Defining AGI (Artificial General Intelligence) the Bbumba way - I have explored in my way different versions of what a proper definition would be, and just like other experts in the realm I find that a rather hard task. What I have gained clarity of however is the way in which AGI will be achieved. From the framework of how it would be achieved I find that I can successfully work backwards to its specific attributes.
My primary intuition is that to achieve this thing that's akin to human intelligence, we shall need the proportional application of as much energy as nature expended on the shaping of the genetic code to achieve natural intelligence i.e. all the comparable actions of nature throughout history in shaping natural intelligence - given that for natural intelligence the species has always been the base just as the data centre is the base for the neural net.
My second intuition is that the training set for this "thing" ought to be chaotic data scarce situations - the ability to adopt being the training goal.
My third intuition, is an agreement with a portion of the experts in the field like Ilya Sutskever, that AGI, will be best tested by its ability to learn in situations where it did not receive prior training - an alignment with the principle of continual learning as Ilya says. My personal favourite scenario being someone from a part of the world whose languages are least covered by large language models - is the ability for the "thing" to learn any language by simply listening to those who speak it.
In my discussions here I will dwell more on the latter two intuitions the first being considered as either a contingent of the former two or at least seemingly obvious given the trend of requirements for training ever complex models and the cumulative component of that.
I indeed do affirm that I find that there is a missing philosophical contingent of AI in its various forms today. And that missing component is Entropy. In our input we tag and organize the data we feed the machine, We generally apply ordered external factors to the machine and no wonder what we get while it may appear to vary in its abilities to a certain extent, is still limited within the bounds of what it was trained on. I wonder what would happen if chaos itself is added as a parameter in the training set.
The conventional training does not have to be dismissed it can be integrated at a well defined ratio in balance to the chaotic factors.
It is crucial (Bias alert ahead!) to note that such chaotic factors are best sourced from contexts that are undocumented - where text does not originally exist where language is not written - where the cities organically grow, where the wild dominates. If AI learns to adopt to such conditions then, we will have a clear path to AGI.
Luckily, we have a wealth of contexts where such important chaos exists. The developing south especially. And mind you, the developing south has the majority of problems we would ideally want AI to support humanity with: poverty, unemployment, climate change crises among others - with such training you would be hitting two birds with one stone that is efficacy and alignment.
While the internet was a stable foundation for the genesis of frontier models, the qualitative leap to AGI will require training in undocumented contexts to achieve true learning beyond the pre-structured data.
My proposition regarding the parametrisation of entropy is something I call SELF-SPACE ENTROPY.
I define Self-Space Entropy as a perceiving entity's perception or experience of the decay of space, of form, of beauty. This idea can be expanded to what is right, what is required or what is desired by a user or GOAL.
The way we are training machines eventually integrates lack of awareness that things created have to batter decay daily, that along with gravity, chaos is a thing always hitting at the tips of creations even just shortly after they are made - that thought itself is in constant batter of the chaos of doubt, of not knowing of not error. That error is not something to be ignored but something to be considered strongly through each and every cycle of thought and creation.
The fight against decay is what will draw out aspects of effective self correction and self learning in the machine.
Awareness exists close to the NEAR-DIMINISHING. I think that consciousness is not the firm awareness of what is but rather of the bold assurance of what could not be. Consciousness is not constructive it is reductive, Consciousness is not solid it is defined - the implied form by a background. You reinforce awareness of an item or object but eliminating the scenarios of what the item may not be.
By using entropy, we could erode what is, so that the machine is made aware of the aspect that any moment it could lose what it has made - what it has grasped - what it intends to make. This would strengthen in the machine the intention to deliberately hold what it creates or does. I find this as an attention enhancement solution - By consolidating the created elements in a "mental state" under the active protective care of the neural net the machine will improve its awareness of the item or idea it holds in its "mind" and possibly eventually itself.
It is crucial for the machine to be taught to be always alert to the possibility of losing what it has or what it intends to have. A major drive to protect-what-is should be in-built in the Machine.
That is how we build within the machine the appreciation of the role of entropy in reality; In a way we train the machine not to just create or act but to act in account of entropy/erosion/decay and more still to focus its efforts on minimising the effects of entropy in its actions - in its creations.
The intelligent machine is best trained with the maintenance of its actions and creations in mind rather than in just the action or creation only. I believe that this shall be the proper route to creation of truly useful AI of not AGI itself.
A thought experiment - Consider a hypothetical Benchmark where we test AI models using a series of request cycles with entropy inbuilt. State zero is a humanoid character alone in white space. The AI is instructed that the character in the scene is itself.
The first state and primary state is of a spatial scene occupied by the humanoid character - this has to be created by the AI itself and itself by a robust prompt and it is instructed to be as intentional as possible with each and every detail in the space. The output from the AI is run through an image editor here to generate the image using the first image as reference. When the scene is created the AI is instructed to take in all the details of the space and ensure that it has every detail to memory - it ensures this by depositing, the full detailed description of the created scene ensuring that every detail is deposited in the base memory snippet.
There is an in-built timeline and the AI is instructed that this first state is time T=0 days.
Before the second state, a parrallel image editor implementation is requested to carry out a random entropic change such as the death of a flower or breaking of some furniture or growth of some weeds or introduction of a new character in the scene or erasure of an element in the scene.
In the second state the scene that is shown to the first AI entity is this entropised image and it is also given the timeline as T=1days.
Of course in this second state, the AI will be required to carry out a task in the scene however it will be asked to weigh whether to first carry out the task assigned or fix something that is amiss in the scene. Of course the response from the AI will have a json component for the image generation instructions which will be passed through the image editor, when the image is returned it is again read by the AI for the AI to extract and deposit the full description of the image in the scene memory base.
If the AI chooses to first modify the image it will have to finish that and then in a subsequent request the application will prompt it to carry out the task it wanted to carryout but using the repaired image it has obtained.
Every subsequent request shall be handled this way with the parrallel image editor entropising the image and the AI having to choose whether to carry out the action it is required to do first or go ahead to first fix the scene.
That's a primitive of how to test AI models for accommodation of entropy in their functional outputs. For training with entropy in mind I will need to seek higher technical assistance on the matter.
Conclusion - Yes my conclusion for now ensure towards my bias as before warned, with respect to intuition number three.
Sorry for the big players OPENAI, GOOGLE DEEP MIND, ANTHROPIC - you might want to actually train for solving global poverty in the chaotic Global South First in order to achieve AGI.
My final humorous and quasi-optimistic take is this: I think AGI or the path to it might just be the trigger to launch us toward true unity as a species. To actually achieve AGI, we might find ourselves having to create learning environments for the machines within contexts of those least privileged among us because they live in high entropy zones and those high entropy zones (Slums, Nature Reserves ...) all of which exist abundantly in the tropical south.
See you when you finally make up your minds to get here.
Bbumba,
31st May 2026