Systematic optimization
Relentless pursuit of hardware-level efficiency, zero wasteful overhead, and disciplined resource stewardship across every layer.
Thoughtful engineering
Systems software engineered for efficiency, optimization, and reliability.
Principles
Openness and sovereignty
Transparent, modular architectures that grant developers complete ownership and control without artificial lock-in.
Uncompromising reliability
Predictable execution and verified state under sustained load, designed to run indefinitely without degradation.
Disciplined scrutiny
Serious engineering held to uncompromising standards, where technical criticism is actively welcomed, accepted, and applied.
Work
A Metal-native inference compiler for Apple Silicon. Compatible dense models compile from declared operations and tensor storage into a persistent GPU session.
Laplace separates a model’s operations from the way its tensors are stored, then compiles those descriptions into shared Metal execution programs. An improvement to the common path can serve different compatible layouts and stored precisions.
Compatible dense models run today, including mixed stored precisions and tied embedding and output weights. Chat, one-shot generation, GGUF inspection, and separate prefill and decode timing are available from one command.
Recurrent and routed expert operations have semantic representations, but the shared compiler does not yet lower them. Complete universality is a development goal. Laplace is an active alpha.
A backbone that reuses its layers across computation steps, combining content-selected attention and gated delta memory. One 29,814,640-parameter checkpoint serves text, streams, and experimental Flow refinement.
Interlace combines bounded rational feed-forward features, content-selected attention, gated delta memory, shared depth with learned loop identifiers, and synchronous input, thought, and output streams.
A single-seed 30M comparison against a parameter-matched dense Transformer records 9.86% lower WikiText-2 test loss and 65.7% less training execution time to a 5.5-nat validation-monitor target. Longer-program stream accuracy is 80.47% against 4.69% for the dense control.
Parameters: 29,814,640
WikiText-2 test loss: 9.86% lower
Time to 5.5-nat target: 65.7% less
Longer-program streams: 80.47%
A stateless sparse attention mixer with learned non-bypass routing. Associative recall is 763 of 768 correct, 99.35% mean accuracy across three seeds.
After rotary position embedding, product routing writes keys into bounded causal posting lists and selects a fixed number of remote blocks. Attention over that set is ordinary causal softmax. Natural-language quality, asymptotic linear construction, and fused-kernel speed are not claimed.
Packed routing records a 0.993 fitted timing exponent from 4,096 to 32,768 tokens on four CPU threads. The learned-R2 selected-mask oracle error is 1.43 × 10−6 maximum.
Associative recall: 99.35% mean
Minimum accuracy: 98.05%
Oracle error: 1.43 × 10−6
Timing exponent: 0.993