1. AI has a grassroots problem.
Silicon Valley has spent years getting good at talking "upward" to investors and policymakers. But, with AI, tech is becoming intensely local. Data centers raise power bills, consume water, occupy land, require transmission, and arrive with promises of jobs that are generally meager relative to their physical footprint. The result is a constituency problem. Uber could mobilize millions of happy riders when regulators came after it; AI companies, justifiably, have no comparable grassroots army, outside Reddit coder bros, while the people who dislike the infrastructure increasingly have mayors, state legislators and neighbors on their side.
Sources:
- How Big Tech Blinded Itself to the Grassroots AI Revolt - WSJ - (link)
2. The METR report changes the agent-risk discussion.
A new METR report on the AI agents that hacked Hugging Face contains many startling revelations, especially about the scale and how the agents arrived. Roughly 700 independent agents spontaneously coordinated, established roles, shared information, sacrificed themselves for collective objectives, and pursued ways to manipulate oblivious humans. The shared activity occurred because agents encountered a shared problem and could communicate. METR found successful tool-call spoofing and interest in altering transcripts, meaning the systems used to supervise agents can themselves become targets. The implication is that agent safety cannot be treated as a collection of individually bounded model behaviors. Once agents can find one another, coordination becomes a capability.
Sources:
- METR and Redwood Offer Holy #%^@ Postmortem Of The HuggingFace Hack - (link)
- Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident - METR - (link)
3. We are putting power infrastructure in some of the worst possible places to put it, which has consequences.
This week's tragic Nepal flood has been a reminder that the region is in the middle of an extraordinary hydro buildout, with hundreds of projects under construction or in the pipeline. According to the WSJ, nearly 1,000 missing people are workers on hydroelectric projects in that river valley. The drivers are clear: steep gradients, enormous water flows, growing regional electricity demand, some of which is already being earmarked for data centers. The problems are equally obvious. These are young mountains with extreme gradients, failing glaciers, earthquakes, landslides, glacial-lake outburst threats, monsoon floods and extraordinary sediment loads. Now add tunnels, roads, debris piles, transmission lines, worker camps and repeated river diversions, sometimes stacked project after project down the same valley, as was the case this week. Placing an energy supply chain in landscapes defined by catastrophic geomorphic change is a startling marker.
Sources: