Research prototype · MIT License

0GLM / Granular Language Model

Coherent speech grown on a laptop, from a corpus of just 16,535 documents — with every sentence traceable to its source, by design.

0GLM reads Granular Language Model — what it does today — and Granular Living Model — the class it is becoming: grains as living states of a navigable field, language one readout among possible projections.

16,535documents in corpus
1.27Mtext grains in field
∼15 minnavigator training, MacBook GPU
100%sentences with named provenance
Examples

Real outputs of the current build (verbatim excerpts from generated answers; ellipses mark cuts). Each answer is assembled by riding the author's own order inside retrieved documents — coherence is inherited from human writing.

How does climate change affect the ocean?
Besides the complexity of coupled interactions between the atmosphere, ocean, ice and land, which can lead to climate change, there is the complexity of understanding the dynamic role of each of the above climate components. What gives the oceans the potential for exerting a strong influence on climate and its variations is the large heat capacity, coupled with the global redistribution of heat and other properties by ocean currents acting over time scales much longer than atmospheric processes. An added complexity is that ocean eddies (e.g., Gulf Stream rings) could be important mechanisms for transporting heat.
22 steps · 4 documents · mean cos 0.440 / max 0.762
Opens with the mechanism (heat capacity → currents → eddies), not an adjacent fragment.
What is a Python dictionary and when should I use one?
A dictionary is a container that stores Python objects and each item of the dictionary consists of a key and value. Python dictionaries are also known as associative arrays or hash tables. Curly brackets distinguish a dictionary from other objects. … The keys of a dictionary should be unique while values may not be.
22 steps · 4 documents
Polysemy case fixed by corpus density: retrieval used to land on a GNU Dico server manual (cos 0.319 → 0.490). Root cause treated, not patched.
How do I write a resume that stands out?
It is important to keep your resume concise. Your resume outline should not exceed two pages. … Most people misunderstand the purpose of the resume. Its true aim is simply to get an interview, not a job. Most people write one resume and send the same version for all job queries. Successful job hunters tailor-make their resumes for each job application.
23 steps · 4 documents
Why do humans need sleep?
Extra sleep helps replenish the brain and boost energy, alertness and attention span after a week of restricted sleep. … Despite decades of research, why we sleep is still a mystery. Theories range from conserving energy, to processing information and memories, to removing toxins that build up when we’re awake.
24 steps · 4 documents · mean cos 0.428
What happens at the event horizon of a black hole?
…whether information behind the event horizon of a black hole is irretrievably lost, or whether it is possible to recover that information from the radiation emitted by the black hole as it evaporates. As the formal wager phrased it: “When an initial pure quantum state undergoes gravitational collapse to form a black hole, the final state at the end of black hole evaporation will always be a pure quantum state.”
24 steps · 4 documents · mean cos 0.366
Reaches the historical Hawking–Thorne–Susskind information-paradox debate.
Ride is not “retrieve chunks and glue them”

Retrieval exists here too — but only to choose where to enter and when to leave. No text is recombined at answer time.

classic RAG0GLM ride
unit of selection a chunk picked by similarity a document trajectory, entered at a chosen grain
order of material chunks re-ranked and glued by score the author’s own sentence order, verbatim
transitions every boundary is a stitch the decoder must survive boundaries only between phases; inside a phase zero join decisions
coherence source hope: prompt instructions + decoder skill inheritance: humans already wrote it in order
Traceability by design

Mechanistic interpretability reads trained networks after the fact. Here the navigation layer is legible up front: --trace prints every candidate, anchor and rejection reason. Real output for the Arctic question:

🔍 HOW THE ANSWER WAS BUILT
  document stage: top-16 docs
  # 4575 dsim=0.675 tail=0.643 fam=1.000 ✅ ANCHOR «causing the temperature to rise further…»
  #  933 dsim=0.656 tail=0.626 fam=0.569 ✅ ANCHOR «but with less ice cover the Arctic Ocean absorbs…»
  #  524 dsim=0.656 tail=0.632 fam=0.645    (lost MMR) «swiftly warming air and ocean temperatures…»
  #12496 dsim=0.633 tail=0.539 fam=0.523    (lost MMR) «The Arctic has been warming disproportionately…»
  #14552 dsim=0.610 tail=0.595 fam=0.550 ✅ ANCHOR «They expect warming to accelerate again…»
  #10361 dsim=0.581 tail=0.571 fam=0.567 ✅ ANCHOR «whereas the Arctic is an ocean surrounded by…»

Post-generation, every sentence maps back to a (document, position) pair in the corpus. This is not probing — it is the architecture being read.

Why it matters
Second modality: audio

The same trajectory renders to sound without any separate pipeline: stylometric transitions map to audio targets via a learned bridge, and the navigator's predicted direction steers each transition (amplitude stays factual). On blind paired listens the z-steered render was preferred in 7/7 cases by the author listener — an independent panel is the open next step. Diagnostics: predicted style direction matches real document motion at mean cosine 0.607 (19/22 boundaries positive).

Baked brain v3: one file, zero retrieval

The declared next stage shipped: brain_v3.pt (1.9 GB) bakes the planner, the granule bank and quality masks into a single artifact. At inference there is no corpus lookup at all — the field navigates baked granules directly. Trained by distilling the ride teacher: 53K demo triples with correction moments plus doc-discriminative negatives ("continue this document's line, not a paraphrase from a neighbor").

blind A/B vs its own teacher
brain v0–v40 : 5 — knowledge baked, behavior not
brain v2 (distillation)2 : 3 — wins multi-point assembly (El Niño), thin-corpus topics (Python dict)
brain v3 (doc-negatives)2 : 3 — question alignment flat at cos .41; loses where one long authoritative passage is required

Honest limitation, stated as architecture not bug: the baked planner is local — it picks the next granule from (context, question) only; the teacher knows in advance that twenty granules of one document form a coherent passage. Roadmap v7: a document-level plan token. Weights + full bake pipeline: huggingface.co/0penAGI/0glm.