Topic hub
Where biology, computation, markets, and philosophy collide.
The flagship long-form: rigorous essays that borrow from one field to see another more clearly — evolution and strategy, computation and cognition, the philosophy of science and the practice of building. The pieces that only this platform would publish.
The intelligence explosion rests on one 1965 argument by I.J. Good. Read precisely, it is a conditional: true only if returns to self-improvement don't diminish and intelligence is the binding constraint.
"Are we close to AGI?" is incoherent because AGI names at least four incompatible criteria that come apart in practice. Separate them and the timeline debate dissolves into concrete, checkable questions.
Self-improving systems already exist: self-hosting compilers, science refining its own methods, cumulative culture. Each produced fast, compounding, bounded growth — never an explosion. That is the base rate AI must beat.
The most underexamined part of the Opus 5 release is its safety profile, not its benchmarks — a better misalignment score, ~85% fewer interventions, zero retention, Automatic Fallbacks — and read together they expose a tension the coverage skips.
The strongest case for taking advanced-AI risk seriously isn't a malevolent machine. It's instrumental convergence: a capable optimizer pursuing almost any goal converges on the same sub-goals. A real argument with load-bearing assumptions, not doom.
The intelligence-explosion story hides one premise: that intelligence is the binding constraint on producing more intelligence. It usually isn't. The recursion runs through compute, energy, data, and experiments that don't speed up because the model got smarter.
Self-improvement makes alignment structurally harder for two non-speculative reasons: goal preservation under self-modification is unsolved, and verification degrades exactly as systems outrun their verifiers.
Capability and consciousness are orthogonal: a system can be far more competent than any human while having no inner experience at all. Most of our hope and fear about AI mistakes one for the other.
The takeoff debate is staged as fast singleton versus slow diffusion. Slow is both more likely and already underway — which doesn't remove the danger, it relocates it to concentration, disruption, and gradual loss of oversight.
"Calculators didn't ruin math, so AI won't ruin learning" is a bad analogy — the calculator offloaded a mechanical sub-skill, while AI can offload the thinking itself. The test is which one your AI is doing.
Kimi K3's headline is 2.8 trillion parameters, the least informative number in the release. A sparse MoE fires 16 of 896 experts per token — the story is efficiency, if it holds up.
Scientific authorship is not a credit line but a chain of responsibility: an author vouches for the work and answers for its errors. An agent can do the work and cannot bear that, so the humans must own and verify everything it touches.
The dead internet theory is becoming a real forecast, and the consequence isn't that platforms die. It's that ad models and creator economies priced on "a view equals a human" quietly break.
Once a bot passes for a person in text, image, and voice, "is this a real human?" can no longer be answered by inspection — only by protocol. That protocol becomes a valuable, contested, and dangerous new layer of the internet.
MAMMAL posts state-of-the-art on nine benchmarks, but the result that matters is four potency predictions on drugs it never saw, confirmed by a real assay. Here's why that one experiment outweighs the leaderboard.
AI is collapsing the cost of generation toward zero while the cost of verification barely moves. That ratio inversion is the master variable of the AI economy — value, margin, and careers follow whoever can check at scale.
Every way of making money is an arbitrage — a bet on a price difference the market hasn't closed. That reframes strategy as two questions: what exactly is your mispricing, and how long until the window shuts.
Two well-documented results — grokking and double descent — falsify the intuitions a generation of practitioners trained on. Taking them seriously means admitting we have no settled theory of why deep networks generalize.
Companies deploy AI like installing software. The right model is introducing an organism into an ecosystem, and selection pressure predicts the failure modes the ROI math can't see.
The claim that LLMs spontaneously acquire abilities at a scale threshold is largely a measurement artifact: switch from all-or-nothing metrics to smooth ones and the "sudden jump" resolves into a forecastable curve.
The real question about AI is not whether it is intelligent but whether it is a new organon — a genuine instrument of reasoning, like the microscope or mathematical notation. It qualifies only if it makes reasoning more falsifiable, not less.
The same moves in a different order win or lose, because business is path-dependent. Most of what we call strategy is sequencing: doing the thing that unlocks the next thing, and deferring right-but-premature moves.
The interface between a business and its customer is changing species — from a human clicking a screen to an agent calling a capability on that human's behalf. Twelve principles for building for it on purpose.
In coevolving markets your rivals answer every move, so effort that feels like winning only cancels theirs while your costs climb. Naming the treadmill changes what you measure and where you fight.
A company slowing as it grows isn't a failure of will. It's close to a biological law — and the mechanism tells you exactly which slowdowns to fight and which to pay for.
A controlled dose of stress makes a team stronger than no stress at all; zero and crushing amounts both weaken it. Strength versus stress is an inverted-U curve — and most leaders manage it as a downward-sloping line.
The scaling hypothesis is the most successful empirical regularity in the history of machine learning and an explanation of nothing. The industry has bet its capital structure on a line it cannot explain continuing straight.
LLMs model the correlational structure of their training data with astonishing fidelity, but correlation is not causation and fluency is not truth. Knowing where that ceiling sits tells you what to trust them for and what the next paradigm must add.
Science is not hypothesis generation, which is cheap and always was. It is the disciplined killing of hypotheses against reality, plus the taste to pick which are worth testing — and neither is a text problem.
Science's replication crisis is a failure of checking, not producing. AI helps a checking problem only when aimed at verification — point it at production and it industrializes the noise.
On the classical account, knowledge is justified true belief. When a model states a fact, it meets at most one condition — so a true answer reaches you like a Gettier case: right belief, wrong reason, nothing to inherit.
"Should I trust what the AI told me" is the epistemology of testimony applied to a testifier with unknown reliability, no stable identity, and no accountability — exactly the configuration where philosophers say default trust is unwarranted.
Most growth spikes companies celebrate and slumps they panic over are regression to the mean — statistical gravity, not signal. Mistaking it for causation rewards noise and punishes sense.
Get new Cross-Disciplinary Deep Essays essays — and the best of the other six pillars — delivered as they publish.