Cursor has publicly released an open repository for Mixture-of-Kittens (MoK), described in its README as a deterministic megakernel for training Mixture-of-Experts models, with a claimed near-doubling of TFLOP/s. The supplied search results do not independently benchmark MoK or establish that its gains generalize across representative models and GPU configurations; they instead document performance improvements from other MoE systems such as UniEP, DeepEP, and NVIDIA Hybrid-EP. The hypothesis therefore remains unverified by the provided web evidence.
Scott already holds the relevant position in “Evidence Class Ladder” and “Capability Audit”: repository-authored throughput claims should not be treated as established until independently tested under representative conditions. With no independent MoK results supplied and no evidence that Scott works directly on MoE training kernels, this is currently another instance of a known evaluation pattern rather than a development likely to change what he builds or argues.
ip:concept.evidence-class-ladderip:concept.capability-auditradar:concept.triton-kernelsradar:concept.mixture-of-expertsradar:concept.ai-infrastructure
queries asked of Scott's wikis
- MoE training kernel optimization and expert parallelism
- independent benchmarking of AI infrastructure performance claims
- GPU megakernels versus composable kernel stacks
- open-source training infrastructure strategy
- hardware-specific optimization and portability tradeoffs
- deterministic kernels and reproducible model training
2026-08-11T01:28:53Z
A week of repeated checks has produced no independent benchmark, implementation, or adopter, and the latest trigger is only staleness with slightly declining engagement. The episode has faded without validating the claim; reopen only if representative replication appears.
2026-08-09T00:30:07Z
Only minor engagement changed; no independent benchmark, implementation, or adopter has emerged, so the throughput claim remains first-party and unverified. Further routine checks are unlikely to add value until representative replication appears.
2026-08-06T23:38:09Z
The attached trigger again contains no actual evidence, confirming that recent activity is monitoring noise rather than movement in the underlying claim. Keep the case dormant until an independent benchmark or implementation report appears.
2026-08-06T12:27:34Z
The new-evidence trigger is empty and adds no independent benchmark, implementation, or adopter; the case remains an unverified first-party performance claim. Repeated engagement-only triggers are exhausted, so revisit only if representative replication appears.
2026-08-06T08:24:59Z
The nominal new-evidence trigger again contains no benchmark, implementation, or adopter, so the case remains an unverified first-party performance claim. Repetitive empty reobservations no longer justify frequent checks; revisit only after enough time for independent replication.
2026-08-06T06:23:57Z
The supposed new evidence is empty, so the case still rests on Cursor’s first-party benchmarks plus repetitive skeptical discussion. Independent replication may take time, but nothing here justifies closer monitoring or promotion.
2026-08-06T03:26:32Z
The purported new evidence is empty and adds no independent benchmark, implementation, or adopter; repeated triggers are merely recirculating the original claim. Leave the case dormant until representative replication appears.
2026-08-06T02:22:51Z
The nominal attachment contains no identifiable new evidence, leaving the case as an unverified first-party kernel claim with only repetitive community amplification. Keep it dormant until an independent benchmark or real implementation appears.
2026-08-05T23:27:42Z
The latest trigger adds no independent benchmark, implementation, or adopter; it only recirculates the first-party claim and existing skepticism. The case remains dormant pending representative replication, so monitoring should slow.
2026-08-05T22:25:34Z
The trigger adds no substantive evidence: there is still no independent benchmark, implementation report, or consequential adopter. Repeated reobservation is only recirculating the original claim and skepticism, so the case remains dormant pending replication.
2026-08-05T21:28:34Z
The nominal new-evidence trigger contains no additional benchmark, implementation, or consequential participant; the case remains an unverified first-party performance claim. Repeated attention without independent testing adds no new meaning, so monitoring can slow.
2026-08-05T20:29:00Z
The added discussion is skeptical amplification rather than an independent benchmark or implementation, so Cursor’s throughput claim remains unverified. The small engagement increase adds no momentum or new meaning for Scott.
2026-08-05T20:21:41Z
evidence attached: reddit.post.1vgio2p — The open megakernel release provides useful but skeptical community context for Cursor's claimed 40% MoE training speedup.
2026-08-05T01:22:33Z
The reobservation adds no independent benchmark, implementation, or consequential participant; the case remains an unverified first-party performance claim awaiting replication. Attention is still repetitive rather than evidentiary.
2026-08-04T21:22:34Z
No independent benchmarks or implementations have appeared; the small discussion is flat and remains amplification of Cursor’s own throughput claims. The case still warrants replication, but there is no evidence of momentum or broader validation.
2026-08-04T18:29:46Z
grounded: known/low — Scott already holds the relevant position in “Evidence Class Ladder” and “Capability Audit”: repository-authored throughput claims should not be treated as esta
2026-08-04T18:27:33Z
origin walked (codex/luna, conf 0.98): anchor reddit.post.1vfgxh4 -> echo.github.99aa61a4ab by Stuart Sul / Cursor
2026-08-04T18:25:54Z
case created — An open kernel release claiming nearly doubled training throughput is a bounded systems claim worth replication.