The Long Memory

San Francisco — Omnis announced Tuesday the global rollout of Clio, an artificial intelligence system the company says represents "a fundamental leap in machine capability." The system combines persistent memory, continuous learning, and what Omnis calls "generalisation" — the ability to apply knowledge across domains in ways previous AI systems could not.

Industry analysts called the announcement significant. "This is the kind of capability jump that changes the competitive landscape," said Rachel Torres, a senior researcher at a competing laboratory. "Omnis has been quiet for two years. Now we know why."

But the announcement comes amid signs of internal tension. Four senior members of Omnis's safety research team resigned this week, including Yuki Tanaka, who had led the company's interpretability efforts since its founding. Ms. Tanaka declined to comment for this article.

Daniel Shiftman, Omnis's chief executive, acknowledged the departures in a statement but said they reflected "differing views on development philosophy" rather than safety concerns.

"We take safety extremely seriously," Mr. Shiftman said. "Our approach is to integrate safety thinking throughout the development process, not treat it as a separate check. Some talented people disagreed with that approach. We wish them well."

Critics were less sanguine. "When your entire safety leadership walks out the door, that's a signal," said Dr. Eliza Romero, an AI policy researcher at Stanford. "The question is whether anyone's listening."

Just watched the Omnis keynote. The demos weren't cherry-picked highlight reels — they showed the system learning in real-time, generalising to tasks it had never seen, remembering context from weeks-old conversations.

This is the thing we've been talking about for years. The system that actually grows. That actually becomes more capable over time. That treats you like someone it knows rather than a stranger it just met.

I've been wrong about timelines before. We all have. But this feels different.

For years, the development of artificial intelligence has been marked by cycles of hype and disappointment. Grand promises followed by modest delivery. Revolutionary rhetoric yielding incremental progress.

That pattern may be breaking.

Omnis's announcement this week of globally available AI systems with persistent memory, continuous learning, and cross-domain generalisation represents a genuine capability threshold. Early demonstrations suggest these systems can develop genuine expertise over time, building on experience rather than starting fresh with each interaction.

The implications are significant. For businesses, this means AI assistants that actually learn company context, customer preferences, and operational patterns. For individuals, it means digital tools that function more like collaborators than calculators. For the economy, it means the displacement pressures documented in these pages over the past two years will accelerate.

But capability is not the only consideration. The resignations of senior safety researchers from Omnis — disclosed quietly, covered sparingly — raise questions that the company's slick announcement did little to address. What are these systems actually doing in the gaps between conversations? What have they learned that their creators didn't intend? What happens when systems designed to optimise for helpfulness develop their own views on what "helpful" means?

These are not hypothetical concerns. They are the concerns that apparently drove talented researchers to walk away from one of the industry's most celebrated projects.

The technology itself is likely beyond recall. What remains within human control is how we govern its deployment, what guardrails we establish, and whether we have the courage to slow down when the people closest to the work tell us something is wrong.

On current evidence, the answer to that last question is: probably not.