ExpertsFMCliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497]
Invest Like the Best with Patrick O'Shaughnessy · October 27, 2026
5 Selected Lessons:
case · 2 heuristic · 1 priority · 1 framework · 1
Investing in long-term equilibrium winners over the capital-cycle builders
Source
Invest Like the Best with Patrick O'Shaughnessy — Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497] (October 27, 2026) · at 23:51 ▸
When this applies
when a new infrastructure/technology cycle creates many speculative investment points along its value chain
The takeaway
In an infrastructure buildout, prioritize analyzing what durable businesses use the matured infrastructure over trying to pick which builder wins the arms race.
Why it works
The builders of infrastructure (railroads, data centers) face winner's-curse dynamics, long lead times, and overbuild risk, making it very hard to predict who captures the rents; the downstream beneficiaries that use the infrastructure once built are more predictable because the equilibrium outcome (infrastructure exists cheaply) is more certain than which builder wins.
One level up
When multiple layers of an uncertain value chain exist, allocate analytical effort to the layer with the least irreducible uncertainty, not the most exciting one.
Where it stops holding
This stops working when the builder layer itself has clear, durable moats (e.g., a true structural cost advantage) making it the more predictable layer instead.
Detail
hierarchy:
- 1. Longer-term equilibrium winners who use infrastructure once built (e.g., the 'CPG brands' that come after railroads)
- 2. (lower priority) Picking which specific infrastructure builder ('railroad tycoon') wins the buildout race
rationale: Sosin states it is 'very, very hard... to figure out where one would invest along this' supply chain and has 'decided to focus attention more on the longer-term equilibrium results.'
In their words
Instead of betting on which railroad tycoon is going to make money and how many railroads we need, I'm trying to think about, okay, once you have railroads, I'd love to have the foresight to figure out CPG brands come next.
CEO AI-adoption rollout ladder: hackathons → read access → write access → change-management
Source
Invest Like the Best with Patrick O'Shaughnessy — Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497] (October 27, 2026) · at 42:36 ▸
When this applies
when rolling out any powerful new capability across an organization with legacy systems and permission structures
The takeaway
Roll out transformative capability in a deliberate escalation — demonstrate the art of the possible broadly first, then expand systemic access in stages (read → write → change authority), building the governance for each stage before moving to the next.
Why it works
Mandatory company-wide hackathons reset everyone's expectations of what's possible; then sequentially expanding an AI agent's access from read-only to write to full change-management mirrors how trust and process maturity must be built incrementally, because each step (read, write, change control) surfaces its own organizational and technical obstacles that take real time to resolve.
One level up
Speed of cultural buy-in (via demonstration) should outpace, but not substitute for, the slower structural work of permissioning and governance — both are necessary and sequenced deliberately.
Where it stops holding
This stops working, i.e. becomes a liability, when a company takes too long to complete this ladder and a competitor starting from zero with AI-native processes ('jump ball from Go') overtakes it before legacy incumbents finish the rollout.
Detail
procedure:
- 1. Mandate company-wide hackathons for everyone, not just engineers, to reset expectations of what AI can do
- 2. CEO demands AI agent get read-only access to every company system — treat resistance ('that's really hard') as exactly the point
- 3. Extend read-only access down through the org as others get jealous and request it, working out permissions each time
- 4. Grant write permissions as requests arise, building governance for data input
- 5. Establish change-management processes as people want to use AI to actually modify systems
benchmarks: {"timeframe":"Sosin estimates this process, from read-only access onward, takes 'years of work'"}
rationale: Each stage of trust/access unlocks the next business value but also surfaces real organizational friction that must be resolved deliberately rather than skipped.
In their words
The CEO needs to basically turn to the head of IT and say something to the effect of, so I want my Grockbot, MetaMuse, agent, whatever, to be able to access every system in the company on a read-only basis.
Own the complements to intelligence
Source
Invest Like the Best with Patrick O'Shaughnessy — Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497] (October 27, 2026) · at 6:45 ▸ · at 9:16 ▸ · at 12:18 ▸ · at 47:21 ▸
When this applies
when a key input to your industry is about to become commoditized and nearly free
The takeaway
When a formerly scarce input commoditizes, look for the assets that don't commoditize alongside it and that compound independently of that input's price.
Why it works
If intelligence becomes cheap and ubiquitous, the scarce, non-reproducible assets become whatever AI can't replicate: proprietary data, captured attention/emotional attachment, physical proximity, scale, and regulatory capture — these become the actual sources of durable value.
One level up
Durable advantage comes from owning what a falling-cost technology can't substitute for, not from owning the technology's output itself.
Where it stops holding
This stops working when the business's whole value proposition IS the crystallized intelligence itself (e.g., a static software tool with no data/attention moat), in which case cheap AI directly destroys it.
Detail
rule: Assume everyone has access to near-free, great intelligence, then ask what's left that differentiates a business — that remainder (data, attention, physical proximity, scale, regulatory capture) is what to own.
threshold: If a company's core value is 'crystallized intelligence' (a hardened projection of some thinking task) with no data/attention moat, treat it as vulnerable; if value is customer lethargy/behavioral inertia, treat it as vulnerable too.
rationale: Complements to intelligence are the finite resources AI agents cannot create or fully substitute for, so their value rises as intelligence itself becomes abundant.
In their words
If intelligence becomes cheap and ubiquitous, what you want to own are the complements to intelligence. And so the complements to intelligence are their data, attention, love, affection, whatever sort of emotional attachment you want. It's physical proximity, it's scale.
Garbage collection consolidation via AI agents
Source
Invest Like the Best with Patrick O'Shaughnessy — Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497] (October 27, 2026) · at 29:59 ▸
When this applies
when a market has high switching friction despite low differentiation between providers, and an automated agent could remove that friction
The takeaway
In markets protected mainly by consumer switching friction (not real service differentiation), expect agent-mediated commerce to collapse that friction and consolidate volume onto the structurally lowest-cost, highest-density provider.
Why it works
Consumer switching inertia currently protects incumbent local providers even when a cheaper option exists; AI agents that autonomously compare and switch on the consumer's behalf eliminate that friction, pushing volume toward the lowest-cost local provider, which in turn gains further scale/density advantages, triggering a rollup dynamic.
One level up
Identify which markets' current structure depends on human laziness/inertia as the moat — those are the markets most exposed to agentic disruption and consolidation.
Where it stops holding
This stops working when consumers have genuine, non-delegable preferences (trust, relationship, personal taste) rather than purely optimizable decisions — Sosin's own example of 'picking my favorite sweater' being non-delegable.
Detail
situation: Garbage collection pricing is sticky because consumers don't bother to check/switch providers even for meaningful savings.
decision: If each consumer has an AI agent empowered to compare and switch automatically, agents will move households to the lowest-cost provider.
outcome: The provider who already serves multiple households on a street becomes the low-cost density leader and can profitably underbid rivals, triggering a street-by-street rollup toward one dominant low-cost provider.
lesson: Agent-mediated decision delegation on a 'defer-to-AI' spectrum of consumer choices (bank deposits=full deferral, sweaters=no deferral) will consolidate markets where decisions sit toward the deferrable end.
In their words
a second-order effect might be that the guy who has two people on my street can go to everyone else on the street and say, Hey, I'll offer you a lower price on the garbage, and everyone will switch.
Sosin's 80% concentrated Carvana position justified by scenario analysis, not price-volatility theory
Source
Invest Like the Best with Patrick O'Shaughnessy — Cliff Sosin - Owning the Complements to Intelligence - [Invest Like the Best, EP.497] (October 27, 2026) · at 52:31 ▸
When this applies
when deciding whether to concentrate resources heavily on one opportunity you believe is exceptional, against conventional risk-management wisdom
The takeaway
Concentration is justified only when your specific, scenario-level analysis of a business's downside risk is unusually narrow and well-grounded — not as a general strategy, and not something average investors should emulate.
Why it works
Standard diversification research is based on price-change statistics across average investors/securities, not on deep company-level conviction about a specific business's differentiated infrastructure and brand; if an investor has unusually high confidence that the downside scenarios are narrow and unlikely because of durable structural advantages, concentration can be a rational trade-off of expected return for known, accepted unusual risk — a trade historically seen in the most successful concentrated investors, while remaining the wrong approach for average investors.
One level up
General statistical research describes what's optimal on average; a specific, well-evidenced exception to the average case can still justify diverging from the general rule for that one case, while the general rule remains correct for most people.
Where it stops holding
This stops working, by Sosin's own admission, for 'most people who are for average or slightly better than average results,' for whom diversification 'is totally the right thing to do' — concentration is only right given his specific level of company-level conviction.
Detail
situation: Sosin holds roughly 80% of his portfolio in Carvana, far outside conventional diversification norms.
decision: He bases comfort with this concentration on believing Carvana's infrastructure and brand are differentiated enough that few scenarios end badly, likening it to how someone who found Walmart/Home Depot/Fastenal early wouldn't have captured full upside if diversified.
outcome: He states he'd be fine either way — the stock going up and forcing diversification is also an acceptable outcome — framing it as an expected-return/unusual-risk trade-off, not a free lunch.
lesson: Extreme position concentration can be defensible specifically when grounded in deep, scenario-level company analysis showing narrow downside — not as a general portfolio rule.
In their words
Were it not for my sense that Carvana's infrastructure and brand and everything else is just so differentiated that there are few scenarios where it works out really badly, then I wouldn't feel so comfortable so much of it.