The image depicts a surreal landscape where whimsical figures attempt to stack bricks to build a skyscraper. The structure defies the laws of physics, twisting and turning in dream-like ways. It's illuminated in an abstract, surreal light, with a color palette of Light Gray, Dark Gray, Bright Red, Yellow, and Blue. Some bricks appear almost melting, falling off as the skyscraper bends impossibly, showcasing the chaotic and doomed attempt.

The always thought-provoking Venkatesh Rao poses the question of what kind of scaling we need for AI. His analogy with building skyscrapers out of bricks-and-mortar is an interesting one. It’s a long read, but worth it.

The part which really resonated with me is when Rao starts talking about governance for AI agents, which needs to be in the form of liberal democracy rather than autocracy. “Regulating [AIs] will look like economic regulation, not technology regulation” he says.This is why you need people who can think philosophically about technology and the future of humanity.

Fascinating.

To keep AI evolving, we need the various heterodoxies to cohere into one or more alternative positive visions of how to build the technology itself, not just creative reframes that make for stimulating cocktail party conversations. Into one or more new idea of what sort of AI we should attempt to built, in an engineering rather than ethical sense of should. As in well-posed, architecturally sound, and conceptually elegant enough to handle whatever we choose to throw at it.

I’m asking the question in the same sense as one might ask, how should we attempt to build 2,500 foot skyscrapers? With brick and mortar or reinforced concrete? The answer is clearly reinforced concrete. Brick and mortar construction simply does not scale to those heights. Culture wars in architecture and urbanism around whether or not skyscrapers are a good idea for society are moot unless you have good options for actually building them.

[…]

The current idea seems to be: If we build AI datacenters that are 10x or 100x the scale of todays (as Sam Altman appears to want to), and train GPT-style models on them that are also correspondingly scaled up, we’ll get to the most interesting sorts of AI. This is like trying to build the Burj Khalifa out of brick-and-mortar. It’s a fundamentally unsound idea. Problems of data movement and memory management at scale that are already cripplingly hard will become insurmountable. Just as the very idea of a 2,500 foot high brick structure is unsound because bricks don’t have the right structural properties, the current “bricks” of modern AI (to a first approximation, the “naked” Large X Models thinly wrapped in application logic) are the wrong ones.

[…]

[Going] back to the analogy to reinforced concrete. [The AIs Rao is arguing for] are fundamentally built out of composite materials that combine the constituent simple materials in very deliberate ways to achieve particular properties. Reinforced concrete achieves this by combining rebar and cement in particular geometries. The result is a flexible language of differentiated forms (not just cuboidal beams) with a defined grammar.

[They] will achieve this by combining embodiment, boundary management, temporality, and personhood elements in very deliberate ways, to create a similar language of differentiated forms that interact with a defined grammar.

Source: Ribbonfarm