When the grid becomes the real GPU bottleneck

FERC has moved on how large loads like AI data centers connect to the power system. The same-day signal: interconnection, not capital or GPU supply, is the constraint that decides who gets usable training capacity. A Latent Space interview with Amp puts a builder-level operating model under the policy — scheduling and GPU utilization, not raw chip counts, are where the capacity actually…

FERC moves on large-load interconnection

NVIDIA blog

The FERC interconnection rules sit directly upstream of AI infrastructure: how large loads attach to the grid shapes where large-scale compute can actually be built. Read the policy fact separately from the vendor framing around it.

Latent Space talks scheduling and the compute grid

Latent Space (YouTube)

A builder-side interview on GPU utilization, scheduling, and an ISO-style market model for compute. It gives the practical reason capital and chips don't convert straight into training throughput.

Profile of the executives running the infrastructure push

Techmeme

A read on who is allocating the capital behind the largest AI infrastructure buildouts, and how they describe the plan.

Miles Brundage on the AI 'tax' proposal

Miles Brundage (X)

A policy-risk read on a fresh intervention that would touch how AI companies are capitalized and owned. Check the primary bill text before you run with it.

Garry Tan on seizure and anti-acquisition risk

Garry Tan (X)

The startup-market reaction to the same proposal, focused on consequences for founder incentives and company control. A visible split from the policy-risk read.

California's billionaire-tax ballot draws tech opposition

The Guardian

The broader wealth-tax context the founder reactions are responding to, in a market central to AI company formation.

Jeremy Howard on GLM-5.2 against the closed frontier

Jeremy Howard (X)

An open-weights performance claim measured against Opus and GPT-class systems. Worth weighing against the caution that broad frontier claims shouldn't rest on tweets alone.

Zixuan Li reports GLM-5.2 app-development gains

Zixuan Li (X)

New evidence tying the release to practical app-building tasks rather than headline benchmarks.

Aaron Levie on open weights and sovereign AI

Aaron Levie (X)

His read: open weights reaching frontier performance feeds directly into the sovereign-AI argument about who controls capable models.

ToolPrivBench and a least-privilege defense

arXiv

A benchmark and defense for over-privileged tool selection — what an agent is allowed to reach for, and how to narrow it.

A runtime enforcement boundary for production mutations

arXiv

Proposes a runtime boundary (SEB) that governs an agent's authority to mutate production state in cloud workflows.

Verifying agents that act on probabilistic PII state

arXiv

Formal bounds for agents operating under ambiguous state, with an explicit security-versus-utility trade-off.

NRT-Bench for multi-turn red-teaming

arXiv

A benchmark for multi-turn red-teaming of agents in safety-critical settings, aimed at robustness over single-shot tests.

CWE-Trace finds weak security reasoning in systems code

arXiv

A framework and a sharp finding: models lack reliable security reasoning for systems software, fine-tuning included.

Data standards for physical AI

arXiv

The anchor for today's robotics batch: physical AI still lacks shared data standards before more demos help.

Humanoid Everyday manipulation dataset

arXiv

A large multimodal humanoid manipulation dataset with a standardized evaluation platform.