AI Engineering Signal #48
Patch or isolate affected dependencies now; remote exploitation risk spans most agent deployment stacks.
Signals
AI-generated CUDA kernels silently corrupt training and inference
audit any LLM-generated kernel before it touches a production run.
KV cache benchmarks: q5/q6 beat q8 and q4
inference memory budgets built on q8 assumptions need recalibration.
Claude Design limits now merge with Claude.ai and Claude Code
parallel Claude workflows will hit shared rate limits faster than current capacity plans assume.
ESMFold2 applies bitter-lesson scaling to protein structure prediction
protein ML pipelines built on ESMFold 1 need architecture reassessment.
Latent Space
Atomically precise manufacturing claims programmable covalent bond placement
if replicated, long-term compute substrate assumptions beyond silicon require revision.
ArXiv
The Take
The infrastructure layer is the attack surface: one vulnerable Python package threads through vLLM, MCP, and agent frameworks simultaneously, while AI-generated CUDA kernels introduce silent correctness failures that benchmarks won't surface. Dependency audits and kernel validation gates are now prerequisite work, not optional hardening.
Subscribe
Related Signals