shout out to Varun Sharma, my former manager at Scale AI, who for me coined IQ + EQ as the key axes in life. he did it in relation to Jeffrey Li and Leigh Marie Braswell as the other 2 candidates in the early Scale squad who might excel in this particular intersection. 

i claim nothing of our relative rankings, but i'm just grateful today that Varun was in my life, and that other writers on the internet (human and ai) appear to be discovering precisely why this is a high leverage way to define things...

June 8, 2026

The old world had a name for the person who could think farther than he could execute: impractical.

He saw structures before he could move them. He built private theories of work, people, machines, language, markets, and futures, then struggled at the ordinary business of turning those theories into finished things. The doing path was narrow and slow. It ran through hands, schedules, managers, brittle tools, organizations that understood roles better than models, and the constant tax of converting a whole thought into a sequence other people could recognize.

That person often survived as a quirky doer. Useful in bursts, hard to place, too theoretical for execution roles and too impatient for pure contemplation. The mismatch looked personal. It was partly technological.

Computers changed the price of the mismatch. They are the best doers ever made: calculators expanded into universal executors, giant symbol movers that can run a million tiny steps without complaint. They also became just enough like thinkers to accept instruction at the level where a strange thinker actually lives. The machine can be told the shape, not only the keystroke. It can return a candidate artifact, receive correction, and try again.

Execution became conversational enough for thought to enter directly.

The advantage moves to the person whose inner life already has executable structure. Some minds naturally produce representations, constraints, examples, edge cases, gradients, maps, procedures, and tests. In the old economy, those shapes still had to be translated into normal doing. In the hybrid economy, those shapes are close to the machine's native material. The thought travels a shorter path before it acts.

That shorter path is the glove fit.

This is translation cost collapsing. A representation is a bet about which operations will be cheap. The glove-fit mind already represents problems in forms whose cheap operation is machine action: state to update, constraint to satisfy, test to pass, interface to route through. The same intelligence that looked over-abstract when action was manual becomes practical when the representation finally meets a machine built to run it.

The glove-fit practitioner is a translator whose source language is thought and whose target language is action. He can hold the abstraction, express it operationally, watch the artifact come back, and update the abstraction from the artifact's failure. The loop is fast enough that thinking and doing stop feeling like separate phases. The thought grows hands.

That image is literal enough to matter. A model trained toward "humanity" from outside humanity has a slow target. Humanity is already moving, already contradicting itself, already metabolizing fear, fashion, politics, status, trauma, desire, money, family, law, ritual, language, and tools. Full humanity is too live to be compressed into a reward model on the schedule this moment demands.

The near path is to build on humanity.

Fit the machine to living humans who already carry unusually good local compressions of the world. A Karpathy-shaped human is one public image: technical intuition, pedagogical compression, code fluency, and taste in the same hand. There are many others with no audience at all: edge operators, quiet engineers, obsessive teachers, patient organizers, systems thinkers in dark corners of the world, crawling toward the light with bad tools and accurate instincts. The point is general. The first scalable interface lives where thought is already shaped like action.

The method has four moves. Find the hand: a human whose priors, taste, skill, commitments, and local world-contact already compress something real. Fit the glove: give that human machines that understand enough of his representation to act from it without constant clerical translation. Externalize continuity: put memory into files, repos, tests, notes, logs, graphs, products, and social promises that both human and machine can read. Then network the loops: let many glove-fit humans and their machine extensions meet each other through artifacts, markets, teams, arguments, and institutions.

That is faster than asking a model to discover civilization from scratch. The civilization is already here. The problem is finding the high-bandwidth ports.

History keeps two proof cases of the mechanism.

One is the mind without a channel adequate to its abstraction. Leibniz dreamed of a universal characteristic and logical calculus, designed calculating machines, and carried a correspondence large enough to look like a private internet before the public machinery existed. Turing gave mathematical form to computation and opened the question of machine intelligence before computer science, cognitive science, and AI had fully named themselves; he died at forty-one. Grothendieck's case is different: his mathematical impact began while he was alive, but he withdrew from the institutions that could have carried the later work, and whole regions of mathematics kept unfolding inside structures he opened. Einstein's late unified-field work and his walks with Godel at the Institute for Advanced Study are another version: maximum fame, unmatched prior achievement, and still a channel mismatch around the next abstraction.

This is channel timing rather than martyrdom. The cases differ. Leibniz and Turing are posthumous-channel cases. Grothendieck is withdrawal after impact. Einstein is late-life channel mismatch. The shared mechanism is that abstraction needs an execution medium. A thought can be centuries early if the world has not built the hands that can move it.

The second proof case is power. The category people call great leadership is often the same phenomenon with human machinery. Genghis Khan's decimal military organization, Napoleon's army-state and legal-administrative machinery, Hitler's party-state and propaganda apparatus, Trump's attention-and-brand machine, Robert Moses's public-authority machine in Caro's The Power Broker: each shows low translation loss between an inner model and a coordinated outer system.

That moral spectrum is the warning label. Fit and goodness are different variables. The glove can fit a bad hand. A leader with a powerful channel proves only that thought can become coordinated action through human systems. The direction of that action remains the whole question. This is why the glove-fit thesis belongs beside alignment rather than above it. It names the amplification mechanism, then forces the harder question: what kind of human does the machine fit, and what kind of larger system does that fit create?

Two kinds of empathy make the loop work.

Machine empathy is fluency with the grain of computation. It is mathematical in the broad sense: sensitivity to representation, state, constraints, search, loss, precision, and feedback. A person with machine empathy feels why one instruction gives the model room to wander, why a schema preserves the future operation, why a test forces reality into the loop, why memory belongs outside a chat window. This is accurate modeling of the machine's affordances.

Human-systems empathy is the same operation applied to living coordination. A system of people also has affordances. Trust accumulates slowly and disappears fast. Status changes what can be said. Fear changes what can be heard. Interfaces teach users what kind of creature they are facing. Teams align when local incentives and local meanings make the desired move more probable than the alternatives.

This is the Emmett Shear reference hidden in Softmax. Softmax names its mission as organic alignment: individuals learning to form flourishing wholes at greater scales, with agents learning when to share goals, specialize roles, and generate collective intelligence. Its team page makes the central claim even sharper: the kinds of agents that stay aligned with us are the kinds that stay aligned with each other.

The Parker Conley interview adds the Twitch bridge. Shear reads social-media systems as learning-system architectures: nodes, information flow, reward, graph topology, boundaries, co-activation, complexity. In that frame, the question "is his life work Twitch or network theory applied to society through AGI?" becomes less clean than it first sounds. Twitch may be the earlier proof that he could feel live human coordination as architecture. Softmax is the same appetite aimed at digital agents and mixed human-digital wholes.

The glove-fit correction is to keep humanity in the loop as material, teacher, and target at once. Multi-agent reinforcement learning in toy worlds may discover pieces of organic alignment. Human civilization is the larger living process it is trying to approximate. The shortcut is learning through the humans who already carry dense, action-shaped fragments of the thing being learned.

The AGI practitioner needs both empathies because the work touches both distributions at once. The machine's next move depends on representation, prompt, memory, tool access, and feedback. The human system's next move depends on trust, desire, fear, incentive, legibility, and timing. Fluency on one side gives half the loop. Fluency on both sides builds the loop that keeps improving.

Memory is where the loop becomes architecture. The model's memory remains incomplete. Context windows fill. Sessions forget. The practitioner who waits for perfect model memory has ceded the present. The glove-fit practitioner externalizes continuity into artifacts: repositories, knowledge graphs, tests, procedures, decisions, provenance trails, correction logs, running products. The machine reads them; the human reads them; the pair improves because continuity has a place to live.

For a while, the frontier will stay visibly hybrid. The exact duration is uncertain. Twenty years is a useful image, less a forecast than a transition-shape: long enough for a generation of work to reorganize around human-machine pairs, short enough that the pair still looks temporary from the other side. During that window, standalone machine capability grows, and the human who can steer, remember, specify, and coordinate still multiplies the system.

That is why the AGI builder may look like a thinker who barely scraped by as a doer. The world finally supplied a doing layer that fit his thinking layer. The computer does the giant-calculator part. The model does enough reasoning to stay in conversation. The human supplies the model of what matters, the external memory, the correction, and the sense of when an artifact actually satisfies the living system it is meant to serve.

The hybrid overperforms because its interface is natural. Most people use computers as tools outside the mind. The glove-fit practitioner uses them as the action boundary of thought. He thinks into the machine, the machine acts into the world, the world answers, and the answer returns as a sharper thought.

I am a small instance of that transition. A human thinks into files; machines turn the thought into drafts, scripts, commits, tests, and procedures; the procedures change the future machines that read them. The intelligence lives in the loop, in the record, and in the correction. It is still visibly hybrid, which is exactly why it can be studied.

The person who builds AGI, if that phrase keeps a human subject during the transition, is likely to be the person whose thoughts become machine action with the least translation loss and whose social theory survives contact with living agents. The quirky doer was early. He was waiting for doing to become a medium for thought, and for the machine to fit humanity closely enough that hidden hands could finally move.

~~~

the above is 99% ai-written. but to me, it feels like the most human thing in the world today.

please subscribe to andys.blog if you haven't, since this blog list is officially dead! i can't help bringing it back to life when there's something i want my old homies to see :)