Fyto: Autoregressive Language-Model Inference Powered Solely by a Living Houseplant

Published plant- and soil-microbial-fuel-cell (PMFC/SMFC) electronics run duty-cycled sensing and a radio; none runs a neural network. We present Fyto, a sealed box in a potted golden pothos (Epipremnum aureum) that generates text with a llama2.c-class transformer on an Ambiq Apollo4 Plus microcontroller, with two hydraulically separated PMFC chambers as its only energy source and output leaving through a galvanically isolated UART. No Fyto hardware has yet been measured: energy figures are emulated or modelled, and the claim that a plant powers inference is a pre-registered hypothesis. Instruction-level emulation of verified Cortex-M4 kernels gives 3.1–3.5 cycles per int8 multiply–accumulate, about twice the usual assumption; the resulting model gives 18–55 µJ per token for a 1.1 M-parameter model (36–55 µJ at the emulated cycle counts; the low end assumes a hand-scheduled 1.75-cycle kernel). Pots delivering the tens-to-hundreds of microwatts reported in the literature still sustain 2×10⁴–5×10⁵ tokens per day (revised chain, Monte Carlo p10–p90). The binding constraints lie elsewhere: an always-on chain idles at 34–69 µW, the specified harvester's fractional-Voc tracking (70–90 %) captures at most 0.61 of the power of a cell whose maximum power point sits at 0.42 Voc; and retained SRAM costs more than the plant supplies. A revised chain with per-chamber constant-voltage harvesting, a decifarad store, power-gating and per-token MRAM commits idles at 1–2.5 µW and wakes in 99.7 % of draws. Running the real stories260K checkpoint, we measure 0.446 words per token and show that per-token commits resume bit-exactly in all 600 fault-injection runs (1,513 brownouts), against 187 of 600 runs (31 %) for a naive ordering. Because a gram of soil carbon could power this load for decades, we pre-register controls (unplanted, sterilised soil, darkness, plant removal, electrode swap, titanium-only, galvanic), an energy ledger, a coulombic-overdraft test and a ¹³CO₂ electron budget that would measure, for the first time in any plant fuel cell, the fraction of charge, and so of tokens, paid for by freshly fixed CO₂. Preprint v1. Code, data, figures and LaTeX sources are included in fyto-paper-v1-code-data.zip. Project: fytopure.online.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-07
DOI
https://doi.org/10.5281/zenodo.24033306
Primary Topic
Microbial Fuel Cells and Bioremediation
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Fyto: Autoregressive Language-Model Inference Powered Solely by a Living Houseplant

Arhaan Khan
2 links
Zenodo (CERN European Organization for Nuclear Research)
Microbial Fuel Cells and Bioremediation
preprint

Fyto: Autoregressive Language-Model Inference Powered Solely by a Living Houseplant

Arhaan Khan
preprint en

Abstract

Published plant- and soil-microbial-fuel-cell (PMFC/SMFC) electronics run duty-cycled sensing and a radio; none runs a neural network. We present Fyto, a sealed box in a potted golden pothos (Epipremnum aureum) that generates text with a llama2.c-class transformer on an Ambiq Apollo4 Plus microcontroller, with two hydraulically separated PMFC chambers as its only energy source and output leaving through a galvanically isolated UART. No Fyto hardware has yet been measured: energy figures are emulated or modelled, and the claim that a plant powers inference is a pre-registered hypothesis. Instruction-level emulation of verified Cortex-M4 kernels gives 3.1–3.5 cycles per int8 multiply–accumulate, about twice the usual assumption; the resulting model gives 18–55 µJ per token for a 1.1 M-parameter model (36–55 µJ at the emulated cycle counts; the low end assumes a hand-scheduled 1.75-cycle kernel). Pots delivering the tens-to-hundreds of microwatts reported in the literature still sustain 2×10⁴–5×10⁵ tokens per day (revised chain, Monte Carlo p10–p90). The binding constraints lie elsewhere: an always-on chain idles at 34–69 µW, the specified harvester's fractional-Voc tracking (70–90 %) captures at most 0.61 of the power of a cell whose maximum power point sits at 0.42 Voc; and retained SRAM costs more than the plant supplies. A revised chain with per-chamber constant-voltage harvesting, a decifarad store, power-gating and per-token MRAM commits idles at 1–2.5 µW and wakes in 99.7 % of draws. Running the real stories260K checkpoint, we measure 0.446 words per token and show that per-token commits resume bit-exactly in all 600 fault-injection runs (1,513 brownouts), against 187 of 600 runs (31 %) for a naive ordering. Because a gram of soil carbon could power this load for decades, we pre-register controls (unplanted, sterilised soil, darkness, plant removal, electrode swap, titanium-only, galvanic), an energy ledger, a coulombic-overdraft test and a ¹³CO₂ electron budget that would measure, for the first time in any plant fuel cell, the fraction of charge, and so of tokens, paid for by freshly fixed CO₂. Preprint v1. Code, data, figures and LaTeX sources are included in fyto-paper-v1-code-data.zip. Project: fytopure.online.

Zenodo (CERN European Organization for Nuclear Research)
Affordable and clean energy
Microbial Fuel Cells and Bioremediation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Fyto: Autoregressive Language-Model Inference Powered Solely by a Living Houseplant — Arhaan Khan · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS