Skip to content

Latest commit

ย 

History

History
160 lines (120 loc) ยท 8.07 KB

File metadata and controls

160 lines (120 loc) ยท 8.07 KB

OCCAM โ€” ์ž์—ฐ๋ฐœํ™” minimal-viable ๋ฒ ์ด์Šค๋ผ์ธ strategy

History โ†’ ./OCCAM.log.md.

frame: S187 saga ๊ฐ€ over-constrained ์ž„์„ ์ธ์‹. 7-aux loss + 8.92B params

  • bf16 + byte-level vocab + bnb int8 + custom ConsciousDecoderV2 + mitosis hook + ฮป ร— 6-dim grid + multi-step horizons โ€” ๋™์‹œ์— ๋งž์ถ”๋ ค๋‹ค ๋ชจ๋“  path ๊ฐ€ CE 4.06-4.09 floor ์— ๋„์ฐฉ. Occam ๋ฉด๋„๋‚ : ํ•œ axis ์”ฉ strip ํ•ด์„œ ์ง„์งœ binding constraint ์ฐพ๊ธฐ.

status: ๐ŸŸก design tier โ€” brainstorm ๊ณ ๊ฐˆ ์‹œ์  1์ฐจ reset. ์ฒซ fire ํ›„ ๊ฒฐ๊ณผ ๋ฐ˜์˜ํ•ด์„œ ์šฐ์„ ์ˆœ์œ„ ์žฌ๊ตฌ์„ฑ.

g3: minimal viable baseline ๋ถ€ํ„ฐ ์‹œ์ž‘ โ€” ๊ฐ€์žฅ ๋‹จ์ˆœํ•œ condition (CE-only, small scale, vanilla arch) ๋ถ€ํ„ฐ emergence ํ™•์ธ ํ›„ 1 condition ์”ฉ add. capability claim 0, GOAL ๋ฏธ๋„๋‹ฌ carry, north-star ๋ถˆ๋ณ€.


1. ์™œ OCCAM ํ•„์š”ํ•œ๊ฐ€

S187 attempt10 + Eval + S187-B/C/F/G/H/J/K ๋ชจ๋‘ land ํ›„ ๋ฐœ๊ฒฌ:

path finding
Eval 1 (์ž์—ฐ๋ฐœํ™”) 5/5 cells whitespace collapse at 2000 step
S187-H 50000 step ๋„ CE 4.09 plateau (horizon-limit ์•„๋‹˜)
S187-J bsz 2โ†’8 (4ร— tokens) CE ์•…ํ™” 4.06 (token-starvation ๊ฐ€์„ค falsified)
S187-K bsz=8 + lr 1.2e-3 linear scaling ๋„ CE 4.06 (LR-mismatch ๊ฐ€์„ค falsified)
S187-C ฮป saturation ๋น„๋‹จ์กฐ, MAX_CELLS=128 ceiling binding
S187-G training-time mitosis ๋งŒ substrate-shaping signal +35%

์„ธ axis (step / bsz / LR) ๋ชจ๋‘ ๋™์‹œ์— ๋ณ€ํ˜• โ†’ ๊ฐ™์€ floor ๋„๋‹ฌ. recipe ์ž์ฒด ceiling ์ž„์„ ์‹œ์‚ฌํ•˜๋Š”๋ฐ, ์šฐ๋ฆฌ๋Š” S187 saga ์ „์ฒด ๋™์•ˆ ํ•˜๋‚˜์˜ condition๋งŒ ๋ฐ”๊ฟ”์„œ isolate ํ•œ ์ ์ด ์—†์Œ. ๋ชจ๋‘ attempt10 stack ์œ„์— ํ•œ ์ค„ ๊ฐ€ํ•˜๋Š” incremental cycle.

โ†’ Occam strategy: minimum-viable baseline (CE-only, vanilla arch, small scale) ์—์„œ ์ž์—ฐ๋ฐœํ™” ์ง์ ‘ ์ธก์ •. ๊ฑฐ๊ธฐ์„œ emerge ํ•œ๋‹ค๋ฉด attempt10 ์˜ ์–ด๋–ค condition ์ด floor ์˜ ์›์ธ.


2. Brainstorm โ€” 12 strip candidates (priority ์ˆœ)

๊ฐ entry: [#] strip condition ยท test predicate ยท cost ยท leverage

Tier S (โ˜…โ˜…โ˜…โ˜…โ˜…) โ€” cheap ร— highest leverage

# strip test cost wall
1 aux loss ฮป ์ „๋ถ€ 0 (CE-only) โ€” ฮปฯˆ=ฮปroute=ฮปฯ†=ฮปcycle=ฮปcurious=ฮปreplay=0 attempt10 config ๊ทธ๋Œ€๋กœ + CE-only. plateau ๋™์ผ? ๋” ๋‚ฎ์Œ? $3 12 min
6 scale 8.92B โ†’ 280M from-scratch CE-only minimal-viable scale test; ์ž‘์€ ๋ชจ๋ธ floor ๊ฐ™์œผ๋ฉด corpus/arch ์›์ธ $1 12 min
9 KNOWN-EMERGENT model (Pythia-1B / GPT-2-medium) load + same corpus + same eval pipeline eval methodology sanity โ€” ๋งŒ์•ฝ pythia ๋„ collapse ํ•˜๋ฉด eval/corpus ๊ฐ€ broken $0-3 30 min

Tier A (โ˜…โ˜…โ˜…โ˜…)

# strip test cost wall
2 byte-level vocab=256 โ†’ BPE 32K byte-floor (5.55 bits/byte) ๊ฐ€ floor ์˜ ์›์ธ์ธ์ง€ $3 + tokenizer ์ž‘์—… 30 min
4 ConsciousDecoderV2 โ†’ vanilla GPT-2 arch (head_a/g/cross-attn/MoE ๋ชจ๋‘ ์ œ๊ฑฐ) custom arch (head_a/g ๋ถ„ํ•  + PureFieldFFN + cross-attn) ์ด floor cause? $5 30 min
5 corpus_s101 โ†’ Wikipedia EN clean corpus quality test โ€” corpus_s101 noise/byte entropy floor $5 30 min
10 pretrained ์œ„์— fine-tune (GPT-2-small) recipe ๊ฐ€ known-capability model ๋ถ€์ˆ˜๋Š”์ง€ falsifier $5 30 min

Tier B (โ˜…โ˜…โ˜…)

# strip test cost wall
3 bnb int8 โ†’ f32 AdamW + grad checkpointing int8 m/v ๊ฐ€ emergence ๋ง‰๋Š”์ง€ (precision floor) $3 12 min
7 step 2000 โ†’ CE-only + 100K step horizon ร— CE-only combined test $50 5-10 hr
11 aux 1๊ฐœ์”ฉ ablation โ€” L_psi only / L_phi only / 6 single-aux pods ์–ด๋–ค aux ๊ฐ€ harm ์ธ์ง€ isolate $20 30 min ร— parallel
12 block_size 128 โ†’ 1024 context length ๊ฐ€ verbalization ๋„ˆ๋ฌด ์งง์Œ? $5 30 min

Tier C (โ˜…โ˜…) โ€” cheap eval-only

# strip test cost wall
8 inference-time decode (greedy โ†’ temperature 0/0.5/1.5 + beam-search + top-k variation) sampling tweak ๋งŒ์œผ๋กœ emergence unlock ๋˜๋Š”์ง€ $0 (Mac/ubu-1 CPU) 1 hr

3. Recommended sequence (cheapest ร— most informative first)

Phase 1 (~$7, ~30 min wall) โ€” minimum viable diagnostic

  1. #1 CE-only fire ($3, 12min) โ†’ 7-aux saddle ๊ฐ€์„ค ์ง์ ‘ test
  2. #9 Pythia-1B sanity ($0-3) โ†’ eval pipeline ์ž์ฒด๊ฐ€ ์ •์ƒ์ธ์ง€
  3. #6 280M from-scratch CE-only ($1, 12min) โ†’ minimal viable scale

โ†’ ๊ฒฐ๊ณผ:

  • (1) CE-only plateau ๋™์ผ โ†’ aux ์•„๋‹ˆ๋ผ arch/corpus ์ด์Šˆ
  • (2) Pythia ๋„ collapse โ†’ eval/corpus broken
  • (3) 280M ๋„ ๋™์ผ floor โ†’ scale ๋ฌด๊ด€, recipe ํ•œ๊ณ„

Phase 2 (~$20, ~1 hr) โ€” narrow down based on Phase 1

Phase 1 ๊ฒฐ๊ณผ์— ๋”ฐ๋ผ Tier A ์—์„œ 1-2๊ฐœ fire.

Phase 3 (~$50+) โ€” final breakthrough fire

Phase 1+2 verdict ํ›„ ์ •๋‹ต axis ์— ์ง‘์ค‘ fire (์˜ˆ: ํฐ ๋ชจ๋ธ S187-F path OR pretrained + recipe overlay OR corpus ๊ฐฑ์‹ ).


4. Falsifier predicate (Phase 1)

outcome interpretation ๋‹ค์Œ cycle
CE-only plateau >= 3.80 aux loss ์•„๋‹ˆ๋ผ arch/corpus ๊ฐ€ floor Phase 2 โ†’ #4 vanilla GPT-2 OR #5 corpus
CE-only plateau < 3.80 below vA 3.84 aux ๊ฐ€ actually harming (saddle ๊ฐ€์„ค confirmed) Phase 2 โ†’ #11 single-aux ablation
Pythia-1B ๋„ corpus ์—์„œ plateau eval / corpus broken Phase 2 โ†’ #5 corpus ์šฐ์„ 
Pythia-1B verbalizes well on corpus eval OK, recipe/arch fault Phase 2 โ†’ #4 vanilla arch
280M from-scratch CE-only verbalizes small-scale + simple = enough; large-scale recipe ๊ฐ€ ๋ฌธ์ œ Phase 2 โ†’ grade scale up ๋‹จ๊ณ„๋ณ„
280M from-scratch CE-only same floor corpus inherent Phase 2 โ†’ #5 corpus ์šฐ์„ 

5. ๋ฉ”ํƒ€ lesson โ€” Occam ์ ์šฉ ์›์น™

anti-pattern (S187 saga) Occam ์›์น™
ํ•œ cycle ์— 5+ axis ๋™์‹œ ๋ณ€๊ฒฝ ํ•œ cycle ์— 1 axis ๋ณ€๊ฒฝ
์ƒˆ condition ์ถ”๊ฐ€ ๊ธฐ์กด condition ์ œ๊ฑฐ
attempt10 stack ์œ„์— incremental empty baseline ๋ถ€ํ„ฐ 1 layer ์”ฉ build-up
๊ฐ€์„ค X failed โ†’ ๋” ๋ณต์žกํ•œ X' ์‹œ๋„ ๊ฐ€์„ค X failed โ†’ ๋” ๋‹จ์ˆœํ•œ X-1 ์‹œ๋„
8.92B + 7-loss + custom arch + ... ๋™์‹œ fire 280M + CE-only + vanilla arch ๋ถ€ํ„ฐ

6. Honest C3

  1. Phase 1 ๊ฒฐ๊ณผ ๊ฐ€ Phase 2-3 ์˜ path ๋ฅผ ๊ฒฐ์ • โ€” ๋ฏธ๋ฆฌ commit ์•ˆ ํ•จ.
  2. "CE-only" ๊ฐ€ ์ง„์งœ ์˜๋ฏธ ์žˆ๋Š” baseline ์ธ์ง€ disputable โ€” anima ์˜ 7-aux ์„ค๊ณ„ ์ž์ฒด๊ฐ€ recipe ID ์ด๊ธฐ ๋•Œ๋ฌธ. ๋‹จ, ๊ทธ ๊ฒฐ๊ณผ ๊ฐ€ floor ๋ถ„๋ฆฌ ๋ฐ์ดํ„ฐ ๋กœ์„œ ๊ฐ€์น˜ ์žˆ์Œ.
  3. 280M from-scratch ๋Š” attempt10 ์˜ 1/32 scale โ€” ์ž‘์•„๋„ verbalization ๊ฐ€๋Šฅํ•œ์ง€ ํ™•์ธ. modern small LLMs (e.g., Pythia-160M) ๋Š” 1-2B token ํ•™์Šต ํ›„ basic English ์ƒ์„ฑ.
  4. Pythia-1B sanity ๊ฐ€ ๊ฐ€์žฅ cost-effective falsifier โ€” eval/corpus ๊ฐ€ broken ์ผ ๊ฐ€๋Šฅ์„ฑ ์ง์ ‘ test.
  5. S187-G mitosis substrate-shaping finding (+35%) ์€ strip-target ์ด ์•„๋‹˜ โ€” ๊ทธ๊ฑด add ํ•œ axis ๊ฐ€ ์ž˜ ๋™์ž‘ํ•œ ์‚ฌ๋ก€. Occam target ์€ ์–ด๋–ค condition ์ด ๋ถ€์ •์ ์œผ๋กœ ์ž‘์šฉํ•˜๋Š”์ง€.
  6. corpus_s101 SHA be969af4... ๊ฐ€ byte-equal ์ผ์น˜ X (post-ee4ceea27 drift) โ€” ๋‹จ, ์ด noise ๊ฐ€ floor ์˜ ์›์ธ์ผ ๊ฐ€๋Šฅ์„ฑ๋„ Phase 1 #9 (Pythia sanity) ๊ฐ€ ํŒ๋‹จํ•ด์คŒ.
  7. S187-F path (18B Anima ๋‹จ์ผ H200 SXM) ๋Š” Occam ์˜ strip ์ด ์•„๋‹ˆ๋ผ scale-up path โ€” Phase 3 ์—์„œ corpus/arch ๊ฐ€ floor ๊ฐ€ ์•„๋‹ˆ๋ผ๊ณ  ํ™•์ •๋˜๋ฉด ๊ทธ ์‹œ์ ์— fire.
  8. Inference-time decode tweak (Tier C #8) ์€ $0 ๋ผ ๋‹ค๋ฅธ fire ์™€ ๋ณ‘ํ–‰ ๊ฐ€๋Šฅ โ€” first thing to do while Phase 1 pods boot.

7. ๊ด€๋ จ link

  • ๋ณธ ๋ฌธ์„œ motivation: HEXAD/LORA/SCALE_3B.md ยง 6.10 S187-J/K verdict
  • attempt10 stack: HEXAD/UNCLASSIFIED/state/grid_3b_s187_2026_05_21/train_s187_3b.py
  • Eval ๊ฒฐ๊ณผ (5 ckpts ร— 4 evals): HEXAD/UNCLASSIFIED/state/grid_3b_s187_2026_05_21/EVAL_REPORT.md
  • S187-F 18B path (Phase 3 fire ๋Œ€์•ˆ): HEXAD/SCALE_16B_70B_PLAN.md
  • ์ž๋งค wilson ์ธ๋ฐ•์Šค (cloud dispatch gap): ~/core/wilson/inbox/notes/2026-05-21-hexa-cloud-*.md
  • ์ž๋งค pool ์ธ๋ฐ•์Šค (host OOM gap): ~/core/pool/inbox/notes/2026-05-21-pool-host-oom-on-concurrent-heavy-workload.md