Same world.
Different word order.
Train two small language models from identical weights. Change the order of English clauses, then compare what the models actually write.
01 Language profiles
same factsDrag the clauses or use the arrows. Each block remains a grammatical English sentence. Only the order changes.
The first run checks numerical gradients and causal masking. Keep this tab open while training.
02 Live corpus preview
Authored micro-stories, not human observations. The action patterns are taught through next-token text, not discovered moral rules.
03 Learning, not simulated progress
Ask. Compare. Inspect.
These are small, from-scratch language models—not pretrained chat assistants. Use the story examples first, then edit the task.
Input task
Relative candidate scores normalise length-averaged log-likelihoods over this list only. They are not calibrated probabilities of real-world actions.
Watch a prefix unfold.
Read the model’s next-word distribution at each input position, alongside actual attention weights. These are model measurements, not a visualisation of human thought.
Next-word probability
full vocabulary normalisationLast-layer attention
Heads averaged. Rows query earlier columns. The empty upper triangle is the causal mask. Hover for token pairs.
Prefix readout
Token, highest-probability next word and distribution entropy.
Nothing behind the curtain.
Actual weights, gradients, source code and provenance. This is a controlled computational language experiment.
What is controlled
- Initialisation
- Byte-identical weight arrays for both arms. No dropout.
- Experiences
- Same event IDs, same minibatch order, same continuation text, same number of non-padding tokens.
- Language treatment
- Natural English clause order. Clauses are not omitted or rewritten. Word counts match exactly.
- Held-out data
- Exact story combinations are disjoint. Templates and vocabulary intentionally overlap.
- Evaluation
- Common prompts are contrasted with native-order prompts. Shared canonical training exposure is configurable.
- Interpretation
- One pair is exploratory. More seeds and same-grammar controls are needed before attributing a systematic effect to the treatment.
What this does not claim
No proof about human brains, national cultures, political beliefs or moral behaviour. No pretrained world knowledge. Word order here means the order of complete English clauses—not a fully constructed SOV or VSO language.
Generated explanations are text predictions, not access to an internal reasoning process. Attention weights are not a causal explanation. Out-of-vocabulary words are explicitly shown rather than silently understood.
Training examples contain authored action consequences. Their equal use across arms is controlled; they are not value-free examples of the real world.
Numerical validation
Identical initialisation • future-token leakage • finite-difference gradients • actual loss reduction • checkpoint and optimizer restoration.
Not run in this browser.
Kernel inspector
CUDA source
WGSL reference
20 CUDA kernels implement embedding, tiled matmul, RMSNorm, causal attention, SiLU, cross-entropy, all gradients, clipping and AdamW. No inference API or hidden CPU training.