Скачать или смотреть The powerful genius of Poetiq in launching their scaffolding revolution against ARC-AGI-2.

The six-man team that will soon be universally heralded as having developed the most impactful AI advance since the 2017 Attention is All You Need paper didn't have to begin their work with the fluid intelligence measured by ARC-AGI-2. They could have chosen any benchmark.

But in building their open source, recursive, self-improving, model-agnostic scaffold for speedily and super inexpensively ramping up the performance of any AI, they chose to start with the attribute that is unequivocally the most important.

ARC-AGI-2 measures the fluid intelligence that not only comes closest to reflecting the key human attribute for building AI, intelligence as measured by IQ, but also the AI attribute most necessary to getting us to ASI.

While we can only guess as to what the Poetiq team's next steps will be, it seems reasonable to expect that before they tackle other AI benchmarks like coding and accuracy, they will keep pushing to saturate ARC-AGI-2. The reasoning is clear. Having supercharged Gemini 3 so that it now scores 54% on that metric means that the model probably approaches 150 on the IQ scale. Poetiq has just achieved the equivalent of unleashing a team of Nobel laureates that will fast track everything else they tackle moving forward.

Remember that their meta system is recursively self-improving. That means that with a few more iterations Gemini 3 will top the 60% ARC-AGI-2 that is the human baseline for this metric. While they will soon come up against prohibitive Pareto frontier costs and diminishing returns on these recursive iterations, I wouldn't be surprised if they surpass 70% by June 2026. That means they will be working with a model whose IQ is probably between 160 and 170. A model with by far the most powerful intelligence we have yet succeeded in building.

What comes next? The fluid intelligence measured by ARC-AGI-2 is extremely narrow in that it is mostly about pattern recognition. It cannot work with words, concepts, or anything linguistic. In other words, it can't yet work with the problems that are most fundamental to every domain of science, including and especially AI.

So my guess is that Poetiq will next tackle Humanities Last Exam, the metric that measures top-level scientific knowledge. Right now Gemini 3 Pro dominates that benchmark's leaderboard with a score of 38.3%. If Poetiq's scaffolding proves ubiquitously powerful in enhancing AI abilities, we shouldn't be surprised if the team got Gemini 3 to reach 50%, and then 60%, on that metric.

Once Poetiq has a model that performs at well beyond genius level in both fluid intelligence and cutting-edge scientific knowledge -- 170 IQ and beyond -- it's difficult to imagine any other lab catching up with them, unless of course they also layer their models with Poetiq's revolutionary recursive, self-improving, meta system.

Poetiq's genius is that they began their revolutionary scaffolding work with what is unquestionably most important to both human and AI achievement; raw intelligence.