Intelligence will become nearly free: the real value is in data and attention, not the model itself
His core thesis: intelligence will become nearly free, so don't bet on companies building compute and models. Instead, bet on data, attention, and physical scenarios that intelligence cannot replace.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Chaos theory reveals the physical ceiling of intelligence
A meteorologist told the story that explains it: run a weather forecast twice, changing only the precision of initial pressure and temperature data—drop a decimal place. Both runs should produce nearly identical results. Instead, one forecast says sunny and 70°F, the other says storms and 20°F—worlds apart. The reason: many systems are inherently unpredictable. Not because intelligence is lacking, but because simulating part of the universe requires simulating the universe entire. No model, however sophisticated, surmounts this constraint.
— Cliff SosinWhen everyone owns best-in-class intelligence, what competitive advantages remain?
His method for analyzing a company: assume all people and all companies own top-tier intelligence, then ask what remains that intelligence cannot buy. The answer is data, attention, emotional attachment, physical proximity, scale, even regulatory moats—these are the ‘complements to intelligence.’ If a company's value comes from consumer laziness and reluctance to think, AI proliferation destroys it. If value comes from these complementary resources, AI proliferation strengthens it.
— Cliff SosinMeta's entire 250-billion-dollar budget flows almost exclusively into intelligence
From a satellite view of Meta: 250 billion in revenue, 250 billion in spending, nearly all flowing to intelligence—data centers are the infrastructure for building intelligence, purchased inference is intelligence, engineering is intelligence, even reviewing video and moderating content count as intelligence-related work. As intelligence costs plummet, this spending theoretically disappears. What Meta truly owns, what intelligence cannot destroy, is massive user data and attention. That is the moat.
— Cliff SosinJevons paradox applies only to one segment of the demand curve
He rejects the claim that lower prices mean infinite demand growth. Every market traces a U-shaped curve: the product is too weak, nobody buys it; it improves and stays expensive, few buy it; price drops and demand explodes Jevons-style, then continues falling and demand actually shrinks, eventually to zero. He cites fiber-optic communications: transmission costs collapsed, markets boomed then imploded. Cogent Communications today carries approximately 30-40% of global IP traffic but has a market cap measured in only billions of dollars.
— Cliff SosinBreeding mice takes a year: automated labs still hit biological clocks
Automated wet labs and robot-driven research cycles do accelerate learning, but they crash into the physical world's speed limit. If an experiment requires mice to breed and mice need a year to mature and die, the experiment takes a year—no model changes that. Even if AI picks experiments that succeed at only 1% but that is ten times better than human researchers, that is impressive. But lab scale and physical time remain hard constraints, not ‘the model suddenly grasped something humans never understood.’
— Cliff SosinUniversal comparison agents force every industry toward winner-take-most
Take garbage collection. Today people lazily keep their inefficient providers rather than switching. Once everyone owns an AI agent that auto-compares prices and negotiates for them, the market overnight shifts to the lowest-cost provider. Further: the agent can persuade an entire neighborhood to consolidate to one provider; economies of scale let that provider cut prices further; one player wins everything, but margins compress to nothing.
— Cliff SosinAdvising portfolio companies: vibe-code hackathon first, read-only access second
His staged rollout advice: First, run a ‘vibe coding’ hackathon where everyone—board to CEO—participates. The goal is not usable software but resetting everyone's expectation of what AI can do. Next, the CEO requires an AI agent have read-only access to all company systems. Then gradually expand to all staff, then write access, then agents that directly modify systems. This takes years. The real question is: do you have enough time to complete this before newcomers starting from zero catch up?
— Cliff SosinSplit stock-picking from position-sizing: AI portfolio management vastly outperforms
A friend ran an experiment: split investing into ‘what to pick’ and ‘how much to allocate.’ Stock-picking remained human. Position-sizing split into two: one branch stayed human, one went to AI. The AI branch vastly outperformed the human's already-mature position-management method. Cliff sees AI entering investing first in second-to-minute cycles, then gradually longer ones, because long-cycle strategies need decades to validate.
— Cliff SosinIn their own words · checked verbatim
What you should really think about intelligence as is more like a ratio, like the roundness of a ball.
Cliff Sosin2:04
We're all going to have a team of 200 IQ people working full-time in our pocket.
Cliff Sosin6:08
the core business at Meta is personalized TV. Personalized TV. It used to be social media, but now it's personalized TV.
Cliff Sosin9:17
if that experiment requires breeding mice and mice need a year to grow up and die, that experiment is going to take time.
Cliff Sosin22:37
Price is a miracle because it's a signal wrapped in an incentive.
Cliff Sosin35:05
it's the best time to start a company until tomorrow.
Cliff Sosin44:27
Think you're the software company that charges bowling alleys. And those guys all get smoked, which sucks for them, but it does lower the cost of bowling.
Cliff Sosin49:36
Figures
| Meta annual revenue and spending | 250 billion USD each | 9:17 |
| Expected intelligence cost reduction | approximately 100-fold over next five years | 10:17 |
| Cogent Communications IP traffic share and market cap | carries approximately 30-40% of global IP traffic; market cap only billions of dollars | 18:31 |
| AI infrastructure consuming US resources | approximately 3% of GDP being redirected to AI buildout | 38:15 |
| Sosin personal portfolio concentration | approximately 80% Carvana, approximately 5% Meta | 42:24 |
Glossary
- Jevons Paradox
- The phenomenon where efficiency improvements in resource use lead to increased total consumption rather than decreased consumption.
- complements to intelligence
- Resources, capabilities, or assets that AI cannot replace and become more valuable as artificial intelligence scales and commoditizes knowledge work.
- diamond-shaped demand frontier
- A demand distribution where effective demand for intelligence expands as capability improves, then contracts again as price drops, forming a diamond-shaped curve.
- vibe coding
- Generating usable programs through conversational interaction with AI rather than writing code directly by hand.
- freestyle chess
- Human-machine collaborative chess competition in which human players use computer assistance, once producing stronger play than pure human or pure AI competitors.
How to listen
Primary market investors concerned whether AI valuation narratives are overheated, and founders trying to figure out where their company's competitive moat stands after AI becomes ubiquitous.
The first ~1 minute contains sponsor ads and can be skipped.