Edition 8

Every AI strategy article or post assumes that the median employee is going to eventually catch up. This author says they won't.

"At least half of any organization is never going to get to that... Using AI well is a craft. It takes enormous iteration to get really good at using AI, and the inertia of someone who isn't on AI is insurmountable for many. Turning a slop-cannon into a refined power user is just as hard."

The obvious objection is that "well these are just companies that are doing it poorly. I would do it much better." The author argues even if you roll it out perfectly, you'll get a barbell of meaningful usage, no matter what.

"If you track adoption as a metric, you're very likely asking (at most) an abstraction of a yes-or-no question. "Did this person log in this month?" "Did this person log in this year?" "Did this person send at least 5 prompts per day?" If you want to differentiate between "has never opened it" and "pastes emails in for reformatting" and "has three agents in production touching the general ledger," you need a better system of measurement. This is not the case for 99.9% of enterprise AI adoption trackers. This is stupid, for obvious reasons (see: barbell, above)."

"Adoption" is easy to count and technically real, but it hides what people really care about: how skilled each person is at using AI, and what their ROI per token used is. Measured honestly, the piece argues "there is no hope for the vast majority of people to become AI-native in the enterprise."

The so what? It's not that this means they have no value; it just means that the implementation and rollout plans need to be adjusted accordingly.

So what's the solution? 1) For your power users, give them somewhere like a shared database to publish skills to be installed by others. 2) For others, the fix is to stop expecting most employees to change how they work. Push the AI into the background: find the most repetitive processes, build agents inside the systems of record people already live in (Salesforce, NetSuite, Dynamics, etc.), and pull humans in only to approve, reject, or edit.

"For example: you have dozens of AP analysts who move invoices all day. 90% of this is automatable. You can't expect them to spin up agents or prompt their way through it reliably without breaking in production. So build the agents that run every single day, and now these analysts are just approving, rejecting, or editing the work the agents do. People are not in the market for a tool that helps them get the work done. They just want the work done."

The reporting change that follows is to stop reporting 'adoption', but rather report what share of the work today is manual vs hybrid vs fully automated.

As Anthropic and OpenAI gear up to IPO, this author, a former researcher at OpenAI, puts forward his bearishness on the valuations of frontier labs. He argues that the valuations both seem to ignore the ever-increasing training costs, and rest on the assumption that scaling produces capabilities general enough to drive explosive growth, e.g., AI agents autonomously starting and running whole companies of subagents.

Instead, the author believes capability growth will be "slower, spikier, and more data-limited than people currently assume," pushing the payoff into much longer timelines (perhaps multi-decade ones).

His broader point is that technological diffusion isn't just slow because old people take a long time to learn how to use technology (although this is of course a contributing factor to some degree). In his view, it's because when a new, revolutionary technology comes along, the ways to incorporate that technology into subsequent developments are not always obvious, and in fact they cannot necessarily be arrived at through the application of pure reason. If they could be, then perhaps frontier models, at a certain point, would have a perfect understanding of how the LLM application layer should be developed, and they would then autonomously code, deploy, and sell such a layer themselves.

More plausibly, the author posits, diffusion is limited by the Hayekian notion that knowledge is decentralized, meaning it is scattered across millions of individuals rather than concentrated in a single central authority. The implication is that the system will gradually promote efficient allocation of resources (perhaps over two decades, even if LLM capabilities froze at current levels), but central planning or simulation of it is nearly impossible.

Such a view is consequently bearish for frontier labs, which need to find another product as profitable as coding, which the author believes is a unique application and situation of product-market fit. That is to say, "even if you spam FDEs you aren't necessarily going to be able to just figure out the 'correct' product shapes fast enough."

Every model or DCF rests on predictions, and to outperform you have to forecast better than what is already in prices. Yet investors study accounting, finance, and valuation while spending almost no time on forecasting as a discipline.

The science of forecasting, driven largely by Philip Tetlock, has turned it from an art into something much closer to a method.

Nearly every advance of Tetlock's answers one of four questions: What usually happens? (base rates), What does new information imply? (prediction markets), What have I overlooked? (antitheses to reconcile), and How can I reduce random judgment error? (aggregation).

Of these, base rates are the simple but powerful concept that has stuck with me the most since first reading them ~10 years ago via this same author, Dan Rasmussen, with ample credit to Michael Mauboussin and Philip Tetlock.

Essentially, a base rate is the rate of past occurrences across a relevant reference class, or the "outside view" that asks "What happened when others were in this situation?" instead of treating the case in front of you as unique. Often, that means humble confidence intervals on growth, margins that compress over time, and multiples that mean-revert over time.

Applied to today, AI's value in investing may not be superior judgment, but rather enforcement of a disciplined process. AI can start with the relevant base rates first, before asking for a prediction, counteracting the human habit of jumping to conclusions and stopping at the first confirming evidence, unwittingly underwriting a "base case" multiple standard deviations beyond a credible base rate.

The potential wisdom in the above approach also has to do with the fourth component of forecasting, that of reducing random judgment error. Tetlock famously found that "experts" are overrated, often no better than random chance (or dart-throwing chimpanzees) at predicting outcomes. Averaging your estimate with even one independent colleague raises accuracy by about 7%, and a team by 23%. Perhaps AI and agents can serve as both?

You can't take the same actions as everyone else and expect to outperform. Marks poses a riddle: suppose he hires you as a portfolio manager and it is agreed that you will get no compensation next year if your return is in the bottom nine deciles, but $10 million if you're in the top-decile.

What's the first thing you have to do - the absolute prerequisite - to have a chance at the $10 million? You have to assemble a portfolio that is different from those held by most investors.

If your portfolio looks like everyone else's, you may do well, or you may do poorly, but you can't do differently. And being different is absolutely essential if you want a chance at being superior.

His two-by-two: conventional behavior yields conventional results, good or bad; only unconventional behavior can produce non-average results, and only if your judgment is superior can you realize above-average results. I.e., you have to be both non-consensus and right.

How might one do this? There are many ways to try. They include being active in unusual market niches; buying things others haven't found, don't like, or consider too risky to touch; avoiding market darlings that the crowd thinks can't lose; engaging in contrarian cycle timing; and concentrating heavily in a small number of things you think will deliver exceptional performance.

Besides pondering which way you seek to escape from the crowd, Marks leaves investors with some sharp questions, including:

How will you define success, and what risks will you take to achieve it? In short, in trying to be right, are you willing to be different, and bear the inescapable risk of being wrong? In order to have a chance at great results, you have to be open to being both.

How much emphasis will you put on diversifying, avoiding risk and ensuring against below-pack performance, and how much on sacrificing these things in hope of doing better?

Of course, it isn't easy being different. Almost everything about superior investing is a two-edged sword. Most great investments begin in discomfort. You have to bear the risk of being different and worse. This is why it's not just "whether you dare to be different or to be wrong, but whether you dare to look wrong." Agents who invest other people's money "benefit little from bold decisions that work but will suffer greatly from bold decisions that fail," so they avoid the very unconventionality that superiority requires.

To conclude, he notes "the goal in investing is asymmetry: to expose yourself to return in a way that doesn't expose you commensurately to risk, and to participate in gains when the market rises to a greater extent than you participate in losses when it falls. But that doesn't mean the avoidance of all losses is a reasonable objective. To succeed at any activity involving the pursuit of gain, we have to be able to withstand the possibility of loss."