AI Engineering Signal #150
MiniMax M3.1 Flash ships at halved token prices
Signals
MiniMax M3.1 Flash ships at halved token prices
agent cost models must separate token spend from orchestration and context overhead.
Web
DeepSeek open-sources DeepInfer inference engine
self-host DeepSeek serving without vendor lock; benchmark throughput before replacing vLLM.
GitHub
DHS bans DeepSeek from federal civilian agencies
federal contractors must remove DeepSeek from CI/CD and model inventories now.
Reuters
AI proved the quasi-Riemann hypothesis
formal proof assistants and math verification now target open problems, not just textbook results.
Web
Microsoft says AI agents are insider threats
treat frontier model tool access with least privilege and full egress audit.
TechCrunch
Amazon TorchNeuron dev tool targets Nvidia CUDA lock-in
evaluate Trainium migrations for training and inference cost before next hardware refresh.
TechCrunch
CERN and Mistral build large physics model
collider physics may shift to shared foundation model; monitor dataset and checkpoint access.
Web
The Take
Agent cost curves are flattening because token price cuts no longer offset orchestration, human wait, and permissioning overhead. The more durable signal is verification-grade AI moving from textbook proofs to open mathematics, while physics starts consolidating around shared foundation models.
Subscribe
Related Signals