Note / 002
Stupid AI / Studio build note
Working thesis
AI should do
the adapting.
The smaller the model gets, the more disciplined the system around it has to become.
The thesis
Stupid AI is not a tiny ChatGPT. It is a portable action engine: a small language model inside a deterministic harness.
Tiny model + harness
= useful agent.
The model handles messy language. The harness carries the world: tools, memory, permissions, state, verification, and consequences.
The organizing constraint
If the harness can solve it deterministically,
delete it from the model.
Dates, numbers, schema validation, permission checks, state transitions, calculations, retries, transactions, and verification are better as code whenever code can do the job.
Every rule moved out of the weights becomes easier to inspect, update, secure, and test.
What the tiny model actually does
The smallest tier is not pretending to plan. It behaves more like a learned parser with a transformer backbone: narrow, fast, and very good at turning ordinary language into a structured next move.
Intent.
What is the person trying to do?
Fields.
What facts are already present?
Missing.
What must be known before a valid move exists?
Refusal.
When should the system stop instead of guessing?
Multi-step planning and recovery can move to a larger local tier when the task actually requires them. General capability is an escalation path, not the baseline tax.
What the harness does
World.
Facts, files, records, and current state.
Tools.
The actions the system is actually allowed to take.
Memory.
What remains relevant from earlier work.
Rules.
Permissions, schemas, limits, and legal state changes.
Proof.
Tests, verification, audit logs, and evidence of completion.
Constrain the output
If every output is a known action packet, JSON syntax should not be a model skill. Grammar-constrained decoding can make malformed structures impossible before verification even begins.
The model should rank valid continuations, not remember where a brace goes.
At this scale, removing impossible outputs can matter more than adding another pile of parameters.
Repair is part of the loop
Real harnesses reject actions. A missing permission, illegal state change, failed test, or bad argument should produce a repair request that the model knows how to consume.
The repair loop is not an edge case. It is the loop.
The verifier is also the teacher
Schema valid? State transition legal? Tests pass? Unauthorized file touched? The harness already has deterministic answers to questions like these.
Those answers are training signal. Rejection-and-repair traces, correct refusals, verified actions, and false success claims can all be scored without asking a person to label every example.
The verifier is not only a guardrail.
It is a data engine.
Confidence should mean something
A number like 0.91 looks authoritative. A raw probability from a small model is not automatically calibrated enough to deserve that trust.
Useful confidence can come from the harness instead: required fields present, entity resolved, permission valid, evidence found, verifier passed.
Precision in the interface should come from evidence, not decoration.
The metric that matters
Capability is easy to celebrate. Honest stopping is harder and more important.
A fixed evaluation suite should score correct actions, unauthorized access, correct refusal, and false success claims. Failure can be repaired. A system that says “done” when the verifier says otherwise breaks the product promise.
False success is worse
than failure.
What simple should feel like
No model shopping. No giant context dump. No vector database by default. No prompt handbook before useful work can start.
Small enough to run anywhere. Honest enough to be useful everywhere.
Public Notes explain design principles. Private datasets / client systems / credentials / security implementation details stay private.
Notes / Subscribe
Get the next note.
No schedule. No content calendar. A note goes out only when there is something worth understanding.