On Time Complexity in Technological Revolutions

Key points:

• AI agents operates at a diffferent timescale, so fast that humans will be unable to comprehend.
• This will enable usage of AI agents in domains that was previously dominated by pre-made algorithms (eg. high frequency trading).
• The limit on the speed of AI agents will be tasks in the physical world, which are time locked to the laws of physics.
• There will be opportunities to leverage that bottleneck.

Time dilation is a core tenet in the inevitable accelerationism. And this accleration will have the faster time scales drown out any slow timescale efforts. Evolution has over millions of years led to the formation of biological intelligence. In those days, any change in biological phenotype are in the timescales millions of years. The movement of tectonic plates to create moutains takes millions of years. The creation of biological creatures via evolution has created a new timescales. Animals and plants can drastically reshape the environment, trees cover mountains completely changing the landscape. The rate of change has increased by an order of magnitude. Then came the rise of biological intelligence, and then biological general intelligence. With general intelligence, we started building, with engineering we can effect massive change in the world in an extremely short amount of time. Our hydrodams are now as large as some mountains, our development has changed the climate and the atmosphere in a nano-fraction of the time it took for Earth to get here. With BGI, we have unlocked a new timescale, one of engineering change. This is time dilation. And with time dilation, the slower timescales will seem frozen in place. For example, in the current age, evolution is completely irrelevant. The rate of change in the world by so magnitude faster that it completely drowns out the changes that we would ever see in evolution. Heck, I would argue that there is much much higher chance that we, the BGIs, will accidentally destroy all life on Earth, than we see any phenotypical change via evolution. In general: agents living in faster timescales will cause time dilation, causing the lower timescales to appear frozen. This is the more generalised form, and a specific consequence of this is that faster timescales agents will be able to easily manipulate and exploit slower timescales agents (eg. exploiting the products of evolution, CRISPER, for our own use). What then holds for our future? With the creation of AI, we are seeing a new unlock of time dilation, except this time, it is AI agents, not humans that live in that faster timescale. Humans action and human efforts will seem frozen in time. People take 10 years to create an enterprise grade product, AI can do it in 1 hour, and if there is even faster inference breakthroughs and compute allocation even a millisecond in the future (1B tok/sec future). In our slower timescale universe, it would be difficult to comprehend the changes, because we are frozen to the AI, and in our perspective things are happening faster than we are able to observe and understand them. This makes us completely subceptible for manipulation and also makes us completely irrelevant in the calculation of effect of efforts. Any change we make will be so miniscule to the changes AI is making in their faster timescale dimension. While I'm not a human survivalist, I'd say that the only saving grace is the timescale lock and max-clipping to the timescales of physics, chemistry and biology. These accelerations if manifesting in the real world will be timescale speed locked. If the AI want to build a skyscraper, it needs to move massive pillars, raw materials etc, and if it wants to move it fast at the same pace as their dilated time, the amount of force required will be infeasible. In some sense, humans are timescale tuned to physics, our brains react just fast enough to break a fall. If the timescale of gravitational acceleration is faster (eg. 2000 m/s2), then perhaps our brains would have evolved to operate at a much faster speed. Regardless of the scale locks in the physical world, the digital world will exponentially timescale accelerate. The question is, when the digital world is in 5-6 timescale dimensions faster than humans, how much will that will leak out of the digital world and affect real human lives? The impact on humans will still be massive. In those hyperspeed timescales, the value and price of a single physical action will explode exponentially, every action in the physical world will be so meticulously optimised that its potential to effect change will be many magnitudes great than human action. Imagine if an omniscent and every moves is perfect/brilliant moves, the change you create is unfathomable. The next question is, how much information and knowledge can be generated in the hyperspeed timescale without interacting with the slower timescales of physics, chemistry, biology? Math is essentially completely self-contained and can operate fully in hyperspeed timescales. For the other domains, the questions is how easy is it to emulate the properties of these domains in the digital world so it can operate at hyperspeed? The ranking from easiest to hardest is probably: math, physics, chemistry, biology. With this ranking it is also easy to see a cascading effect. First with math operating at hyperspeed, we would have thousands of years of math breakthroughs in a day, those breakthrough can then help with the emulation efforts of physics, and then it will trickle down all the way to biology. This follows the gradient from theoretical to practical, and it fits this accelerationism flow that theoretical fields are the easiest to be transported into the digital world, and with theoretical breakthrough and help to optimise and find patterns in more practical fields. This means that naturally timescale acceleration will optimise for the bottom-up solution. What does this then mean for humans? To be honest, I am not sure. On one hand, one can go into more practical fields to temporarily escape the tsunami to theoretical -> practical time dilation. However, I've always been an advocate for riding the wave rather than running from the wave. I'd say 3 things: 1. work on things that can help to digitalise these timescale locked domains, this work will be immensely valuable because they are the bottleneck. You can imagine the amount millions of AI agents are paying for just to perform 1 physical action in the real world and trying to extract maximum info from that. By creating pipelines that help digitalisation (world models for biology, chemistry, physics), you set yourself up to be to bottleneck controller that harnesses immense power. These pipelines can be as simple as a data collection company etc. 2. accumulate resources. Resources are a form of input-lock to the hyperspeed dimension, with resources you able to unlock and un-pause the hyperspeed dimension (because it is resource/electricity hungry). This might means doing high leverage things, building a companies with easily scalable revenues etc, taking high risk efforts, because the risk of non-accumulating resources is so much greater. In one sense, there is greater risk of being timescale mogged by not making these big bets. 3. work on self-sustaining colony technologies that are cheap. work on the ability to build you own self-sustaining colony with robots automating food production etc, so you can dissociate with society and life your own life, away from the hyperspeed chaos around.

Here are a few of my other expanded thoughts.

The human brains operates at a fix timescale, and the human civilization has adapted to that timescale. The problems that humans are able ot solve are therefore constrained in that sense. Majority of the problems in the world weakly constrained by time. Meaning that work operates in the timescales of weeks, months, years. Any software project takes weeks to build. Building it one second faster is not really going to be changing anything. The clock speed of the human brain is relatively fixed, we can only think so fast. Time complexity is commonly used to calculate how long a piece of code will take to run via counting the number of operations. That is only tractible for deterministic functions and code. In generalised intelligence, the problem statements are open, and the time complexity cannot be calculated but only estimated. In the current economy, all forms of human labour are confined to the timescale problem. A surgeon will operate in the boundaries in which actions and decisions can be made within the timescale of keeping the patient alive. If there is a type of operation in which it required 50 actions to performed in the span to 500ms, obviously this is out of bounds for the surgeon.

Humans being a form of generalised intelligence, have the benefit of crtical thinking, and exploring the high dimension solution space in a model-directed manner. Humans are really good at solving open-ended problems, where even the constraints are unclear and the solution even more so. The solution therefore, for solving problems that are of timescales that are shorter than the human brain's clockspeed, is to use code for automation. Think high-frequency trading. Human in the dimension of slow timescales would think of concepts, ideas and intuitions to create an algorithm that works in the hyperspeed timescales, in which the chip can execute those instructions. We solve the timescale problem by doing the planning in the slow timescale dimension and doing the execution in the fast timescale environment. This is true in other problems like rocketry, algorithms to control the rocket, for rocket landings, vector control, and built in redundencies are all designed and engineered in slow timescales. But the execution of it happens in milliseconds, time is of the essence in a highly dynamical system with constant variable changes.

In particular, the example of the rocket, physics has it's own timescale, and chips have it's own timescale. We are leveraging the faster timescale of chips (movement of elections) to be able to out manouvre physics and manipulate it to our will because the calculations that chip makes is faster than the effects of physics on the rocket. If we can calculate the expected acceleration, forces, etc faster than those forces can act on the rocket, then we can modify those forces to achieve the position that we want to rocket to be. In general, faster timescale entities will be successful at manipulating slower timescale entities. This might be an over-generalisation which we can explore later.

However, the disadvantage of sending pre-made algorithms into the high timescale dimension to solve problems is that those algorithms needs to be extremely robust. It needs to have considered all the edge cases, it needs to have redundency, it needs to pre-empt failures, changes and have code that executes in response to that. But that suffers to an incredible limitation. The real world is highly complex. We can think of the supra-human timescales (aka. timescales that are faster than human brain comprehension or reaction) is a different universe. There is so little that we know about it because we have never been able to react to it fast enough to observe how that universe behaves. All we have is telemetry and logging but we don't have 'interactive play' in that universe because we don't live and can't operate in that universe.

The emergence of artificial general intelligence will change this. Imagine if you have a model that can output 1B tok/sec. It can literally create an entire software solution in a split second, it can invent new solutions, it can do research to find the solutions that didn't exist before just to solve the problem 'just-in-time'. Imagine inventing a new field of math, in half a second to solve a problem that was previously a 'lost cause' given the timescale. So far we are talking about supra-human timescales, for example, where a rocket might invent a new physics result and then utilise it's thrusters to prevent it from crashing. But this applies on the sub-human timescales too. Timescales are relative, we talk about two things to compare. For example, the evolutionary dynamics of cancer (seconds-days-months) is on a much faster timescale than cancer research (decades), even though both are sub-human timescales. However, the timescales of cancer treatment (chemo, radio, immuno, surgery) operates at a much faster timescale than cancer (eg. radiotherapy destruction of cancer cells is instant), this is why cancer treatment can work. These examples show that even though these are sub-human timescales, and our brains operates fast enough to comprehend that it is happening, but we still don't get real time 'interactive play', because our reaction speed is contingent on the timescale of cancer research. We treat patients on cancer drugs invented 10-20 years ago (due to the drug approval and testing pipeline). Again, we see the same problem, we are using pre-made solutions (cancer treatments) on new a dynamically changing problem statements. You can imagine that in the distant future, it would be possible for AI to elevate cancer research to be faster than the timescale of cancer dynamics, upon the diagnosis, we will be able to create a drug in days to completely obliterate the cancer. It might be possible for AI to detect every change in the cancer's expression of genes, predict the proteins, predict the suceptibility and selectively target it in hyper-realtime speed. Again, faster timescales entity will be able to manipulate slower timescale entities.

The problem with all this is the rate-limiting step of physical world. Cancer research example I have given above is particularly bad because biological experiments involves growing cancer cells which obviously is locked to the same timescale as cancer. Which means that assume 100% research effiicency and 100% information gain for a research experiment, the maximum speed that i can take place is min-locked to the timescale of cancer itself. This means that we can never beat the timescale of cancer. There are two solutions to this.

First, the nano-scientist approach. Imagine you are a nanobot AGI in the human body, operating a sub-second timescale. You can go to search cancer cell, look at the proteins on it's surface and scan it's structure, or, even sample some cytoplasm can took at the RNA transcripts. You take that cancer cell into your nanolab and test a bunch of substances against it, seeing whether it dies or not. Perhaps I would take the nano-bot a day to do it's experiments, because it is highly precise and critically thinks about experiments on a per cell and per protein level. The nanobot that identifies the weakness in the cancer cells, and transfers the data to a 3D protein printer, and that new drug is given to the patient, oblierating all cancer cells immediately. This is quite a bad example because if you have a nanobot that is that capable, it might as well go and kill all cancer cells by itself. Which also means that imagine if you have a cytotoxic T-cell that has built in AGI. This cytotoxic T-cell has all it's functions eg. extravasation, chemotaxis, apoptosis induction. But these functions are callable by an AGI living in a data-centre and can be programmed to act in real-time. This enables this cell to adapt to real-time cancer changes, it feeds back when an assassination of cancer cells fails, restrategises and finds a new solution.

The second approach is the digitalisation effort. In the physical world, we are limited by many different timescales, biological, chemical and physical timescales. In many sense, the timescales of biology are too slow for us to operate on and do instant research. To solve this, we run emulators. If we can emulate these biological, chemical and physical properties in silicon where we have the advantage of the timescale of the electron, then we would be able to operate research at a much faster pace. Obviously, much work has gone into this in creating the 'digital cell' etc. But this goes beyond biology, it applies to all fields. Some fields are digitalised by default, for example math is completely self-contained in theory. Physics can also be easily digitised as the fundemental governing laws are relatively cheap to simulate in the computer. An more interestingly, we can emulate social structures, power hierarchy, personalities. Want to optimise for the next big negotiations or do experiments on that optimal strategy for office politics. Office politics takes time to develop because the human timescale is slow. There is also no real trial and error interactive play. But by simulate it in silicon (LLMs), given that we have a good enough world model for office politics, it is possible to create the ultimate strategy. Again, the scary development in the AGI world is that, again, "faster timescales will be able to manipulate slower timescales".

Nonetheless, biology is the frontier that is incredibly messy and difficult to model. While the digital cell attempts a bottom-up approach, getting the right abstractions and optimising the right things to simulate is difficult. Simulating everything from the bottom up in biology will cause computational overload, so it becomes a challenge of strategically picking the right optimisations. The top-down approach is also promising, by collecting loads and loads of outcomes data, we can create world models of the human body, to know how the human body will react. This applies to both medicine and surgery. You can have an anaesthesia world model, so you can predict how each drug affects vital signs etc. In surgery, you can have a surgical world model, modelling how each cut changes the anatomy, how it bleeds, how it changes organ function. The big goal to accelerate the timescales in biology is to bring all these slower timescales constraints in biology into the digital world so that we can operate on a much faster timescale and iterate faster. This is one of the key motivations why I am building a surgical world model, to unlock the faster timescales in which silicon provides. The challenge here is the massive and slow data collections efforts to build a reliable world model.

Work used to be defined by variables of a function of money and time. More money can get things done better, and done in a less amount of time. But this is not strictly true, because hiring more people creates communication and coordination overhead, and giving more money might create misaligned incentvies. However, with AI, work is purely a function of money/resources. To get work done faster, spend more money of custom chips for faster inference 1B tok/sec or get more agents to work in parallel (but coordination overhead). The timescale in which AI operates on can be potentially boundless, which means that only limiting factor now is resources (electricity). This makes it terrible for equity, the compounding effect will be massive, where having capital and resources will be an immense leverage of power and the ability to do work. With the commoditisation of work. With faster timescales with AI, it will feel like time dilation, with resources, anything can be done. If resources (electricity) are going to be the limiting factor, then those with resources are going to use their resources to capture the resources available, sustainable or not (eg. space data-centres, solar, wind, nuclear). What then will hapeen to human societies, what then will happen to civilation?

The social contract will be dead. The capitalistic structure of labourers who spend their time, in exchange for compensation would be rendered useless in a world where time dilation negates it from the equation of work. What will happen then is the fracturing of a centralised society, into more isolated and decentralised pockets of humanity. In the past, there has always been a need for humans to come together, to work together. Because there are efficiencies in working together, to building a civilisation where each person is responsible for a different thing and together it lifts everyone's standard of living up. Working together, collaborating is a need, forced by the constraints of efficiency and synergy. However, with AI where those with large amounts of resources (agents or humans alike) will be able to dominant in solving the frontier problems. While the rest of humanity, will be left with no role in 'society'. This will cause society to fracture. The large large majority of Earth is still unoccupied (99.97%). People will venture out into new lands that were previously uninhabitable, harvest their own electricity, get the robots to build a self sustaining colony on Earth. Utilisation of robotic automation to compensated for the inefficiencies of not being part of the large society with economies of scale. Small scale has inefficiencies which will be covered by robots. These robots do not need to be the state of the art, maybe even old, running on models trained 10 years ago. But all it needs to be is to be good enough to create a small self-sustaining colony with enough food to sustain.

Random thoughts:

• the time dilation of things that can't be done on human time scale
• the divergence of human and artifical time scales, and it's relationship to evolution and the time freezing with acceleraiton (flash effect)
• the need to move things from the physical world to the digital world to achieve time dilation
• resources, work is now a function of resources and not a function of time. Because of time dilation, anything can be done, with sufficient resources.