volatility \(\sigma\), the same Sharpe \(S=\mu/\sigma\), constant correlation \(\rho\), and we put 50% into each.
Then
\[r_p=\frac{r_1+r_2}{2}\]Expected return stays
\[E[r_p]=\mu\]while portfolio variance is
\[\operatorname{Var}(r_p) = \frac14 \left( \sigma^2+\sigma^2+2\rho\sigma^2 \right)\]so
\[\sigma_p = \sigma\sqrt{\frac{1+\rho}{2}}\]Therefore:
\[S_p = \frac{\mu} {\sigma\sqrt{(1+\rho)/2}} = S\sqrt{\frac{2}{1+\rho}}\]For two Sharpe 1.2 strategies:
\[\rho=0 \quad\Rightarrow\quad S_p = 1.2\sqrt2 \approx1.70\]At correlation 0.2:
\[S_p = 1.2\sqrt{\frac{2}{1.2}} \approx1.55\]At correlation −0.3:
\[S_p = 1.2\sqrt{\frac{2}{0.7}} \approx2.03\]That is just standard diversification math under equal means, equal volatilities, fixed correlation, and no transaction costs.
Obviously, people are not portfolios :)
But I like the analogy.
Some of the people I value most are useful precisely because their strengths and mistakes are different from mine.
One is better at execution. Someone else is calmer when things go wrong. Another catches assumptions I tend to miss.
The value is not that everyone has a higher individual “Sharpe.”
It is that we are not perfectly correlated.
And the first calculation gives me a related reminder:
a bad stretch is evidence, but it is not automatically a verdict.
Sometimes the process is broken.
Sometimes you are seeing variance.
I think knowing when to tell those two apart is probably more useful than trying to avoid bad months entirely.
]]>