← TinyGPT · docs · devlog · roadmap · speedup
source: docs/sessions/2026-06-17-stepback-inventory-roi.md · view on GitHub ↗

Step-back: what we have, the wall, and the ROI menu

Date: 2026-06-17. A deliberate pause to inventory the assets, name the wall, and rank what’s next by ROI — before committing to another build. Narrative of the arc lives in the 2026-06-16 session doc and journey §8; this is the ledger + the fork.

The arc that led here

Distillation hit frontier-parity on file-ops agentic (4B 58→100, beats Gemma-12B at ⅓ size) → the real gap was breadth, where specialists regress (negative transfer) → teacher-free ReST lifted breadth 60→65% with depth held → then we widened the lens: measured Apple’s on-device model as a candidate (it can’t ground actions), mined VibeThinker’s training recipe (SSP/MGPO/specialist-merge), and closed the distributed-boundary thesis gap with a working data-parallel PoC. The honest read: the learning compounded; the product (agentic tool-calling) is near its cheap ceiling.

What we have (the inventory)

Models (~/.cache/tinygpt/models/)

modelwhat it isnumbers
mt4b_fusedQwen3-4B distilled, file-ops agentic specialist100% hard, beats Gemma-12B at ⅓ size
mt4b_rest_fused+ teacher-free ReST iterationdepth 100%, breadth 65% (stock 60%)
mt4b_mb_fusedmulti-backend gold-distill (the failed one)depth 100 / breadth 31% (negative transfer)
vibethinker-3b-mlxconverted reasoning specialistGSM8K 40/40 = 100%; no tool-calling

Infrastructure (scripts/)

Findings / intel (durable)

The wall

  1. Cheap distillation is maxed on trained domains. File-ops saturated at frontier; breadth is a long-tail grind (+5pp from ReST, and each further point costs much more).
  2. No forcing function to an outcome. Each result spawns three experiments — rich for learning, but nothing was shipping. That’s why it felt like it wasn’t going anywhere.
  3. Shipping ≠ wiring. mt4b_fused speaks BFCL tool_calls; Pace’s planner needs the v10/v11 intent-envelope (spokenText + 7 intents incl. safety). Shipping it = re-distillation on Pace’s action surface, gated by the v11 ship gate — a real project, not a plug-in.

The ROI menu (what we can do)

Ranked by payoff/effort; honest about which axis each serves.

#optioneffortpayoffaxisnote
1Specialist-merge experimentLOW–MEDMEDresearchcheapest shot at the breadth wall; infra exists; either lifts >65% or closes the lane honestly
2MGPO auto-curriculum patch to ReSTLOWMEDresearchre-weight sampler to the 30–70% frontier band; cheap efficiency gain
3Model radar (systematic small-model discovery)LOWMEDintelturns VibeThinker-style luck into a pipeline; on-thesis
4Re-distill mt4b on Pace’s action surfaceHIGHHIGHproductthe real shipped outcome; gated by the v11 ship gate (hard unhappy-path dims)
5Serving/throughput (continuous batching)MED–HIGHHIGHboth~10× rollouts (unblocks all future RL) + Pace latency win
6Pivot: distributed boundaryMEDlearninglearningbuild on the PoC (toy ZeRO, 2-Mac Thunderbolt); positions for scale
7Consolidate & stopdoneclosurethis doc is the capstone; reassess fresh

Honest recommendation: if the goal is a clean verdict on the agentic lane, do #1 (specialist-merge) — cheap, attacks the measured wall, and gives a yes/no. If the goal is a shipped outcome, #4 is the only one that ends with a Pace capability, but commit to it as a real project (it needs the v11 corpus + ship gate). #5 is the best “unblocks everything” bet if we intend to keep doing RL.

Pending (in flight as of writing)