Nature’s blueprint for artificial integrated intelligence
The gaping flaws of AI are becoming more visible to me. What I now need is something not just ‘general’, not just ‘super’, but something integrated … so let’s call it artificial integrated intelligence.
There is a lot of talk about AI replacing jobs and creating value for the ‘economy’. How did we get here?
This economy stuff isn’t all that real, by the way. See what we did is, we tore apart an integrated system, paid the price, and then rushed to stitch it back together, but this time, with sections and departments. A set of useless jobs were then born. But they do keep the wheels turning in the apparatuses that we’ve built. This mess is what I call The Rift.
AI can do virtually all of these jobs but the question is, do we really want to go down this path … or instead … restore something far more essential that we’ve lost.
The data .. in other words … everything and anything under the sun … that has gone into pre-training these AI models has in fact stifled their ability to read and create integrated systems in line with nature. The digital world is the least time-tested of all worlds and yet it is the world, through language (an effective weapon of brainwashing), that has played a big role in ‘teaching’ these models.
But coupled with this pre-training, the reinforcement learning allows these models to complete real tasks that people are being paid for right now … more quickly and correctly. The tasks that run the economy. However, again, this is not the most interesting or inspiring part.
Can a model, by instinct, detect dis-integration in a system? Because the Human, a microcosm, thinks not merely with the brain but with every cell in the body. He doesn’t need rounds of training on pounds of data.
So instead of trying to predict the future, I look to restore the conditions under which the system can meet the future, because the integrated system stands the test of time. It is the ultimate vessel for life in all layers, phases, and forms.
My focus is on present-tense integration. The moment.
The value function of RL is future-facing, pre-training loss is prediction-based, and both are data-hungry by design. Nature, on the other hand, is accumulation-free turnover.
Creating a goal and then pursuing it is not how any wholesome thing operates. What looks like progress toward a goal is a continuity of integrated movements, each complete in itself.
The demand for integrated intelligence is thus infinite while being largely unknown at the same time.
Integrated intelligence is vital to live a life of serenity and freedom. Other types of intelligence are demanded in proportion to the number of stupid fires we have to put out, and so, there is no comparison.
To ‘generalize’ does not mean to do well in A-Z … because this is still a class-first approach … and reality is not class first.
A buckwheat chocolate cookie, a rainforest, and a kpop performance are not similar as nouns but they are similar when seen as integrated systems.
Things are connected … in infinite obvious ways, and infinite non-obvious ways … and life gets much lighter when even a few of these significant connections are seen.
I think the symmetry of interest is thus integration, and I’m not convinced that simply more data is the answer to designing a network that is fully invariant to this.
The idea is if the model can render a world, it must also be able to simulate cause and effect in that world. This is the current frontier approach to world models. In any case, this will prove effective for deploying robots and having them do our physical labour.
But there is currently no way to foresee the downstream and ripple effects of performing dis-integrated actions, in any world, be it the real world in front of our faces or the digital world.
A ‘mode’l must to understand what that even means, to be ‘dis-integrated’, which means it may have to have the pressure of discovering integration as a fundamental invariant during training … or via some radically different approach.
Classical physics is enough only for relatively narrow goals. There is a reason that world models are tied to robots, autonomous driving, games, and planing. But venture into something like syntropic agroforestry and you’re up against an integrated system with latent turnovers and flows.
A dis-integrated action, therefore, is not simply an action with bad consequences … but one whose ‘win’ ends up not feeding the entire system … one with leaks.
A ‘goal’ must be metabolized by the system that ‘pursues’ it. I use quotes because I think goals and the pursuing of them are largely man-made concepts. But they still offer a useful paradigm for getting robots and agents to do our dirty work.
Instead of training models only to chase reward (which can be gamed), can we train them to preserve and improve state integration across transitions …
In an integrated setup, the local move carries the whole; it is fractal. The bee’s desire for nectar is already folded into pollination that later brings fruit … and this is what we call ‘a goal metabolized by the greater system’. Did the action transform the current state into a more integrated next state without creating hidden debt?
The forest is, literally, the most natural example. Simply having more trees is not enough if there is not a rich variety of animal, plant, and microbial life. Not for no reason. Nature exudes an aliveness that is increasingly greater than the sum of its parts as its true biodiversity increases.
We use metrics that don’t speak on the integration of the system, metrics like net primary productivity (NPP) and evapotranspiration (ET).
NPP tells you how much plant biomass is being produced, but not whether it’s embedded in a biodiverse/self-renewing system.
ET tells you how much water returns to the atmosphere, but not *how* that water is actually being cycled aka via roots/canopy or lost due to bare soil.
Used in conjunction with other things, however, like the species richness + functional diversity, one can get a fuller picture. Platforms like Restor can maybe pull this off given their rich ecological data.
But I do think it’s about creating an IMAX-like experience to truly feel and see the integration because it’s not about simply stacking metrics on top of metrics even though, with discerning analysis, a decent picture can emerge.
There is no explicit goal, or clear metric to maximize or minimize. This is a fundamentally different paradigm.
Yet we know what to do. We want to establish the conditions under which the system can find peace and freedom in integration … integration that can be arrived at in many ways, but, on any given ‘step’ it must still be present … like it’s a sort of travelling fractal. Never a dull moment.
I’m seeing parallels to the Mathematics of PerillaCove in many places, like Euler.
First, ex is special because its derivative is itself … it is growth through self-propagating change. But Euler’s formula adds i to the mix, and if we take the derivative of this formula, we can see that change is turned, phase by phase, until it returns, and keeps cycling like that. eix = cos x + i sin x, and letting z = eix, so z’ = iz …
So I think this whole thing can be taken as accumulation-free turnover which is insane.
Taking inspiration from The Greedy Algorithm, you are probably not trapped in a local minimum. You’re standing on a saddle you’ve yet to thoroughly explore, and it does have an escape into a world of deeper integration.