Base Compute hands kernel tuning to agents; the carryover claim is the unmeasured part

The load-bearing sentence in the August 20 write-up is not in the cost table. It is the statement that Base Compute is trying to preserve what its agents learn on one optimization job and apply those findings to later models and processors ^[4]. That is the difference between a consulting practice and a product. It is also the part the published experiment cannot speak to, because all three agents were handed the same already-completed Nemotron port, the same evaluation setup and the same eight-hour limit ^[5]. What was measured is how cheaply a model searches a fixed problem. Whether anything it finds survives contact with the next processor is asserted, not shown.

The reason that labor exists at all is architectural. BaseRT, introduced on July 1, is a native C++ runtime that goes at Apple's Metal API directly with handwritten kernels, fused operators and hardware-specific dispatch rather than routing through MLX, PyTorch or Core ML ^[11]. Declining the framework layer buys control and hands you the bill: every new architecture can otherwise mean another round of operator implementation, numerical verification and device-specific optimization by systems engineers ^[3]. The first public target, NVIDIA's Nemotron 3 Nano 30B-A3B at 52 layers, was a model BaseRT did not support ^[13]. Agentic kernel work is the repayment plan for the framework Base Compute chose not to use.

Now the arithmetic the write-up leaves on the floor. Claude Fable 5 took the top score with 50 experiments at $265; Kimi K3 reached 97 percent of it in 21 experiments for $77 ^[6][7]. So the final three percentage points cost $188, about $63 a point ^[15]. Per experiment, Fable ran at $5.30 against Kimi's $3.67, roughly 44 percent more for each attempt, and it made 2.4 times as many attempts ^[16][18]. Line the three runs up and the curve is steep then flat: going from 12 experiments to 21 bought 11 points, and the next 29 experiments bought 3 ^[17]. Any budget set by "run it to the eight-hour limit" is mostly paying for search that repeats itself.

GLM 5.2 is the awkward row. It reached 86 percent in 12 experiments through BaseRT on an M3 Ultra Mac Studio, and Base Compute assigned that run near-zero marginal cost ^[8]. "Marginal" is doing heavy lifting there: the Mac and the eight hours it was occupied are capital and opportunity, not zero.

One detail cuts in Base Compute's favor. All three agents flagged early changes involving state-space-model prefill, decode dispatch and the Mamba convolution path before their results diverged ^[9]. If the profitable moves are that legible, they look like properties of the model-hardware pair rather than of the agent, which is exactly the shape a carryover story needs.

Distribution is the least speculative part: the public BaseRT repository ships command-line tools, the model format, a public C API and Python, Node, Rust and Swift bindings under Apache 2.0, and it can serve an OpenAI-compatible API ^[12]. B:OS itself has no disclosed revenue, contracts or pricing, and the experiment does not establish what anyone will pay for automated optimization ^[10].

The number that would settle this is not a faster kernel. It is a second port whose agents start from the first port's findings, reported as experiments-to-parity. If port two still needs 21 experiments, Base Compute has cheap labor, not accumulated knowledge.