AI Engineering Signal #130
Xiaomi MiMo-V2.6-Pro tops open-weight model rankings at $3M training cost
Signals
Xiaomi MiMo-V2.6-Pro tops open-weight model rankings at $3M training cost
Rerun local-inference capacity plans and proprietary API budget assumptions now.
Latent Space
Cloudflare Python Workers now generally available
Python edge functions can replace latency-sensitive microservices without server provisioning.
Simon Willison
Meta Muse blocked from using Amazon.com
Agentic shopping routes fail unless merchant APIs or access agreements exist.
TechCrunch
AI coding made CI a bottleneck at Linear
Plan runners and merge queues for agent-scale commit volume, not human cadence.
Web
OpenAI AI resolves over 100 open math problems
Research workflows need formal proof verification before accepting machine-generated results.
TechCrunch
Alibaba plans 5T–10T parameter model and new chip
Compute procurement must model Chinese domestic accelerators and power ramp, not only NVIDIA supply.
Reuters
China breaks ground on supercritical CO2 molten salt plant
Novel thermal storage can firm renewable power for datacenters; watch capex and dispatch models.
Web
The Take
Open-weight capability has become cheap enough to force a rethink of proprietary API budgets, while agent-generated commits move bottlenecks into CI, edge runtime, and merchant API access. Frontier math and non-Western compute/energy milestones are arriving faster than most evaluation and procurement plans assume.
Subscribe
Related Signals