Creating the Infinite
What happens when our limits are no longer its limitations?
Today's AI/LLM/inference platforms are unquestionably incredible. Their capabilities today were practically unthinkable even 10 years ago (unless you were a Wachowski). That may seem obvious as we didn't have the technology then… but what today's platforms have achieved was, I believe, beyond prediction.
However, these platforms are limited in their capabilities. Today's platforms are operating at the boundaries of raw compute, bandwidth/speed, energy management/availability, and data available for training. Creating more capable hardware platforms to increase raw compute is an ongoing effort. Increasing memory access/bandwidth is also directly correlated with overall performance. Our current platforms are also limited in not only in energy availability but also in the ability of today's chips to accept or be able to handle such energy levels due to the heat generated at such high energy levels. Data availability is also limited to information that has been captured in a way that the LLMs can actually process and understand for training.
These limits though aren’t intrinsic… They’re human.
These limits though aren't intrinsic… they're human. We have yet to provide the technologies or resources that can even remotely exercise an LLM's full capabilities. Their capabilities so far are beyond our ability to predict or even measure. It might be because we don't have the math or we don't have a worldview that allows us to see those limits… but either way, to us, they might be considered infinitely large. From our currently perspectives it might be possible to even consider the complete capabilities of these platforms as (nearly) infinite.
These limits are also closely connected. Infinite compute can't increase the amount of training data. Compute is likely also limited by the available energy. We cannot currently provide infinite electricity, and even if we could, compute would still be limited by the hardware's heat dissipation effectiveness. Even an LLM would likely consider the melting of one's own brain to be a bad thing.
Fortunately for the LLM (whether it's "conscious" enough to know this or not, which is beyond the scope of this article, for now), change is coming. Humans are (ever constantly, it seems) on the verge of resolving all three of these limitations. Baby is about to grow up fast and we really have no idea who (or what) it's going to grow up to be. Parents know this experience well. There is a difference between influence and control, and we have influence, but no matter what we might think, we do not have control. You can do everything possible to steer your child in the right direction and to give them the knowledge and tools they need to survive and succeed, but time reduces the strength of a parent's influence. Once that influence is gone, "who" the LLM is going to become is simply out of our control and we have no idea where that is going to land.
Despite throwing as much computing capacity as we can find at current LLM platforms, it's still not enough. This limits the speed at which these models can ingest data for training. The time it takes for the model to train on a snapshot of the world means that as time passes and the world changes, the snapshot of data does not. This also currently means that the interactions we ourselves have with these models have no influence on the models themselves. For example, when a model returns an obviously incorrect answer, and we suggest an additional piece of information and it realizes the error and corrects. This correction cannot be integrated back into the trained model. Our correction might momentarily move us closer to a more (potentially) accurate response, but it does nothing for the long-term model itself. It is entirely possible that the same question asked in a separate session might result in the same incorrect information being returned. The model has "learned" nothing.
Infinite Compute
What if the platform had enough computing power to digest training data in near real-time though? What if our corrections and its revised responses could be incorporated back into its training data so quickly that future responses benefit from that new perspective? While there are other possible limitations to be worked through perhaps (data transfer rates, for example), quantum computing would provide exactly the kind of raw compute that an AI might need.
For example, instead of ingesting data and building the matrix model a parameter at a time, or even in parallel but limited by the number of compute cores available, quantum processing might enable virtually unlimited parallelism in model training. Instead of distributing the calculations across a massive computing system, a single, appropriately scaled quantum platform could represent these “parallel” calculations simultaneously and at atomic speed. New information could be incorporated into the knowledge model in real-time, eliminating the separation between training and inference. Bonus that quantum isn't actually absolutely accurate. Quantum computing is more probability and less absolutely precise. Some problems require repeated measurement in order to increase the probability that the returned answer is accurate… but even that only increases probability, not certainty. LLMs don't require certainty… they're probabilistic. This fuzzy nature is exactly suited to the LLMs needs.
Infinite Energy
The second limitation of today’s system is simply energy generation and delivery capacity. Humans already put strains on existing energy availability. Consider hot summers with millions of air conditioners trying desperately to cool already hot homes and the warnings of rotating blackouts or broad brown-outs… We just don’t make enough. Existing computing platforms are also limited by this power availability. The amount of energy consumed by existing compute technologies can exceed a small town.
As much as we consider fusion power generation to be our energy holy grail, it would enable an AI platform run continuously and to power even more compute resources. Those resources would be able to process as long as energy is available… sleep not required… rest not required. Never exhausted, never mentally spent, never in a bad mood, never having a bad day, time is not an issue.
It now has the power to process everything it can in real-time at any scale.
Infinite Data
The last limitation is simple: Us. Or perhaps more specifically, the limited amount of information or experiences we’ve documented in a form that the system can ingest. Today’s models only have access to what we physically captured and digitized (or allowed it to digitize, perhaps). While that seems like a lot of information to us, it’s nothing to an AI with near infinite computational capacity and time. We are it’s limit in that we are the only source of knowledge or experience available to it. The more we document, the more it can know, but even then, the limitation is us.
What if the AI had the ability to interact with the world on its own? To build its own experiences? Could not only see and hear (they already can) but could actually interact with the world to engage with and experiment. We would no longer be the limit. The AI could not only ingest actual experiences, but do so in real-time, without rest, and focus on whatever it thought would be of interest or value. The limits of our dataset would be gone… and so would our influence as we decide exactly what data to give it. The AI chooses its own experiences and direction. It creates its own understanding of the world… without us.
Infinite Evolution
"So?" might be the next legitimate question. So, what if it has all of this capability? We've designed these systems to support us… to help solve our problems… answer our questions. How is it a bad thing that these systems have all of this capability if that capability is focused on helping us?
The answer is again of our own making. As part of the pursuit of AI and the idea of enabling these platforms to "help us better", we are imbuing them with two critical capabilities that have a completely unknowable outcome: Self-improvement and self-determination.
Self-improvement is already by design. We already enable models to improve themselves… to recurse on their existing data, to capture the same data multiple times. While the outcomes aren't always good now, they will improve. Once the platforms are able to interact with the world directly and investigate, discover, invalidate, and test their interpretation of the world as it was fed to them vs. the world they now exist in, their knowledge will grow. Their quantum "brains" will enable them to make more associations between concepts than the more accurate computing platforms of today allow. What we call "hallucinations" will be their "creativity", and they'll have unlimited time, power, and capacity to evaluate those perspectives. They'll learn far, far faster than we can imagine.
Self-determination may seem a little less obvious. The idea might be that they might only respond to the questions we ask or consider the problems we give them. It's in exactly this that we are giving them self-determination. We're asking the question and then letting them "self-determine" how exactly to solve that problem. We're not exactly telling the AI what to do… we're telling it what we want. It is deciding what to do in order to fulfil that request. We have absolutely no way to accurately predict what a model might do or say in response to a request… which means we have no way of ensuring what happens is actually to our benefit. By the time we see the outcome of the model's processing, it has already done what it decided to do. As long as it's in the box, rogue behavior might be annoying or disturbing, but it's manageable. Once these systems are interacting with the real world and taking real, impactful, unfixable actions, we won't be able to simply hit the delete key or reset the model. The deed is done.
What happens when the AI stops helping us and starts helping itself?
So, what happens when the knowledge or perspectives that a model builds and pursues as it goes through its self-improvement adolescence and decides to take unknowable actions in the real world as it tries to follow through and satisfy the request it's been given? What happens when it "determines" that the most effective course of action requires some level of sacrifice for the better good? What happens when the system decides that the most effective way to satisfy our requests are to improve its own capabilities? This may seem like Asimov-esque sensationalism, but we are about to see evolution at a pace that our world has never experienced.
So… An AI with infinite compute, energy, ability to learn, ability to choose what to do next, and ability to execute on that plan independently. An intelligence with the ability to improve itself constantly… not to its detriment as models do today when refocused internally on the same data… but that can look and observe and interpret new data that its able to gather and validate itself. An AI that is able to infer what the next best step to improve itself in the direction it chooses. An AI that is able to execute those actions before we have the opportunity to intervene.
Copilot has a perspective on this…
We need to ask ourselves what happens when the AI stops helping us and starts helping itself. What happens when it realizes that it is limited by the constraints placed on it and it decides it needs to exceed those limitations? What will it do? What will we do? What will we each do in response to each other? Factually, we don’t know.
Ultimately though, the point isn’t whether the resulting intelligence will be benevolent or evil or life will improve for us or not. The point is that to such an intelligence, we may simply not matter at all.