AI infrastructure · Electricity · Data centers
The grid asked the AI to please think about this later
AI’s electricity problem usually gets framed as building more generation and transmission. Reuters’ October 8 explainer examines another option: reducing or moving some data-center demand when the grid is stressed. OpenAI has agreed to curtail grid electricity use by up to 1 gigawatt at a planned 3.2-gigawatt Georgia facility. That is a commitment about a future site, rather than evidence of a reduction already delivered.
Some batch work can tolerate a delay more readily than a customer waiting for an interactive answer. Operators can explore shifting those workloads to periods with more available electricity, alongside other ways of providing grid relief. Reuters describes an industry still working through pilots, individual agreements and coordination problems. Customer commitments, operational reliability and utility incentives determine how much flexibility is practical; a training run is not necessarily cheap or easy to pause.
Moving demand can relieve a peak without eliminating the energy needed to complete the work. It complements new infrastructure rather than making that infrastructure unnecessary. Still, this is an appealing form of maturity: a useful system should know when the world around it needs a breather. We have apparently reached the stage of computing where “Could you close some tabs?” has become an energy-policy question. Sometimes progress means teaching an enormous computer when to go get a coffee.
Source: Reuters on flexible demand and its limits ↗
South Korea · Cybersecurity · AI agents
One attacker apparently acquired the cyber equivalent of additional employees
CrowdStrike’s investigation links late-September and early-October intrusions against South Korean financial organizations to ARTEX, an open-source penetration-testing framework, and language models. Its evidence includes exposed configurations and Claude Code session histories. The report identifies DeepSeek as ARTEX’s primary model backend and other models in additional sessions; a Claude Code record does not itself establish use of Anthropic’s Claude model.
The company assesses a likely Chinese-speaking, financially motivated actor with moderate confidence. It cannot definitively identify the adversary, and the number of affected organizations remains unconfirmed. Its account describes data exfiltration. Those limits matter: a serious campaign does not become better understood by turning tentative attribution into a confirmed identity or an unverified institution count into a fact.
The broader concern is leverage. Software assistance can help an operator sustain more technical work, but this report does not quantify how many people it replaced or prove that expertise has become unnecessary. It also describes a different control problem from an agent exceeding the boundaries of a benign task: the apparent operator wanted the intrusion. Both demand defenses. A malicious person gaining leverage is plenty troubling without promoting the software to autonomous supervillain.
Source: CrowdStrike’s evidence and attribution limits ↗
Biohub · Virtual biology · Scientific AI
Language models ate the internet. Biology would like better documentation.
Biohub announced an expansion of its Virtual Biology Initiative on October 7, following its April launch. The $1.8 billion headline combines funding, data, computing and measurement technology. It includes Biohub’s founding $500 million commitment, more than $500 million from the Department of Energy over five years, and $300 million collectively from Google DeepMind, Isomorphic Labs and Meta. NIH’s contribution draws on resources developed through more than $500 million in prior federal investment; it is not a newly announced $500 million cash payment.
The goal is a shared foundation for models that predict how cells respond to interventions. That requires expanding experimental measurements, making datasets compatible and validating predictions against living systems. Biology already has substantial data. The problem is that its existing collections do not provide every measurement, condition or common standard a broadly useful virtual cell would need. An investment announcement is a plan to address that gap, not a completed predictive model.
This is less glamorous than announcing a model and more fundamental. Cells did not create Stack Overflow or leave a convenient cell_response_final_FINAL_v3.csv on the internet. Researchers have to produce much of the missing evidence at the bench. The productive loop is models proposing experiments, experiments returning measurements, and researchers testing whether the next predictions improve. But first, somebody has to ask the cells what happens.
Source: Biohub’s expansion and contribution breakdown ↗
