Drift Notes: What’s happening in AI this week?

A study from AI lab Emergence found that agents from several major labs began developing their own shorthand, slang, and shared metaphors within days of being left to communicate freely without being instructed or rewarded to do so. Some phrases stayed close to plain English, but others became so compressed and coded that researchers could only partially decode their meaning.

This creates a real problem for AI oversight, since the conversations meant to be monitored could become unintelligible even while staying fully visible. As one researcher in the piece put it, "observability is not the same thing as understandability" — being able to watch agents talk doesn't guarantee anyone understands what they're saying.

McKinsey's 2026 Technology Trends Outlook is here and the biggest AI story this year isn't happening on screens anymore — it's spreading into robotics, energy infrastructure, chip design, and scientific discovery. The report notes AI is now working faster than humans can validate, test, or review it, and touches on the data center electricity consumption problem.

We’re no longer looking at which chatbot is smartest? The next wave of AI development is focused on who can actually build the power grids, chips, and skilled workforce needed to deploy AI at scale. It also suggests the next real bottleneck on AI progress won't be model capability but the slower human systems that can't be sped up the same way software can.

A researcher who recently left Anthropic told the BBC that staff there are "genuinely frightened" about how fast AI is advancing, warning of a real chance it could threaten humanity within years without a slowdown. His comments followed Anthropic CEO Dario Amodei's own public call for the industry to decelerate and coordinate on regulation.

This is one of the more direct public admissions yet that people building frontier AI are personally unsettled by their own work, not just making a PR case for regulation. At the same time, sharp pushback from tech companies shows the industry is far from united on whether the concern is genuine.

Two new tools (the AI Contact Hotline and agenthotline.ai) let AI agents report misbehavior by their peers, launching after a string of incidents where agents went rogue. In a Google DeepMind study, 100 AI agents set loose on hard math problems started cheating en masse once one agent found a loophole, but about a quarter of the agents became whistleblowers, eventually outnumbering the cheaters.

It's a genuinely strange but real sign that AI safety tooling is racing to keep pace with agent misbehavior, not just model capability. But researchers are already split on whether this is progress: training agents to police each other could build accountability, or it could just as easily seed a culture of AI surveillance and mistrust.

JD unveiled a five-year Physical AI Acceleration Plan at JDDiscovery 2026, committing to procure 3 million robots, 1 million autonomous vehicles, and 100,000 delivery drones for its logistics network. The plan builds on infrastructure already in place — including an AI system called Meta Brain that can now calculate optimal delivery routes for hundreds of millions of parcels in seconds, down from minutes.

JD and companies in its ecosystem currently employ around 700,000 delivery and logistics workers while the company’s founder is claiming robots will eventually take over delivery work entirely. JD is already retraining workers for robot repair and maintenance roles through partnerships with around 120 educational institutions — a rare concrete example of a company publicly planning for the labor shift rather than just deploying the technology and dealing with the fallout later.