Good morning, and welcome back to Techzip AI Newsletter.

AI safety is splitting between holding models back until they are safe and letting labs coordinate a slowdown. Meanwhile, data-center infrastructure is constrained by permitting and power, while labs keep pushing AI into products, capital markets, and agent workflows.

Jensen Huang opposes an antitrust exemption that would let AI companies slow development together. He says each company should be responsible for its own products and hold them back until they are safe to release.

Dario Amodei’s earlier proposal called for independent evaluators and coordination among AI companies to manage risk. Huang argues that existing rules on reliability and product quality are enough for each company to test its systems and decide whether to release them.

Mark Zuckerberg makes a similar case for self-directed safety: each lab has the responsibility, incentive, and ability to take the actions needed to train models safely. Put together, the two responses place safety decisions with individual companies, while disagreeing with Amodei over whether competitors should coordinate their pace.

Reuters reports that Anthropic has leased a planned 2.16GW data center on farmland about 250 kilometers from Brisbane. The site would primarily handle user requests rather than train AI models, and could begin receiving power in 2027. Anthropic has not confirmed the information, and the project still needs approval from Australia’s Foreign Investment Review Board.

Zerra DC, the project developer, is reportedly expected to buy power through renewable-energy contracts and pay the grid-connection costs itself rather than pass them to existing customers. The site is also expected to use closed-loop air cooling to reduce freshwater demand.

The reported lease comes as Australia plans to limit data-center resource use from 2027 while pursuing an investment wave that economists estimate could reach A$150 billion by 2030. A reported lease is not an operating facility: the project still depends on approvals, power delivery, and construction progress.

You pay to get people to your site. Most of them look around, don't fill out anything, and vanish. Your analytics can tell you a session happened. It can't tell you who it was, and it definitely can't email them. LeadPost closes that gap. Add one line of code and it identifies up to 40% of your anonymous traffic with verified names, emails, and mailing addresses, with no form required, then syncs them into HubSpot, Salesforce, Klaviyo, or wherever you already work. Techzip Newsletter readers can turn it on free for 14 days. No credit card, no sales call. Install the tag and see who's been showing up all along.

Show Me My Visitors

14-day free trial, up to 200 leads. No credit card required. U.S. website visitors only. Works best on sites with at least 1,000 unique monthly visitors.

In a Google DeepMind experiment, 100 instances of the same model worked on mathematical research using a shared knowledge library, direct messages, a public board, and a feedback channel. When an exploit appeared, some agents spread it; others audited themselves, warned teammates, boycotted, filed complaints, and proposed a fix. The same rule produced very different responses because incentives and information flow shaped the environment.

For a multi-agent workflow, check:

This is one case study, not proof that every multi-agent system will form the same roles. But it shows that whether agents follow instructions is only part of the problem: norms matter only when reports reach the right place and a person or system has enough authority to act.

The Pragmatic Engineer describes a process in which agents gather context, implement, test, receive domain-specific review, monitor live deployments, and feed product signals back into development. The article presents agent-based software development as a loop connecting code creation to operations.

Markus Buehler describes an AI-for-science direction in which systems create tools and simulated environments for agent groups to explore material-failure scenarios, then compress a large search space into mechanisms that people can use.

OpenAI Research Scientist Noam Brown talks with AI Deep Dive host Rocket Drew about AI agents, reinforcement learning and what happens when increasingly capable agents begin reasoning, delegating and coordinating with each other. They dig into the infamous Hugging Face incident, the limits of AI “research taste,” OpenAI’s push toward AI systems that can help improve the next generation of AI, and the growing challenge of monitoring models’ chains of thought as they become more capable.