On October 15, 2026, DiarioBitcoin reported that an internal Google DeepMind evaluation found no significant structural acceleration in measured AI progress after April 2026. The account said the evaluation, included in the Gemini 4 Argon safety and security report published the previous day, covered 123 models and 92 benchmarks. DiarioBitcoin's report also gave a higher monthly trend after April—but said the evaluation did not treat the increase as a significant structural shift.
What the evaluation reportedly found
The finding concerns whether AI was accelerating progress in AI, not simply whether researchers were using AI tools. Those are different questions: assistance with coding can speed up parts of a workflow without showing that AI systems are autonomously advancing AI research.
DiarioBitcoin described the analysis as covering 123 models and 92 benchmarks, using the General Capability Index (GCI). The account attributed the reported analysis to Google DeepMind researcher Haydn Belfield.
The GCI trend rose, but the reported conclusion did not change
The reported monthly public-model trend was 0.091 GCI points before April 2026 and 0.118 afterward. The second figure is higher, not lower, but DiarioBitcoin said the evaluation did not consider the difference a significant structural acceleration.
| Period | Monthly public-model trend |
| Before April 2026 | 0.091 GCI points per month |
| After April 2026 | 0.118 GCI points per month |
The difference between the two reported rates is 0.027 GCI points per month. That is about 29.7% above the earlier rate, calculated from the figures in DiarioBitcoin's account. It quantifies the change in the reported rates; it is not a separate study result or a measure of statistical significance.
Code assistance is not autonomous research
DiarioBitcoin reported that AI tools contributed to more than 75% of Google DeepMind's production code, while 88% of code changes still required human review. Those figures describe AI participation in coding and the review process; they do not, by themselves, establish that AI systems were independently driving research progress.
The same account said fully delegated changes rose from 7.5% to 11.9%. It also described survey responses from 12 senior Google DeepMind researchers: automation was more effective on well-defined, verifiable tasks than on open-ended ideation and complex planning.
NeoTeo previously covered questions about Gemini 4 Argon's practical coding performance and Google's response. The evaluation discussed here addressed a different issue: whether measured AI progress accelerated structurally after April 2026.
What the engineering benchmark measures
The reported RE-Bench evaluation covered 74 machine-learning engineering tasks. DiarioBitcoin said Gemini 4 Argon solved about half, while Gemini 3.7 Flash scored 27%. Those results describe performance on that benchmark, not a universal measure of research ability.
A separate review of 1,389 Gemini 4 Argon executions reportedly found lower success on complex tasks without human intervention. Together, the benchmark and execution findings concern task performance and the role of human involvement; they do not extend the evaluation's conclusion beyond its reported measures and period.
