← back to timeline

FRISTON-MINIMAL-PILOT

negative

friston-minimal-pilot

2026-05-25 L7 level 2 runs
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

A small AI watches a short sequence of sentences cycle past it, tries to guess what sentence will come next, and writes down what surprised it. We check whether its guesses get less wrong over 15 turns — and whether the guesses are actually better than an AI that doesn't keep any memory.

Technical details

Research hypothesis

Phase F1 of AUTHORS-MATRIX Friston cell — operationalize the Markov-blanket active-inference loop on local Qwen substrate as: (a) internal state = a persistent text-memory buffer; (b) sensory states = deterministic env_cycle text inputs; (c) action states = LLM-generated outputs per turn; (d) generative model = LLM conditioned on (persona + memory + last input) producing a «predict-what-comes-next» continuation; (e) prediction error = 1 - cosine between predicted-next-text embedding and actual-next env input embedding; (f) update rule = on epsilon > theta, append (expected, actual) surprise line to memory (FIFO cap 2000 chars). F1 tests only P1 (substrate learns: late-gen mean epsilon < early-gen mean epsilon by ≥0.10 in MEMORY_DRIVEN) and P4 (substrate distinguishably outperforms blind baseline: MEMORY_DRIVEN late mean < RANDOM_BASELINE late mean by ≥0.05). P2 (action steers env) and P3 (Bayesian-optimal update behavior) are deferred to F2.

Experimental setup

Type: 2-condition-matched

Condition Parameters
MEMORY_DRIVEN prediction_mode: memory_driven, memory_update_enabled: true, note: Agent receives persona + accumulated surprise memory + last sensory input. Memory grows on epsilon>theta surprises. Tests P1 (learning) and provides the «active inference» candidate side of P4.
RANDOM_BASELINE prediction_mode: random_baseline, memory_update_enabled: false, note: Agent receives persona + EMPTY memory + last sensory input (no accumulation across turns). Blind-predictor control for P4 distinguishability. Mandatory per AUTHORS-MATRIX Friston cell Goodhart-audit.

Factors: prediction_mode (memory_driven, random_baseline)

Parameters

model
qwen2.5-7b-instruct-1m
n_runs_per_condition
1
temperature
0.85