Sam Mausberg

An experimental LLM compiler and runtime, designed to search GPU kernels, batching, and cache policy together. I worked on it from February to September 2026.

An experimental systems language for CPUs and NVIDIA GPUs, designed to be written by AI agents. It compiles to C++20 and CUDA, and its compiler refuses GPU kernels with shared-memory races, or with barriers that some thread does not reach, before they run.

Claude Code agents wrote most of the compiler, runtime, tests and documentation to my design.

Optimized Blackwell kernels for normalization, FP8 projections, attention backward, and convolution with gating.

Official B200 evaluations on nine problems. The six listed cover 94 workloads, with up to 2.94× geometric-mean speedup per problem over NVIDIA’s agent-optimized scoring baseline.

A GPU kernel index that matches an exact workload to compatible implementations and their published benchmarks.

One catalog serves the website, REST API, command-line client, and immutable bulk exports.

For a specified finite class, one-sided safe guidance can cost linearly many queries where pointwise guidance needs only logarithmically many.

Known finite classes on n-bit strings. The safe-guide results assume bounded stopped reciprocal capacity.

A certified output head reproduces the engine’s greedy token, including BF16 rounding and tie rules, while recomputing about 1.8 vocabulary entries per decision.

Qwen3.5-4B in SGLang on one GH200. Against tuned stock configurations, throughput rises 1.3–4.6% at one request; with 2 to 8 concurrent requests, the fastest configuration is 1.0–3.0% slower.

Returning a quantum memory’s reduced state does not mean it can supply fresh independent randomness.

Closed machines with exact or approximate return of the memory’s reduced state.

Attention outputs can admit compact approximations while a specified class of certificates still needs many cached rows.

The separation holds for the stated certificate class. The checker that passes evaluates the dense reference and is slower than dense decoding.

A shared support can entangle two interferometers. Force replay and response bounds give a criterion for excluding that mechanical channel.

The criterion applies to a specified linear Gaussian apparatus. The mounting example is modeled; no instrument was measured.

Information tasks give a conditional Bell bound. Additional capacity and measurement requirements recover the quantum value.

The capacity bound and the reversible setting interchange are added premises.

For entropy-preserving tasks, possibility is equivalent to a uniform work bound along a vanishing-noise family. A stationary quantum model realizes the resulting entropy order.

The result uses modified task clauses and an explicit quantum completion.

An exponential separation between distribution-independent SQ learning and ordinary, exact probabilistic, and expected-error dimension. It was obtained independently of Patel (2026), who posted a separation first; a qualitative separation also follows from earlier results.

The learner runs in time polynomial in the size of the explicit table. It is not claimed to be polynomial in the input length. Lean checks supporting lemmas only.

At bounded effective time, logistic SGD’s expected error is controlled by label correlation with a fixed Gaussian ReLU feature map.

The bound requires effective time ηmT ≤ 12 and large enough width and step count.

Certified privacy accounting for multi-epoch shuffled Gaussian training, including the full adaptive transcript.

Data are shuffled once per epoch into fixed-size batches with Gaussian noise, and each query may depend on all earlier releases.

Exact softmax attention requires Ω(nd + n²/M) memory transfers in the stated real-arithmetic model, with a stronger bound for score-polynomial exponential arguments.

Bounded deterministic programs using +, −, ×, ÷ and exp; n ≥ 2, d ≥ 2 and an M-word cache with M ≥ d². The unrestricted bound is tight for fixed d.

A randomized algorithm computes three-dimensional softmax attention in near-linear word-RAM time, giving optimal exponent α(3) = 1.

Logarithmic-bit rational coordinates bounded by n¹⁰, scalar values in [−1, 1], and simultaneous additive error at most n⁻¹⁰.

Exact autoregressive sampling with an appendable attention index has polylogarithmic context cost under fixed-rank key geometry.

Each layer's keys share one affine space of fixed dimension, with positive RMSNorm stabilizers. Expected work includes key and value creation, insertion and every precision refinement; the polynomial degree grows with rank.

At every fixed logit rank, chosen-prefix next-bit samples suffice to learn an accurate standalone generator in polynomial time.

Fully supported T-bit distributions whose centered next-bit logits have magnitude at most T and rank at most d at every history cut. The learner may query any prefix.

Executable memory for SQL agents: relations learned from correct queries are reused only after a later task corroborates them.

Transfer and retention gains are not yet shown; the larger pilot is unfinished.

Bounds that use exact row counts let a cached summary certify some attention results without reading the rows. A refinement forest and transactional writes carry accepted results into later decoder state.

On captured model runs, no complete attention call is certified. There is no pretrained speedup and no proof of equivalence to a deployed attention backend.

Exact boundary tests and coupled moment bounds certify some attention outputs from moment summaries in constructed cases.

No coordinate passes on the trained Qwen trace, and every measured complete GPU path is slower than fused dense attention.

Supporting results and constructions for Erdős problems 20, 61, 119, and 885, formalized in Lean 4 and mathlib.

Partial results and unfinished proof developments; the main conjectures remain open.