Bernardo Kastrup on the anticipation for Google's Gemini 3 release (mastodon.nl)

Bernardo Kastrup weighed in ahead of Google’s much‑anticipated Gemini 3 release, arguing that massively scaled neural networks can spontaneously develop structures equivalent to reasoning. He suggests that even without explicit rule‑based engines, sufficiently complex statistical networks naturally generate logical patterns that resemble the operation of reasoning, which could mean current scaling trends lead to genuine AGI — and potentially super‑intelligence — within a few years despite mainstream skepticism. Technically, Kastrup’s point centers on emergence: deep models don’t need hand‑coded rules to exhibit logic‑like behavior if their parameter counts and training complexity reach a threshold where statistical patterning organizes into stable, reasoning‑capable structures. For the AI/ML community this sharpens focus on interpretability, benchmarking, and alignment — small incremental scale or architecture changes might produce large, hard‑to‑predict capability jumps. The perspective elevates urgency around safety research, monitoring of capability growth (especially in large multimodal models like Gemini), and policymaking to manage risks from rapid, emergent advances rather than incremental feature additions.