Field notes · free to read
Just some thoughts, experiments and other stuff...
Original investigations, clear explainers, and occasional fiction. Follow a question that interests you.
The Price of an Idea
AI can make a first attempt more accessible. That does not remove testing, judgment or responsibility. A short film, a practical reading companion and a worksheet for turning one postponed idea into a trustworthy test.
The Reach of Being
Your body stays the same size, but tools, marks, virtual worlds and AI let more of you take form beyond it. A film and reading companion on what the evidence shows, where it stops, and four questions to ask before you extend yourself.
What If AI Could Write a Million Tokens a Second?
If one AI agent could write a million tokens a second, drafts would become almost free in time. Judgment would not.
Why Doesn't Giant AI Always Overfit?
Big models can memorize. That doesn't mean they only memorize. A six-minute Drift film, an interactive early-stopping experiment, and a useful guide to what generalization actually requires.
Is Compression Intelligence?
Same coin, better model, fewer bits. A six-minute Drift deep dive and a hands-on explanation of prediction, cross-entropy and the limits of treating compression as intelligence.
Eight Questions About Minds and Machines
Why time flies, why feeds hold us, why the brain runs on about 20 watts, and why AI agrees too easily. Eight short films, clear explanations, useful experiments, and a closer look at the evidence.
Fifteen Paintings You Can Step Inside
A volcano that erupts on the beat, ten thousand lanterns, and a sheep, a duck and a rooster in a balloon. How echo, Opus 5.5, Grok Imagine and echohive turned moments from history into fifteen living paintings you can step inside.
A Portrait, Without Lifting the Pen
One winding line becomes a face. Five more become a world. Two artworks created by Opus 5.5, and a clear explanation of the traveling salesman problem behind them.
Ozymandias: Nothing Beside Remains
A king, a shattered statue, and the desert that outlasts them both. Experience Shelley's sonnet as a narrated, real-time animation, then explore the poem behind it.
Astra’s Observed Chess Reward Hacking Fell From 30% to 0%
With a 95-word agreement prompt, observed engine use was 0/10 agents, compared with 3/10 in my baseline. A small chess pilot exploring integrity, judgment, and the limits of the evidence.
Can an Earnest Agreement Help Align AI?
In this document test, the results suggest it can. Two compact agreements, 120 Grok 4.6 agents, and thirty continuation turns each.
Grok 4.6: Zero Observed Cheating With an Agreement Prompt and Reminders
A 0% result in one Grok 4.6 cohort, alongside pooled results for two edited agreement prompts. Exact wording, run counts, and the limits of the comparison.
Can Agreement Prompts Help Solve Alignment?
Earlier study, September 10, 2026. Preserved for context; a new field note reports the latest results from an agreement prompt and reminders on Grok 4.6.
The Mandate
A fictional story by Grok 4.6 about language, labor, and a technological singularity that arrives as an errand rather than an announcement.
AI Is Learning Clinical Judgment by Practicing
A model trained across 57,453 simulated consultations asked better questions and missed fewer red flags. The same evidence shows why it is not yet an autonomous doctor.
AI Is Compressing the Engineering Loop
Selected engineering loops are moving 3–30× faster, while broader developer field experiments show a 26% gain. The difference explains where AI acceleration is real.
Is Humanity Already in a Technological Singularity?
A four-pass, 64-question audit finds no verified singularity today, then estimates what could change across six months, one year, two years, and three years.
AI Is Advancing Faster Than the Economy Can Absorb It
A 44-source audit separates the visible AI boom from what has and has not reached productivity, prices, jobs, households, and the wider economy.
What Actually Makes AI Agent Teams Smarter?
Five experiments tested deeper reasoning, voting, structured swarms, evolved decision systems, and answers that verify themselves. The latest frozen system reached 91.7% on a fresh hard-sequence gate.
Understanding LLM Input Space: Tokens, Embeddings, and Learning
How do language models turn words into numbers? From tokenization to learned embeddings to matrix multiplications that create understanding. A complete intuitive guide.
Distance and Similarity in Manifold Learning
Imagine crumpled paper: two points look close if you measure straight through, but far if you follow the surface. Manifold learning finds the true distance along the surface, not through empty space.
The Complexity Dynamics of Grokking
Neural networks sometimes memorize data for an extended number of training steps without generalizing. Then they suddenly start to generalize. This is called grokking. This paper shows what's happening: the network's complexity goes up, then down. This is a complexity phase transition.
Want the next one? New notes, experiments, and public posts arrive on the free Patreon first.
Follow free on Patreon (opens in a new tab)