Current AI is built on text. Text is the compression format for sharing thoughts. I’m interested in what’s underneath: the structure that emerges when an agent experiences reality one moment at a time and learns what those moments mean to each other.
Experiences happen sequentially — that’s an inevitable fact of living in a linear-time reality. Any embodied agent produces temporally ordered memories, and that ordering turns out to be a powerful, universal signal for learning associations. No labels, no supervision, no annotation. Temporal co-occurrence alone produces faithful associative recall where similarity-based approaches score zero.
I’m also investigating whether the physics of holographic storage provides a natural mechanism for continual learning. Both programmes come from the same place: the conviction that intelligence is grounded in reality and that we’re leaving too much of reality on the table.
The same contrastive learning method that improved text retrieval and discovered narrative structure in 9,766 novels also works in molecular biology. Trained on protein interaction data from the STRING database, the model recovers functional gene relationships that are invisible to standard expression similarity — achieving strong discrimination (AUC 0.908) in precisely the regime where similarity-based methods fall to chance (0.518). A second gene dataset confirms the result. Two drug experiments identify where the method fails and why. Three findings emerge that the text experiments alone could not have revealed: biological associations transfer to completely unseen genes where text associations do not, improvement concentrates on understudied genes with the greatest need for functional characterisation, and tighter association quality outperforms larger but noisier training data — reversing the pattern seen in text.
PAM trained at scale on 9,766 Project Gutenberg novels (25 million text chunks, 373 million temporal relationships) discovers hierarchical narrative structure without supervision. The model learns what passages do rather than what they’re about, grouping a Victorian chase scene with a Russian chase scene because they perform the same structural beat, despite sharing no vocabulary. Clusters range from broad narrative modes at coarse resolution to specific registers like courtroom cross-examination and sailor dialect at fine resolution. Held-out novels receive coherent assignments through single-pass inference, demonstrating inductive transfer. An interactive demo is live.
The foundational paper. A JEPA-style predictor trained on temporal co-occurrence learns to retrieve items linked through shared experience rather than representational proximity. On a synthetic benchmark, the predictor’s top retrieval is a true temporal associate 97% of the time, recovering associations across representational boundaries where cosine similarity scores zero. A temporal shuffle control confirms the signal is genuine co-occurrence structure, not embedding geometry. A held-out evaluation confirms the predictor remembers what it experienced, from the perspectives at which it experienced it.
The applied paper. A lightweight MLP (4.2M parameters) trained on passage co-occurrence reranks dense retrieval results for multi-hop question answering. On HotpotQA, Recall@5 improves by 8.6 points without evaluation-set tuning, with gains concentrated on the hardest questions (+28.5 points where the dense baseline fails). On MuSiQue, +10.1 points. Training on similar-but-not-associated pairs actively degrades retrieval; association and similarity aren’t just different, they’re sometimes opposite. An inductive variant shows no significant improvement, confirming the method captures corpus-specific co-occurrences. The method adds 3.7ms per query and trains in under two minutes.
Investigating photorefractive crystals as neural network inference engines. Weight matrices stored as angle-multiplexed volume holograms inside iron-doped lithium niobate crystal, with light propagation physically performing matrix-vector multiplication. Simulation modelling the full wave physics has established single-layer classification at 92.6% accuracy and identified width scaling as the natural axis for holographic architectures. A physical proof-of-concept is planned.
The mathematical framework behind holographic storage provides a mechanism for continual learning independent of physical implementation. Multiple weight matrices encoded in a shared volume can be updated without catastrophic forgetting: four independent matrices maintained worst-case error of 0.014 over 200 update rounds with no drift. The problem that biological memory solves through consolidation and rehearsal, holographic superposition solves through the mathematics of interference patterns.
The dual-JEPA cognitive architecture that the published PAM and AAR work is building toward. Two complementary predictors operating over a shared embedding space, one learning similarity structure (world model) and one learning associative structure (episodic memory), with concepts emerging from compression during a sleep/consolidation cycle. The concept discovery results provide early evidence that the compression mechanism works. Integration of the Outward and Inward predictors and empirical testing of specificity-through-intersection is the current milestone.
A reliability-weighted learning mechanism where component adaptability is determined by predictive accuracy. Plasticity reactivates automatically during distribution shifts. An investigation into managing the handoff from bootstrapping to intrinsic cortical function.