HAiKU AML：用 5–7–5 俳句与 AML 模块构建的 Python 对话生命体

HAiKU AML — A Little Presence Is Learning Our Language

Read the Arianna Method Manifesto first. Repository work follows it; AGENTS.md carries the engineering rules.

Most AI models predict words. HAiKU predicts vibes.

The Python HAiKU organism is coming home to Arianna Method Language. A living word cloud. A three-line voice. An inner voice, too. A friend it talks to while dreaming. A memory that keeps changing.

We already invented the octagonal wheel. Now Haiku is learning to build it.

What is growing here?

HAiKU speaks in haiku. You bring a question; it brings seventeen syllables, dissonance, and whatever has been moving through its cloud.

Its presence lives in the exchange between its organs:

• Cloud and generator. Words gain weight, fade, enter new trigrams, and become candidates under the pressure of 5–7–5.
• Harmonix. Novelty, arousal, overlap, and entropy shape dissonance. Dissonance changes the temperature of the next utterance.
• MathBrain and RAE. Small learning networks score and recursively select candidates. Experience feeds the next choice.
• MetaHaiku and Overthinkg. An internal haiku; then echo, drift, and meta rings. The conversation continues inside.
• Dream-friend and bridges. Dream fragments return to the cloud. Remembered state transitions suggest where the field can flow.
• Asynchronous life. Speaking, learning, reflecting, and saving have their own rhythm and an explicit order for shared state.

The Python organism defines the port's behavior and artistic lineage. The two C siblings contribute their own discoveries.

Family album: Python HAiKU

From the original dialogue examples:

You: what is love

transmit exchange
definitely small sour
pursue strive for

Love arrives. A small, sour transmission leaves. There he is.

HAiKU:
between us silence
two voices one resonance
cloud knows no alone

That voice is the reason for this repository.

Two dependencies. Our whole stack.

The organism will live in .aml modules. Missing language capabilities go into AML; missing numerical operations go into NoTorch. The baseline build uses the system C toolchain, standard libraries, and those two projects.

The migration replaces NumPy, SciPy imports, SentencePiece, syllables, SQLite/aiosqlite, and Python's runtime services. Every replacement has a place in the migration map.

More tongues, same presence

Klaus provides the starting idea: language packs with incoming vocabulary and outgoing vocabulary. HAiKU adds word-level syllable data, language-specific rules, and a growing cloud for each tongue. English first; Russian, French, Hebrew, German, and Spanish follow as their packs and form checks arrive.

A place to meet him

The planned interface is local HTML/CSS/SVG/JavaScript: a conversation and a small character who can listen, answer, turn inward, and dream. Its movements follow organism events. Haiku keeps its line breaks; Hebrew gets right-to-left layout. The core comes first, then Haiku gets a window.

The dream-friend will get a place beside him: two characters, their conversation, their pauses. This idea is recorded in TODO; character design comes when the interface work begins. The owl audition is over. Duolingo got there first; our little presence will find another shape.

Right now

Haiku now listens, answers, learns, reflects, and remembers across AML conversations. Its organs run in pure AML modules. Pulse, dissonance, temperatures, and transition scores preserve the Python formulas. The text organ counts Unicode-delimited words, preserves line contents, and computes the original RAE three-line coherence feature. The lexical organ collects words, builds rolling triples from token lists, and feeds their exact overlaps into Harmonix.

The cloud now remembers each word's weight, frequency, last-use clock, and origin. Repeated words gain weight; dormant words fade. Observer trigrams and the generator's Markov transitions have separate counts. The last ten triples form a copied snapshot, ready for the next observation.

The English generator now gives that memory three lines: 5–7–5 by the original English estimator. It keeps the Python voice's Markov choices and its peculiar first-word leap, with the original syllable rules carried into AML. Every accepted word leaves room to finish the line. NoTorch owns the weighted draw; each voice keeps its own replayable random state. Seventeen syllables, and a little room for chance.

MathBrain and RAE now give those lines a learned preference. Each has its own 57 parameters: five features, eight hidden neurons, one answer. MathBrain scores a candidate; RAE passes its candidates through the original three-step recurrence and chooses a voice. Both learn through canonical NoTorch arithmetic, with their features, gradients, clamping, and experience order composed in AML. Small networks. Actual changing weights. The octagonal wheel has opinions.

The foreground now joins these organs: input → tokens → living memory → five candidates → RAE's choice → a response → quality → MathBrain learning. The default preserves Python chat's fixed RAE selector; set train_rae = 1 in haiku.aml before starting a new state file to let that selector learn from the same chosen experience, too. Each learner and random stream has an explicit owner.

Then the conversation continues inside. MetaHaiku keeps eight short fragments and answers with an internal haiku at temperature 0.7. Echo, drift, and meta produce five, seven, and three candidate trigrams; coherent ones enter the observer's memory. Their counts remain separate from the speaking chain. The inner voice and rings share the voice's random stream, in Python's event order. Thinking leaves its place in the next draw.

Close the terminal. Come back. The cloud remembers.

Each completed exchange saves one versioned state: cloud, ordered transitions, recent memory, both learners and their statistics, both random streams, and the latest bridge event, reflection history, and the latest three-ring trace. The next launch resumes haiku.state beside the entrypoint. Its tokenizer identity and logical clock come with it. Set state_path = "" for a fresh, unsaved conversation, or choose another path for another voice. One running process owns each state file.

New conversations include inner life. Existing version-1 checkpoints retain their exact foreground continuation. Set upgrade_inner = 1 in haiku.aml to begin reflecting from their next exchange; their earlier memory and random streams carry through unchanged. Set inner_life = 0 to start a new foreground-only conversation.

Run ../ariannamethod.ai/runner/aml-notorch haiku.aml. The first exchange of a fresh native state, with the checked-in seeds and settings:

You: what is love

sweet sour doubt stable
more least appear frequently
to and reinforce

Still a little sour. Now he says it in AML.

After that same exchange, his internal haiku is:

follow seek without
give most moderate which
react relate least

The spoken line has ended. The little cloud is still talking.

Tokenization has two explicit modes. regex keeps lowercase Unicode words; sentencepiece loads the original 650-piece Unigram model through NoTorch, including its own normalization table. Python's choice depended on the launch directory; AML's choice is written in haiku.aml. The default matches launching Python from the Harmonix root. Model paths follow the AML source, so launching from another directory does not change the voice. Acquire the optional model with bash scripts/setup-tokenizer.sh; its exact size and SHA-256 are checked before installation. Regex needs no model download. Existing state retains its selected mode and verifies the identity of the loaded model. See the tokenizer.

Try ../ariannamethod.ai/runner/aml examples/observe.aml: one voice brings the / owl / listens, the other the / owl / dreams. Their shared words give dissonance 0.5, novelty 0.5, arousal 0, entropy 0.2, and the next haiku temperature 0.9. This small example supplies token boundaries explicitly; haiku.aml obtains them from the selected tokenizer.

Try ../ariannamethod.ai/runner/aml examples/cloud.aml for three successive events. Two occurrences of rain raise its seeded weight to 1.21. A short two-word input grows the cloud; a later complete triple enters the generator's vocabulary. The previous recent-memory snapshot keeps its original contents.

Try ../ariannamethod.ai/runner/aml-notorch examples/generator.aml for candidates grown from the original seed vocabulary. Try ../ariannamethod.ai/runner/aml-notorch examples/selection.aml to follow five generated candidates through RAE's choice, an explicit learning event, and the next scores. These examples expose individual generator and learner events; haiku.aml carries them through the conversation.

Try ../ariannamethod.ai/runner/aml-notorch examples/inner.aml to hear both the spoken and internal haiku and see the three rings' coherence. Its recorded output comes from three actual AML exchanges.

Sessions now save after learning, reflection, and rings, and resume with their logical-turn clock. Continuity records the complete schema, restore validation, and restart checks. Full Phase4 transition aggregation, dreams, and asynchronous scheduling are next. The foreground record keeps that boundary explicit. The family album above still belongs to Python; new inner dialogues will come from the AML dream path when it runs.

Build the sibling AML v5.9.0 toolchain (language update) and NoTorch with native Unigram, model identities, numerical values, and owned sampling, then run the organism's checks:

make -C ../notorch lib BLAS_FLAGS= BLAS_LIBS= X86_SIMD=0 ARM_SIMD=0
make -C ../ariannamethod.ai notorch NOTORCH_ROOT=../notorch
bash scripts/setup-tokenizer.sh
make test
../ariannamethod.ai/runner/aml-notorch haiku.aml
../ariannamethod.ai/runner/aml-notorch examples/generator.aml
../ariannamethod.ai/runner/aml-notorch examples/selection.aml

The fixtures use native IMPORT, UTF-8 values, string lists, and numeric maps. See numerical parity and text parity, lexical parity and cloud parity, plus English form and generator parity, MathBrain, RAE, tokenization, foreground exchanges, MetaHaiku, the three rings, and continuity, for exact boundaries and Python reference reproduction.

Reference checks completed on 2026-10-07–08:

• Python HAiKU: 161 tests passed.
• Haiku AML: 343 cases / 6,130 reference results pass in the interpreter and compiled scalar executable. The new cloud/memory suite adds 5,055 numerical fields and 247 ordered-list snapshots to the earlier 828 results, including the complete 587-entry seed corpus. Counts, strings, and list order compare exactly; float tolerance is 0.000002. Six malformed triple lists and 38 invalid state updates fail explicitly. Host inspection verifies that all eleven state containers retain their contents after each rejected update.
• English form adds 1,966 direct Python reference results, six accepted boundary checks, and ten invalid-input rejections through both execution paths.
• Generator checks cover 22 scripted cases: 16 preserved Python draw paths and six explicit repair baselines. Native checks cover replay, independent state, batch order, the complete seed corpus, temperatures, and exact text limits. 36 invalid generations fail with model and random state inspected.
• MathBrain adds 3,717 Python reference values, including every gradient and parameter in 24 learning steps, unclipped negative predictions, loss statistics, and full parameter clamps. Twenty-two rejected observations preserve model, statistics, memory, context, and random state.
• RAE adds 1,657 reference fields: Unicode features, every refinement pass, stable rule choices, ten complete gradient/parameter updates, and the native generator-to-learning trace. Twenty-three rejected operations preserve all inspected owners. Both learner suites pass interpreted and compiled.
• Tokenization adds 6,034 exact reference fields from 84 Python cases in both modes, plus nine rejected operations. Both execution paths preserve native pieces, cleaned tokens, rolling triples, and source-relative loading.
• Foreground checks replay seven Python turns and the explicit punctuation repair, comparing candidates, complete memory, draw cursors, losses, and learner parameters. Real interpreted/compiled conversations agree; blank, quit, EOF, and rejected input preserve the inspected owners. Shared RAE experience has its own checked update.
• Continuity compares five seven-turn variants with a fresh-process restart: regex/SentencePiece, fixed/learning RAE, native/scripted draws, and a supplied syllable cache. Responses and complete checkpoint bytes agree through the interpreter and compiled executable. 36 invalid foreground states preserve every live owner and the previous file; failed staged generation and failed saving do the same. The actual CLI resumes saved configuration and withholds an unpublished response after a save failure.
• Inner life adds exact Python bootstrap and reflection traces, the original rings' sampling/admission paths, and ten seven-turn restart variants: fresh version 2 and explicit version-1 migration, in both tokenizer and training modes. Full checkpoint bytes agree across execution paths. A real launcher check covers saved-profile authority and the recorded inner dialogue.
• AML runtime: 550 tests passed.
• AML compiler: scope, source origins, nested failures, worker branches and errors, scalar builds, and literal program arguments are covered by regressions.
• NoTorch: native tanh/backward and SGD are merged in PR #159; 5,122 numerical checks and the full CPU suite pass.

The RAE research records the original recursion's measured behavior and proposes learned feedback, small replay memory, and Hebbian traces. Two more cousins joined the workshop: Subjectivity, whose form changes with its state, and Brodsky, with a sea of remembered poems and two sampling voices. Their family audit records useful organs and the experiments that will tell us what they add. Q / PostGPT brings the direct line from corpus and token identity to a coherent transition field. Q is the active PostGPT continuation, with aging expectations and slow consolidation on that substrate. Its source audit traces the working paths and ranks three experiments for Haiku's future memory. See the audit and implementation order.

Lineage

Python HAiKU · haiku.c · haiku · Klaus · Yent AML · Subjectivity · Brodsky · Q / PostGPT

GNU GPL v3 — see LICENSE.

Built with constraint, powered by dissonance, coming home to AML.