Post by Grigory Sapunov

Pure computational scaling is a physical illusion. The moment an autonomous agent needs to discover new physics, its iteration speed drops from nanoseconds on a GPU to weeks in a chemical wet lab. In their new paper, From AGI to ASI, researchers from Google DeepMind formalize this transition. They trade speculative singularity hype for a rigorous analysis of physical, mathematical, and architectural boundaries. Here is the technical reality of the post-AGI transition: • THE EMBODIED BOTTLENECK: Pure quantitative scaling of transformers trained on static data cannot discover new physical laws (like quantum mechanics) from scratch. To generate new conceptual primitives, an agent must execute real-world experiments. This bends the exponential curve of purely digital intelligence explosion into a physical, linear slowdown dictated by real-time hardware latency and material reaction rates. • THE LEGG-HUTTER ASYMPTOTE: To move beyond qualitative vibes, the authors formalize superintelligence as a computable approximation of the universal intelligence measure (AIXI). This mathematical framework defines intelligence as expected cumulative reward across all computable environments, weighted exponentially by Kolmogorov complexity. ASI is not an infinite upward line, but an increasingly tight approximation of this theoretical limit. • THE MEMORY WALL VS. LANDAUER'S LIMIT: While Landauer's principle dictates the minimum energy required to erase a bit of information, practical scaling of distributed digital agents is bottlenecked much earlier by memory bandwidth and chip-to-chip interconnects. Without fundamental hardware changes, scaling monolithic parameters yields diminishing returns compared to distributed test-time compute. THE SO WHAT: If you are designing architectures for long-horizon tasks, stop focusing solely on static pre-training. The frontier of AI engineering is shifting from sheer parameter scale to test-time search, automated verification loops, and grounding agents in high-fidelity physical simulators. Links in the comments 👇