architecture-selection
Skill0 starsChanged 56 days agoMITView on GitHub
Choose an architecture and input representation for a narrow task — classical baselines, CNNs, GRUs, tiny transformers, and the representation design that matters more than any of them. Use when starting a tiny-model experiment, when a model has plateaued, or whenever someone reaches for a transformer by default.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- dataset-synthesisSkill
- design-evalSkill
- export-pipelineSkill
- quantization-strategySkill
- tiny-model-triageSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
Before touching architecture, ask:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Architecture and representation
The representation is the product. Most of the win in a tiny model comes from framing the input so the task becomes easy — not from the model.
Representation first
gpu-lexer is the reference case. Instead of tokenising per language, a CPU pass splits text into word runs, whitespace, newlines, and symbols, then records language-neutral features: part kind, length, edge characters, hashes, neighbouring symbol pairs. Only then does a model embed those sparse features into learned channels. The cleverness is upstream of the network.
Before touching architecture, ask:
- What structure does the data actually have — locality, order, symmetry, hierarchy?
- What can a deterministic pass compute for free, so the model does not have to learn it?
- What can be removed from the input without hurting the task? Smaller input, smaller model.
Spending a day here routinely beats a week of hyperparameter tuning.
Inductive-bias ladder
Pick the weakest thing that fits the structure. Stronger bias means less data and fewer parameters.
| Structure in the data | Reach for |
| Independent features, tabular | Logistic regression, then gradient-boosted trees |
| Local / spatial / translation-invariant | 1D or 2D convolutions |
| Sequential, modest range | GRU, or convolutions plus a scan |
| Long-range, content-dependent pairing | Attention — and only here |
Gradient-boosted trees still beat neural networks on most tabular problems. They are also the fastest thing to try. Note their inference latency is often higher than a small network's, which matters if you are targeting the instant band.
On the bitter lesson
Sutton's argument is that general methods leveraging computation beat handcrafted structure. At tiny scale with narrow tasks and little data, you are in the opposite regime — for now.
The useful distinction (Cranmer): bake in inductive biases grounded in the mathematical structure of the data — locality, symmetry, invariance. Do not bake in arbitrary human heuristics about how the task ought to be solved. The first kind ages well; the second becomes the ceiling.
Hyung Won Chung's framing is the one to hold: add the structure your current compute and data require, and plan to remove it later, because it will eventually bottleneck you.
Hybrids are usually right
gpu-lexer and JPU both pair a learned classifier with a deterministic composition step that guarantees valid output structure. Let the model handle the fuzzy part and let code enforce the invariants. You get learned flexibility without learned nonsense.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

