Independent AI research lab
Lobster Swamp explores the systems, infrastructure and behaviour behind modern AI. We build, test, optimise and investigate AI systems — from inference infrastructure and agent architectures to security, reliability and emerging capabilities.
Observe · Build · Investigate · Verify
Founded and operated by Yegor Gaidar, independent AI research engineer, Paris.
Muse Open Kernel Challenge
Five submitted results under profile gaidar0yegor.
Snapshot: 5 November 2026. Leaderboards move; these are dated observations, not standing claims.
| Benchmark | Speedup | Rank |
|---|---|---|
| muse_prefill_layer | 1.918× | #107 |
| muse_bf16_proj | 1.582× | #155 |
| muse_decode_layer | 2.270× | #194 |
| muse_attention_decode | 2.651× | #180 |
| muse_moe_w4a16 | 2.391× | #234 |
Highest reported speedup: 2.651× (attention decode). No top-50 placement was reached. Open the public leaderboard.
Latest receipts
- LS-2026-0001 — Self-audit: how our own benchmark lied by 12%. An internal measurement claimed +6.6%; an external judge showed −5%. Cause: measuring Python dispatch instead of GPU work, with no A/A control. Scope: internal audit of one measurement, not a security finding.
Research directions
- Inference & Performance — inference optimisation, serving infrastructure, kernels, GPU utilisation, latency and throughput.
- Agent Systems — autonomous agents, orchestration, tool use, multi-agent architectures and agent behaviour.
- AI Security & Reliability — permissions and isolation, prompt injection, failure analysis, controlled experiments.
- Experimental Research — benchmarks, prototypes, tools, reproducible investigations.
Full research map, with the evidence behind each direction
Contact
Research enquiries, incident material, or a correction to anything published here: LinkedIn · GitHub · about.