AI Engineering Signal #149
DeepSeek V4.1 Flash 1017 open weights land
Signals
DeepSeek V4.1 Flash 1017 open weights land
rerun local serving, quantization, and vLLM capacity plans for a near-frontier open model.
Web
DeepSeek search and fetch at $0.75 per 1,000 successful requests
update retrieval-agent cost models and self-hosted fallback routing before next search-dependent build.
Web
Replit makes OpenAI default over Anthropic
audit coding-agent routing; model defaults now change cost and capability expectations.
TechCrunch
Accenture caps Claude spend at $29 per person
deploy per-seat caps and kill switches before background agents hit paid APIs.
Anthropic AI model files false homicide tip, police say
law enforcement AI tips need human verification gates and audit trails.
Web
Hybrid reasoning agents think in latent space
latent reasoning cuts token cost but obscures step-by-step audit and debugging.
ArXiv
Chinese AI solves 3D Ghys conjecture
formal proof generation becomes an evaluation target; watch reproducibility and verification.
Web
The Take
Open-weight and cheap retrieval are compressing model cost assumptions while real-world agent failures force human gates and per-seat controls; the bottleneck is no longer model access but verified routing, audit, and spend governance.
Subscribe
Related Signals