The Next Web’s October 17, 2026, account says the Gemini 4 Argon Safety and Security Report found no statistically significant end-to-end acceleration of AI research and development (R&D). The report rates the risk from AI-driven AI R&D as currently low. That’s the central distinction in recursive self-improvement (RSI)—AI systems helping develop their successors: assistance with research is not the same as an accelerated research process.

What the Argon assessment measures

The assessment sets separate thresholds for acceleration and automation. ML R&D acceleration level 1 means substantially accelerating AI progress beyond historical rates; ML R&D automation level 1 means fully automating a Google research team. A synthesis of the assessment says Argon meets neither threshold. The Google DeepMind model card provides the formal framework.

Those thresholds concern substantial progress across AI research and development or automating a research team. They leave room for AI to assist with particular research tasks without meeting either level.

The assessment’s internal index, built from 92 benchmarks, rose by 0.091 points per month before April 1, 2026, and 0.118 points per month afterward. The report says the difference was not statistically significant. The Next Web reports that Google attributed Argon’s position above the earlier trend to additional investment and training compute, rather than AI-driven acceleration.

Argon also solved around half the challenges in an internal AI research-task test, according to The Next Web’s account. That account says a near-perfect score would be needed for full automation.

What Sergey Brin says about models building models

At a Gemini 4 Argon roundtable, Google cofounder Sergey Brin says researchers and models work together on the next model version. He characterizes Google as beginning to get onto an RSI treadmill, as The Next Web reports. His description concerns collaboration; the assessment measures whether that work has accelerated R&D end to end.

Google DeepMind’s Vihan Jain offers a more emphatic characterization of Argon’s place in that process, describing the RSI flywheel as having reached a new gear. That is Jain’s characterization, not a measurement of research acceleration.

NeoTeo previously covered questions about Gemini 4 Argon’s practical coding performance. The Argon assessment addresses a separate question: how AI contributes to developing future models.

Where people still steer the work

The Next Web reports that AI wrote more than 75% of Google DeepMind’s production code, while code changes remained subject to human review before submission. People also continued to choose research directions, design experiments and diagnose training problems.

Fully delegated pull requests accounted for 7.5% in April and 11.9% in September, according to the account; about 88% still required repeated human steering or human-written code. The figures describe a process with extensive AI help and continuing human involvement.