Does fitting only on the windows an assistant explored at random fix learning from its own feedback?
research/ambient-exploreonly
Fitting only on the windows an assistant explored at random fixes the uninformed-prior failure of learning from its own feedback (11.99 vs −13.08 utility/h at ε = 0.02), at an exploration cost (stylised)
What it found
A stylised follow-up to ambient-labelmatch and the reweighting and doubly robust fixes after it. The deployed learner already draws a random sample for free: the windows it spoke on only because its ε coin came up. Fitting on those windows alone, unweighted, reaches 11.99 utility/h at ε = 0.02 and 12.25 at 0.05 with an uninformed prior, against −13.08 and −71.85 for the fit on every spoken window (oracle 12.71; 20 seeds), and as a learner it earns 6.49/h online at ε = 0.005 against −2.04. Exploration costs dismissals (online utility falls at larger ε), and with an informed prior the plain fit is better at small ε. Stylised simulation only.
A figure from its own output
Paper, code and results
More in Ambient AI assistant
- How many false wakes a day does an always-on trigger get when its frame scores are correlated, and what does requiring m frames in a row buy?How many false wakes a day does an always-on trigger get when its frame scores are correlated? Exact grid-chain rates for "m frames in a row" on AR(1) scores, checked by Monte Carlo: a threshold set by the independent-frames rule gives 613 / 4,793 / 20,993 wakes a day (not 1) for m = 2 / 3 / 5 at phi = 0.9, while a single-frame threshold is nearly unaffected. Stylised.
- How bursty are an always-on trigger's false-wake counts, and does a lockout help?How bursty are an always-on trigger's false-wake counts? On AR(1) frame scores the Fano factor is near 1 (1.05–1.5), but a 5% window-to-window drift in noise scale lifts it to about 9 for m = 3 (38 at 10%); a 10 s lockout cuts it to 2.5 but also cuts the mean by a third. Stylised.
- How does the burstiness of an always-on trigger's false-wake counts grow with the counting window when the noise level drifts for hours?How does the Fano factor of false-wake counts grow with the counting window when the noise level drifts for hours? At 5% persistent drift (about 3 h correlation time) it grows from 1.9 at 100 s to about 150 at 28 h, against a flat 1.9 for independent drift. Stylised; the drift is assumed.
- Do a dual-MAC energy loop and a divide-free threshold make the Cortex-M4 voice gate cheaper?Do a dual-MAC energy loop and a divide-free threshold make the Cortex-M4 voice gate cheaper: 17-37% fewer instructions with the block-minimum floor, slower without it, RAM doubled; decisions unchanged (emulator counts, synthetic audio)
- What does the adaptive voice gate cost on a Cortex-M4, in flash and instructions per frame, and is the block-minimum floor worth it?What does the adaptive voice gate cost on a Cortex-M4?
ambient-gatefw's C gate built with clang for Thumb-2 and run in an emulator: bit-exact, about 250 bytes of code, 3237 instructions per frame exact and 2096 with the block-10 floor, because the energy sum dominates. Instructions, not cycles; stylised. - Can a noise-adaptive voice gate fit in about 120 bytes of RAM on a pendant microcontroller, and does a cheaper noise floor cost battery life?Can a noise-adaptive voice gate fit in about 120 bytes of RAM on a pendant microcontroller? C port of
ambient-gateroc's adaptive energy gate, bit-exact with the Python reference; a block-minimum noise floor cuts state from 1020 to 120 bytes and comparisons per frame from 249 to 24, moving modelled battery life by at most 0.07 h. Synthetic audio, host build only.
All 46 Ambient AI assistant projects →
Other research topics
Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.