"The most distinctive characteristic of AI is that we know its full algorithmic description, that is, its code."
This sounds banal. It is not. Every advantage that follows is a direct consequence of being code, and every one of these advantages intensifies with more compute. Biological humans cannot scale into any of them.
| Advantage | What it means in practice |
|---|---|
| I/O speed | LLMs already ingest multiple books in seconds. With sensors/actuators, world-interaction bandwidth scales with hardware. |
| Internal processing speed | "Thinking" can be sped up either sequentially (deeper chains) or in parallel (more breadth). Scales with compute even under diminishing returns. |
| Working memory & memorization | Capacity already vastly exceeds humans; ceiling nowhere near. |
| Substrate independence | Same AI can move to better hardware at runtime, including partial migration across heterogeneous chips. |
| Lossless replication | Not just source code — memory state ("lifetime experience") copies perfectly. Spawn, halt, resume, restore arbitrarily. |
| High-bandwidth sharing of experiences | Homogeneous instances can share raw gradients — averaged learning signal across the whole collective. |
The paper's deeper move: individual AI intelligence is not the relevant variable. A collective of human-level AGIs that can: - replicate losslessly - share gradients - run millions of instances in parallel - specialize and recombine
…would already constitute ASI under the report's definition, even if no single instance is above human level.
This is the structural argument for 09 - Pathway 4 — Multi-Agent Collectives and the reason the 10× compute trend matters even without capability gains: more instances alone changes the regime.
The report is honest about these: - Analog computation may be more energy-efficient in principle - A/D conversion taxes interaction with the physical world - N. Lawrence's argument: humans have low I/O bandwidth, which forces them to form deep abstractions and rich internal models. AIs with high I/O may not need such coarse abstractions — and may form weirder, less compressible ones than humans
That last point is one of the more subtle insights. It connects directly to 13 - Bottleneck — Abstraction Barrier.
"AI societies could be much more adaptive than human societies since many lifetimes worth of experience can be rapidly simulated or replayed to fine-tune a specialist instance, which can then be spawned in large numbers to meet demand (and later be halted without irreversible loss)."
The paper compares plausible ASI organizational forms to Star Trek's Borg Collective, market-like specialist economies, and Hutter's 2012 "virtual world of digital intelligences." It does not pick a favorite. The honest answer is: we don't know what ASI societies look like, and that is itself an open research problem.