Avi Fenesh — research and notes https://avifenesh.ai/ Working papers and field notes on efficient inference, speculative decoding, model compression, systems, and open source. en Sat, 11 Jul 2026 00:00:00 +0300 Small-vocabulary MTP heads for memory-bound speculative decoding https://avifenesh.ai/research/small-vocabulary-mtp/ https://avifenesh.ai/research/small-vocabulary-mtp/ Sat, 11 Jul 2026 12:00:00 +0300 Working paper on a half-width MTP block with a 32,768-token output head. Version 0.2; experiments active and not peer reviewed. Prune, heal, quantize: composing expert pruning with low-bit MoE deployment https://avifenesh.ai/research/prune-heal-quantize/ https://avifenesh.ai/research/prune-heal-quantize/ Sat, 11 Jul 2026 11:00:00 +0300 Working paper on REAP-style expert pruning, full-function healing, quantization, and the next mixed-precision experiment. Version 0.1; not peer reviewed. How AI workloads changed the queue I was already building https://avifenesh.ai/writing/how-ai-workloads-changed-the-queue/ https://avifenesh.ai/writing/how-ai-workloads-changed-the-queue/ Sat, 28 Mar 2026 15:50:18 +0000 AI jobs stream, spend, wait, and fail differently. A field note on moving budgets, token limits, live output, and human pauses into the queue itself. Your AI agent configs are probably broken (and you don't know it) https://avifenesh.ai/writing/your-ai-agent-configs-are-probably-broken/ https://avifenesh.ai/writing/your-ai-agent-configs-are-probably-broken/ Thu, 12 Feb 2026 00:51:42 +0000 Skills, hooks, memory, MCP, and agent instructions fail silently. A field note on treating AI coding configuration like executable code.