Skip to content

Commit 2d65210

Browse files
0bserver07claude
andcommitted
docs(tasks): add task 015 - V2/V3 instrumentation cross-repo index
Tracks the ByteDMD (v2) and simplified-Bill-Dally (v3) instrumentation effort across hinton-problems, schmidhuber-problems, sutro-problems. Captures current state (PR #59 merged, #60 and #18 open), ownership, and the open IR-expansion question. SutroYaro acting as the lab dispatcher/index per its scoped role. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 parent b536a0a commit 2d65210

2 files changed

Lines changed: 77 additions & 0 deletions

File tree

Lines changed: 76 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,76 @@
1+
# Task 15: V2 / V3 instrumentation of Hinton + Schmidhuber stubs
2+
3+
**Priority**: HIGH
4+
**Status**: IN PROGRESS
5+
**Source**: Telegram chat-yad, 2026-05-09 to 2026-05-12 (Yaroslav's v2/v3 staging clarification, Andy + Seth's first comparator pairs, the IR-expansion question)
6+
7+
## Context
8+
9+
After the hinton-problems (53 stubs) and schmidhuber-problems (58 stubs) catalogs shipped, the next step is measuring them under data-movement metrics. Yaroslav's staging (2026-05-11):
10+
11+
- **V2** = the subset of stubs feasible to instrument with **ByteDMD** (Python-level LRU stack tracer). Output: per-stub ByteDMD cost + reuse-distance distribution.
12+
- **V3** = the further-restricted subset feasible under the **simplified Bill Dally** model (the 2D-grid v0 IR used in `sutro-problems/matmul`). Output: per-stub read-distance histogram in the matmul `doc/access_distance/` format.
13+
- Survivors of both filters become the **candidate pool for the next hill-climbing competition** (matmul-style leaderboard).
14+
15+
This is a sideways move on the [3-axis cube](../research/active-threads-2026-05-09.md): stay on the Hinton/Schmidhuber problem rung, climb one rung on the metric axis (ByteDMD → simplified Bill Dally).
16+
17+
**Relevance to SutroYaro**: the code lands in the per-purpose repos (hinton-problems, schmidhuber-problems, sutro-problems), not here. SutroYaro's job is to be the cross-repo index for the effort — that's what this doc is.
18+
19+
## Where the work lives
20+
21+
### V2 (ByteDMD instrumentation)
22+
23+
| Repo / location | State |
24+
|---|---|
25+
| `hinton-problems/v2-bytedmd/` | Live. Framework by Andy Zhang (zh4ngx). |
26+
| `hinton-problems` PR #59 | **MERGED** 2026-05-10. encoder-8-3-8 backprop vs Boltzmann CD-1/CD-5. Backprop wins 49–135x. |
27+
| `hinton-problems` PR #60 | **OPEN**. bars-rbm (CD-1) vs bars (wake-sleep Helmholtz). Seth Stafford (SethTS). CD-1 wins ~20x. |
28+
| `schmidhuber-problems` | No `v2-bytedmd/` dir yet. The cleanest comparator candidate is `linear-transformers-fwp` vs `fast-weights-key-value` (Schmidhuber proved mathematically equivalent; would confirm whether the ByteDMD numbers also match). |
29+
| Tracking issues | [hinton-problems#45](https://github.com/cybertronai/hinton-problems/issues/45), [schmidhuber-problems#17](https://github.com/cybertronai/schmidhuber-problems/issues/17) |
30+
31+
### V3 (simplified Bill Dally / matmul-style IR)
32+
33+
| Repo / location | State |
34+
|---|---|
35+
| `sutro-problems/matmul/` | The template. `doc/access_distance/` is the histogram + CDF format to follow. |
36+
| `sutro-problems` `dev` branch, `wip-boltzmann-shifter/` | Partial port of Hinton's shifter into the matmul-style submission framework. |
37+
| `sutro-problems` PR #18 | **OPEN**. symmetry (6-bit palindrome) as a third sutro-problems challenge. Seth Stafford. Surfaces the IR-expansion problem: optimal baseline uses `cmp eq`, not in the v0 op set. |
38+
| Andy's hinton→dally agent | Hit the IR roadblock (only IR-trackable algorithms work). Not PR'd. Paused pending the IR-expansion decision. |
39+
| `simplified-dally-model` | Yaroslav's scoring engine. He is still iterating on the minimum instruction set. |
40+
41+
## Open question: does the v0 IR need more ops?
42+
43+
Raised by Andy's roadblock and Seth's symmetry PR. Yaroslav (2026-05-11):
44+
45+
> If agents can figure out a way to solve the toy problems presented by Hinton or Schmidhuber by only utilizing the IR, [good]. If they can't then we should probably increase it by adding ops.
46+
47+
Concrete evidence: PR #18's symmetry baseline (cost 20) uses `cmp eq` to test patterns rather than evaluating the polynomial form, and `cmp eq` is not in the v0 set (`add`, `sub`, `mul`, `copy`).
48+
49+
This decision is tracked separately as an issue on `simplified-dally-model` (or `sutro-problems`) — see the link once filed. It blocks the V3 subset definition.
50+
51+
## Ownership (current)
52+
53+
- **Andy Zhang (zh4ngx)** — V2 encoder framework (done), hinton→dally V3 (blocked on IR)
54+
- **Seth Stafford (SethTS)** — V2 bars pair (PR #60), symmetry V3 challenge (PR #18)
55+
- **Yad** — Schmidhuber V2 (the `v2-bytedmd/` dir that doesn't exist yet) is the open gap; also the cross-repo coordination (this doc)
56+
- **Yaroslav** — IR-expansion decision; will look at V3 seriously after the AI Council talk (2026-05-12)
57+
58+
## Next actions
59+
60+
- [ ] Review + merge `hinton-problems` PR #60 (bars pair)
61+
- [ ] Review + merge `sutro-problems` PR #18 (symmetry challenge)
62+
- [ ] File the IR-expansion issue on `simplified-dally-model`, linking PR #18's `cmp eq` finding
63+
- [ ] Start `schmidhuber-problems/v2-bytedmd/` cloning the hinton-problems pattern; first pair = `linear-transformers-fwp` vs `fast-weights-key-value`
64+
- [ ] Once Yaroslav decides the IR scope, unblock Andy's hinton→dally V3 work
65+
66+
## Acceptance
67+
68+
- A V2 `v2-bytedmd/` dir exists in both hinton-problems and schmidhuber-problems with at least 2 comparator pairs each
69+
- The IR-expansion question has a public decision (either "v0 is enough" or "here is the minimal expanded set")
70+
- At least one V3 read-distance histogram exists for a Hinton or Schmidhuber stub, in the matmul `doc/access_distance/` format
71+
- This doc stays current as the cross-repo index for the effort
72+
73+
## Dependencies
74+
75+
- V3 subset definition is blocked on the IR-expansion decision
76+
- The hill-climbing competition (the eventual payoff) is blocked on having a V3 candidate pool

docs/tasks/INDEX.md

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -18,3 +18,4 @@ Feedback from Meetings #8-16, Telegram, and Google Docs. Updated 2026-05-09.
1818
| 12 | 2D-grid Dally model extension to sparse-parity | HIGH | OPEN | [012-2d-grid-dally-sparse-parity.md](012-2d-grid-dally-sparse-parity.md) |
1919
| 13 | Accuracy vs joules graph for sparse-parity solvers | MEDIUM | OPEN | [013-accuracy-vs-joules-graph.md](013-accuracy-vs-joules-graph.md) |
2020
| 14 | Sparse-parity over time ("morse code" framing) | MEDIUM | OPEN | [014-sparse-parity-over-time.md](014-sparse-parity-over-time.md) |
21+
| 15 | V2 / V3 instrumentation of Hinton + Schmidhuber stubs (cross-repo index) | HIGH | IN PROGRESS | [015-v2-v3-instrumentation.md](015-v2-v3-instrumentation.md) |

0 commit comments

Comments
 (0)