Research into how models actually learn — architecture, training, memory — and an attempt to build something that learns like a child. Draw an architecture out of tensor primitives, compile it straight to the GPU, and watch the neurons light up.
Most people use models. This is about understanding them — from the tensor up. If you can draw an architecture, compile it, and race it against another, you build real intuition for why one design wins and another doesn't.
One app where people and AI draw, document and run systems from the same shared Views. Built for Windows, Linux and Mac.
Hands-on AI training built into the app, aimed at graduates and people over 50 who are finding it hard to get into work. Learn AI by building real things with it.
Knowledge written as small, checkable XML Views instead of paragraphs: English, maths, physics, chemistry and biology. Generators add more every night.
Models learn from text that was written for people. Most of it is noise for a model. Thinking DNA writes knowledge as small structured Views instead: the facts, the steps, and a check. For maths, physics, chemistry and biology, Python generates the Views, so every answer is computed and then checked a second, independent way before it is kept.
Early result. A small model trained to read a View, rather than memorise its answer, got 100% on facts it was never trained on, including 10-digit sums, when it was handed the exact View. When it had to pick the right View out of four similar ones, it fell to 26%. So the job splits in two: a deterministic lookup fetches the exact View, and a small model reads it.
Next. Train small models from scratch on the same knowledge as Views and as prose, and compare what they learn and how fast.
A tiny transformer’s neurons placed on a sunflower (Fibonacci) sphere, lighting up as it reads a sentence. Turning the vectors inside a model into a picture you can actually watch.
A deterministic, human-owned loop that keeps a model learning: it trains, tests itself, rolls back mistakes, prunes memory, and sleeps to consolidate — while the loop itself can never be rewritten by the model. Status: the train-and-test loop runs today; rollback, memory pruning and sleep-consolidation are designed and not all running yet.
Attention drawn by hand from three matmuls and a softmax, raced against the same network without it. On matched compute, attention won — val loss 2.40 vs 2.50. A real result from a hand-drawn network.
Papers pulled from arXiv, rebuilt as runnable architectures and raced against a matched transformer baseline at two scales, negative results included. Paused at the end of August to focus on Thinking DNA; kept here as they ran.
Every comparison holds these fixed: parameter count matched within a stated tolerance, tokens seen rather than steps, same tokenizer and same train/val split, three seeds with the spread reported. A run that cannot meet these is reported as failed rather than compared. Wall-clock and FLOPs-per-token are logged alongside each result — matched parameters is not matched FLOPs, and the FLOPs column shows which. These are fixed-budget results, not converged ones.
The first table publishes this week.
In progress. The pieces below are being assembled into one system. Some are running, some are designed and not yet built.
The pieces are being assembled into one living system: a “brain” and its organs. The rule that keeps it safe: three kinds of memory (the shared model weights, private per-user memory, and short-term working memory), and only the human-owned Control Plane ever writes to the shared weights.