DOGMA / Start Here

DOGMA in One Hour

One-hour guide3 min702 wordsMarkdown source

DOGMA is not DNA stored in a computer, not a wet-lab molecular computer, not a renamed transformer, and not AGI. It is an experimental architecture and evaluation program. “DNA-inspired” describes a source of hypotheses; it does not prove that the hypotheses work.

DOGMA Pipeline

Six local stages replace one monolithic attention operation

Advance through the canonical non-transformer computation from sequence to realized output.

Genome

Encode

HelixHash creates structural sequence representations.

G = H(sequence)

The Four Layers

DOGMA becomes easier when four different subjects are kept separate.

1.: Molecular biology studies how cells store and express information. DNA is transcribed into RNA; RNA can be translated into protein. Regulation determines what is expressed, where, and when.
2.: Molecular computing constructs physical computations from molecules. Adleman encoded a graph problem in DNA [Adleman, 1994]; strand-displacement systems later implemented logic and neural-like circuits [Seelig et al., 2006, Qian and Winfree, 2011].
3.: Genomic machine learning learns statistical structure from biological sequences. Transformers and state-space models are useful baselines here, but they are software models of sequence data, not molecular computers.
4.: DOGMA is a software research architecture inspired by genomic organization. Its current canonical line deliberately avoids transformer self-attention. It asks whether local interaction, recurrence, selective expression, persistent bounded state, and evolutionary search can produce useful learning behavior.

What DOGMA Is Not

One Worked Example

Suppose the byte sequence ACGTACGT enters a small DOGMA model. Overlapping four-symbol windows are \[ (A,C,G,T),\ (C,G,T,A),\ (G,T,A,C),\ (T,A,C,G),\ (A,C,G,T). \] Each window produces a local representation. A causal convolution mixes nearby windows; a recurrent state carries information forward; a gate chooses how strongly each state is expressed. The model predicts the next byte. This is a neural computation on digital symbols. No molecule is synthesized.

If the correct next byte is \(C\) and the model assigns it probability \(0.5\), the loss is \[ -\log (0.5) \approx 0.693. \] Across \(N\) held-out bytes, average cross-entropy \(\bar {\mathcal {L}}\) gives perplexity \(\operatorname {PPL}=\exp (\bar {\mathcal {L}})\). Lower perplexity means better teacher-forced prediction on that split. It does not guarantee readable free generation, factual answers, reasoning, or general intelligence.

The Recursive Research Loop

A DOGMA candidate is never allowed to rewrite production weights directly. The controlled loop is: \[ \text {parent} \rightarrow \text {one bounded mutation} \rightarrow \text {train} \rightarrow \text {frozen probes} \rightarrow \text {archive or promote}. \] The candidate must retain real-data anchors, avoid evaluation overlap, run with multiple fixed seeds, and pass separate sequence, explanation, and self-critique probes. A quality-diversity archive retains useful stepping stones; only a reproducible global improvement replaces the live checkpoint.

A Twelve-Week Route

Week	Reading	Question to answer
1	Chapters 1 and 2	How do biological and molecular computation differ?
2	Chapter 3	What does regulation compute?
3	Chapter 4; Appendix A	What do loss, gradient, and perplexity measure?
4	Chapter 5	What problem does self-attention solve?
5	Chapter 6	How are genomic foundation models evaluated?
6	Chapter 7	Which DOGMA mechanisms are canonical, historical, or hypothetical?
7	Chapters 8 and 10	How can local structure become a computational prior?
8	Chapters 9 and 11	How are regimes, types, and programs represented?
9	Chapter 12	Why separate exploration, archives, and promotion?
10	Chapters 13 and 14	Why can lower perplexity coexist with worse generation?
11	Chapters 16 and 17	What evidence would count as progress toward AGI?
12	Chapter 18	Reproduce one baseline and write a failure analysis.

Evidence Checklist

Before accepting any DOGMA claim, ask:

1.: Is this a biological fact, an external result, an implementation statement, a measured DOGMA result, or a hypothesis?
2.: Are the data source, split, checkpoint, evaluation version, and random seeds recorded?
3.: Is the comparison matched for parameters, training bytes, and compute?
4.: Does the evaluation test free generation and downstream behavior, not only teacher-forced likelihood?
5.: Could contamination, cherry-picking, or synthetic-data feedback explain the result?

Starter Exercises

S.1.: Explain in two sentences why the central dogma is a biological theory, not proof that DOGMA will work.
S.2.: Compute perplexity for average losses \(0.5\), \(1.0\), and \(2.0\).
S.3.: Design one probe for canonical DNA generation and one probe for a readable biological explanation. State exactly what counts as a pass.
S.4.: Give an example of a candidate that belongs in the archive but should not replace the live model.