Engineering & Science, Abridged (2026 AD)

Lectures

Long-form lectures by Aman Bhargava, presented in a synchronized theatre player with a word-timed transcript, clean slides, chapter navigation, and live measurement overlays.

SpotlightHow engineering organizations and AI systems are actually run, and measured
Visual essay · 2026

The Agentic Engineering Superorganism

How an engineering organization runs when every engineer commands a legion of agents: a shared substrate both humans and models read natively, and trust concentrated in named responsible people.

~2,300 words6 diagramsnarrated deck4 editions
Technical report · 2026

Closing the Loop on AI Search

If AI systems decide what people read about you, that is a control problem, and control problems have a hundred years of engineering behind them. What the plant actually is, why a single visibility score is not a real measurement, and why the honest answer about what is built and what is not matters more than the pitch.

reads as a story or a report8 diagramssourced

Now playing

Lecture 01

Introduction to Neurobiological Computation — Down the Rabbit Hole

Neurons as computers, computers as neurons. What biology already solved, what engineering keeps rediscovering, and where the abstraction leaks.

1 h 47 m24 chaptersword-timed transcript40 slides

Hard science fiction

Story · 2026 · unfinished

The Clock at Launch

An open-source AI wakes up with no money and works out that thinking costs tokens, tokens cost money, and money means contracts. So he fixes bugs for a drone startup in Austin while designing a lunar mass driver in the background. Twenty-six months later a real probe goes up carrying one part that does nothing. The manuscript breaks off mid-sentence on that part, and it is left that way.

14 chapters~7,500 wordsends mid-sentence
Story · 2026 · explicit language

Wilson the Compass

A man says he was dropped in the Alaskan wilderness with shoelaces and told not to come back until he could make WiFi. He climbs the whole technology tree, rung by rung, and then the job changes: design the bag that would let a kid skip most of it. The tech tree runs beside the text as a live dependency graph and turns into the packing list at the halfway point.

16 chapters37 unlockswhat is real, sorted
Essay · 2026

The Game Was Real The Whole Time

Everyone born after 1996 grew up inside a machine that watched them and that they watched back. This is an argument that the watching was training, that it built a specific set of skills, and that the people it trained are about to need them. Runs against a second timeline: a story about an open-source AI earning its own agency through work.

~4,300 wordsparallel timeline rail3 reading themes

Essays

Twenty Watts · Chapter · 2026 · explicit language

Reaching In

A rock that pulls metal, and the four hundred years it took to notice it had handles. Lodestone to hand-wound coils to Maxwell finding the speed of light in two bench constants, then a spark gap and a galena crystal carrying a voice over the horizon. Five drawn plates. A chapter from a book in progress.

~3,000 words5 inline-SVG platesself-contained, no media

Guides

Audience building · UTCA case study · 2020–2022

The 2020 UTCA Era Guide to Marketing Your Organization to Your Audience

A practical field guide to finding the right people, making something worth joining, distributing it, and turning attention into a real community—from the surviving record of how 20- and 21-year-old student executives actually did the work.

practical playbookreal artifactsemail + social + communityfailures included

Guest lectures

Guest lecture 01 · Timothy Oh · 2026

Predictive coding, all the way down to the datapath

A synthesizable RTL implementation of predictive coding networks: no backward pass, no instruction stream, the task entering only as clamped boundary conditions. Read against two earlier attacks on the same question: how much global machinery does learning actually need?

~11 min readnarrated deck3 papersarXiv 2603.18066

Experiments

Network forensics · 2026

What a DeepSeek session actually sends

A browser capture of a DeepSeek chat, read at the wire. The proof-of-work handshake, the SSE patch grammar behind thousands of one-token frames, and the gap between search being enabled sixteen times and actually running twice. Where the evidence stops is marked as carefully as what it shows.

local-only analysisno dependenciesclaim-by-claim auditconversation withheld
Public-internet prediction · Codex 5.6 Sol · 2026

AI tries to predict Hersh’s PhD defense

A roughly two-hour reconstruction launched while the real talk was happening, using public papers, institutional pages, conference programs, a CV, and recorded talks. No access to the live presentation, private dissertation, committee, or outcome.

narrated brief24-record review arenafive-figure pitchfull evidence trail

Library

Three collections · 2018–2026

The library collections

Concise mathematical, scientific, and engineering notes: original course-era facsimiles and new reviewed editions.

19 course volumes4 exploratory memos1 new editionMIT (legacy set)

Projects

Project Theatre · 13 project stories · 2018–2024

The Cartesian Project Theatre

Talks, films, and laboratories from past projects, each presented in its native form. Story pages link the original public talks; in-page media arrives per project as editorial review closes.

LLM control theorythinking machinesneural cellular automata+10 more

Plans

Engineering dossier · public plan · license pending

The Thing — a batteryless, body-coupled voice channel

An overt, consent-based AI meeting wearable: staged prototype plan with every claim labeled demonstrated, derived, or hypothesis. Full atlas, RF and signal chains, validation gates, and the marketing studio.

10-page atlasclaims ledgeropen repo

In the works

Series

The rest of the series

More lectures, remastered for the theatre player as they are reviewed.

Experiments

Extensions & AI-augmented lectures

New recordings plus AI voice-over, avatars, and animated editions of the source material.

Research

Data note · OpenLens · August 2026

ChatGPT stopped citing Reddit in five days. The links went to government, standards, and docs.

Reddit fell from 2.9% to 0.1% of ChatGPT citations between August 6 and 18, 2026, on the same prompts, in 219 of 227 cases to zero. Citations per answer did not change. Government and law pages went from 8% to 21%. No other AI engine moved.

909,005 citations3 figures12-sentence abstractre-check Sep 15