Synthetic creatures with one continuous life learn where food is, not what it is, and fail a word test
DesignEddi · reddit · 2026-11-18
A developer shares measured results from "Synthients" — small synthetic creatures living a single continuous life with no episodes, resets, or checkpoints, and no language model inside.
Key findings (small samples, pre-registered pass bars):
- Barely self-sufficient: 7 of 9 new lives kept themselves fed over 300k steps while exploring 843 of 900 grid squares — exactly at the bar.
- They learn where, not what: 7 of 8 creatures conditioned to fixed food locations largely ignored ripe food appearing in new places; behavior was bound to place, not food.
- Curiosity trap: over 1.2M-step lives, all 9 spent growing time (from 3% to a median 10%) in a pure-noise patch — the classic "noisy TV" problem re-emerging slowly within one life.
- Failed first word test: 12 creatures trained 200 lessons each to pick one of three objects after a sign scored 1/3, chance level, with none passing. The author suspects missing episodic/single-moment memory, which they'll build next.
The author asks: is learning "where" instead of "what" a bug or the rational strategy when food locations have been reliable — and how would you design a test to tell them apart? Write-ups at hesperan.com/research.
More from Research
- CADMAGEDDON: open-source agent rebuilds editable CAD locally on a single GPU — Oleg Pavlov · 2026-11-18
- MillBill: local 27B model beats frontier models at generating machining programs — Gennadii Savrasov · 2026-11-18
- Ex-DeepMind's Ryutaro Tanno joins AI biotech Pathos as Chief AI Officer — davidstutz92 · 2026-11-18
- VRST 2027 Goes Fully Distributed With Three Equal Sites Across US, Europe and Asia — plopesresearch · 2026-11-18
- Decision models beat LLMs at structured output: Liquid AI d1 runs 2.7-3.6x faster than LFM3 — mervenoyann · 2026-11-18
- Researcher pulls 3 all-nighters for ICLR rebuttal, then gets 10 days of silence from AC — IanArawjo · 2026-11-18