xonecas.io
768
La

learning-ai-lua-code-model / commit ed2252d

feat: adds discussion and experiement idea about RLM

xonecas2026-06-16 ed2252d740479855bceb710e1bfa7d6d146602c7 parent ebb854b

3 files changed

README.md+1−0
@@ -115,6 +115,7 @@ The script can load the current bigram, MLP, and GPT checkpoints.
115115 - `notes/005-session-handoff.md`: current project state, learned concepts, run commands, and recommended next steps
116116 - `notes/006-inference-sampling.md`: loading checkpoints and sampling without training
117117 - `notes/007-distillation-future-experiment.md`: planned teacher/student distillation experiment
118+- `notes/008-future-rlm-harness.md`: future RLM-style harness/sampler experiment
118119
119120 ## Next likely step
120121
notes/008-future-rlm-harness.md+40−0
@@ -0,0 +1,40 @@
1+# Future experiment: RLM-style harness
2+
3+This is a reminder for a much later experiment inspired by the Recursive Language Models paper.
4+
5+## Idea
6+
7+Build a small educational harness around one of our Lua language models or a stronger external/local model where:
8+
9+```text
10+large input corpus lives outside the model context
11+model/sampler receives metadata and selected snippets
12+code handles chunking, looping, storage, and aggregation
13+semantic work happens through model calls
14+```
15+
16+The point is to explore the systems idea:
17+
18+```text
19+do not load everything into the model context;
20+keep a handle to the data and process it in chunks
21+```
22+
23+## Possible toy version
24+
25+Use a Lua corpus or documentation bundle and ask questions like:
26+
27+- find sections about tables/metatables
28+- summarize all examples of functions
29+- collect idioms from many small Lua files
30+- compare generated summaries from chunked vs direct prompts
31+
32+## Learning goals
33+
34+- understand context windows as limited working memory
35+- practice chunking and map/reduce-style processing
36+- separate exact code operations from fuzzy model calls
37+- see where recursion/sub-calls help or become expensive
38+- connect classic software memory-management intuition to modern LLM systems
39+
40+This is intentionally a future milestone. We should first keep improving the basic training, sampling, evaluation, and distillation experiments.
pi-session-2026-06-16T08-14-44-950Z_019ecf7f-44d6-7517-9eda-a8d6da0293ca.html+1−1
@@ -1158,7 +1158,7 @@
11581158 </div>
11591159 </div>
11601160
1161- <script id="session-data" type="application/json">eyJoZWFkZXIiOnsidHlwZSI6InNlc3Npb24iLCJ2ZXJzaW9uIjozLCJpZCI6IjAxOWVjZjdmLTQ0ZDYtNzUxNy05ZWRhLWE4ZDZkYTAyOTNjYSIsInRpbWVzdGFtcCI6IjIwMjYtMDYtMTZUMDg6MTQ6NDQuOTUwWiIsImN3ZCI6Ii9ob21lL3hvbmVjYXMvc3JjL2xhbmd1YWdlLW1vZGVscy9sdWEtY29kZS1tb2RlbCJ9LCJlbnRyaWVzIjpbeyJ0eXBlIjoibW9kZWxfY2hhbmdlIiwiaWQiOiJjNzMxODIwZCIsInBhcmVudElkIjpudWxsLCJ0aW1lc3RhbXAiOiIyMDI2LTA2LTE2VDA4OjE0OjQ0Ljk2NVoiLCJwcm92aWRlciI6Im9wZW5haS1jb2RleCIsIm1vZGVsSWQiOiJncHQtNS41In0seyJ0 …
1161+ <script id="session-data" type="application/json">eyJoZWFkZXIiOnsidHlwZSI6InNlc3Npb24iLCJ2ZXJzaW9uIjozLCJpZCI6IjAxOWVjZjdmLTQ0ZDYtNzUxNy05ZWRhLWE4ZDZkYTAyOTNjYSIsInRpbWVzdGFtcCI6IjIwMjYtMDYtMTZUMDg6MTQ6NDQuOTUwWiIsImN3ZCI6Ii9ob21lL3hvbmVjYXMvc3JjL2xhbmd1YWdlLW1vZGVscy9sdWEtY29kZS1tb2RlbCJ9LCJlbnRyaWVzIjpbeyJ0eXBlIjoibW9kZWxfY2hhbmdlIiwiaWQiOiJjNzMxODIwZCIsInBhcmVudElkIjpudWxsLCJ0aW1lc3RhbXAiOiIyMDI2LTA2LTE2VDA4OjE0OjQ0Ljk2NVoiLCJwcm92aWRlciI6Im9wZW5haS1jb2RleCIsIm1vZGVsSWQiOiJncHQtNS41In0seyJ0 …
11621162
11631163 <!-- Vendored libraries -->
11641164 <script>/**