The Synergistic Levers of AI Acceleration: Why Superintelligence is Closer Than You Think

The Compounding Frontier of Artificial Intelligence and Why Interconnected Progress Shortens AGI Timelines

It’s truly incredible how many different facets of artificial intelligence are making progress simultaneously. Here I want to explain a few of those facets to give you an idea about what to expect in the future. I’m assuming that your timelines for AGI and super intelligence will shrink after reading this. I don’t think many people understand that these things will be working together, multiplicatively and synergistically creating a runaway feedback system.

Imagine that your brain was growing in size at an exponential rate but not only was it getting bigger, but dozens of your mental attributes were expanding and becoming more refined on a daily basis. I believe that the synergistic compounding of its improvement vectors is one of the most underappreciated aspects of modern AI evolution. These multipliers aren’t just parallel upgrades; they reinforce and accelerate one another in reinforcing loops. 

Everyone knows that hallucinations are decreasing, benchmarks are being saturated, cost is coming down, and speed is going up. However, many people forecast that this will end soon and that artificial intelligence is in an economic bubble that is slowing down. I know where they’re coming from, because I also believe that language models are only an early incarnation of true artificial intelligence. But I also believe that the synergy between many different vectors of improvement is now pushing us towards a much more advanced stage. It truly looks like, well before artificial consciousness is reached, large language models and all of their new bells and whistles will deliver an intelligence explosion and the technological singularity.

Synergy with Robotics. As the simplest example of merging complementary technologies, let’s consider the simultaneous progress in language models and that in robotics. Today, robots routinely run language models making them much more useful. These two, previously separate, technologies have recently combined to create something much more than the sum of their parts. In fact, all major robot manufacturers now use language models as the intelligent engine guiding, not only how the robot speaks to humans, but how it interacts with its world.

Now let’s focus specifically on language models and their various components so we can get a wide perspective on how several different areas of research are all converging to accelerate progress towards artificial general intelligence. Keep in mind that there are thousands of researchers around the globe working on making each of these components more effective.

Pre-Training Scaling. The developers of frontier LLMs are using more energy, more computing resources, and more memory (parameters) to train each new model. They are also providing it with more text to read (data). As each year passes, these companies have been and will be able to increase all of these to generate more performant models with greater emergent capabilities.

Test Time Compute. In the last year we’ve seen language models become much more capable as they have been trained to use the ability to think to themselves. This happens after training, during inference (user interaction), and that’s why it’s called "test time" compute. Today’s models are very quickly learning how to reason more effectively, and it’s clear that increasing reasoning, especially with difficult problems, will continue to yield better results.

Reinforcement Scaling. Language model developers are finding that they can use positive feedback in many different ways to improve model performance on a variety of different tasks. Now they are even using reinforcement training to improve reasoning itself. Moreover, AI’s are moving from human feedback to model-generated feedback loops.

Tool Calling. Language models, generally just produce text, but these text outputs can be used to orchestrate actions within software, on the internet, and by robots. This allows them to outsource uncertainty to deterministic, verifiable systems. For example, often a language model will resort using a calculator to make sure it doesn’t make any mistakes. Today they can invoke Python, web search, or domain specific APIs. Language models are steadily becoming integrated with a wider variety of tools and they’re getting better at selecting tools as well as implementing the output from those tools.

Coding. As language models' mastery of English (and every other language) has improved its coding abilities have improved as well. In 2021 Open AI found that simply continuing to train the GPT 3 model on examples of written code not only gave it the ability to become an elementary programmer, but also improved its abilities with language. Today when you ask a language model a difficult question it will often write computer code to help it work through the problem. It’s abundantly clear that AI’s faculties with language and code complement each other and as both continue to grow, we can expect this complementarity to grow along with them. Aside from code, AI is learning how to interact better with the command line and is being further incorporated into browsers, shells, operating systems and productivity software such as Word, PowerPoint, and Excel.

Context Size. Context size is extremely important for language models because it dictates how many specifics, related to the problem at hand, it can keep in mind. In 2003 a common context window for a large language model was around 8000 tokens. Today you can use 1 million tokens with Google Gemini. That allows you to reason over an entire book and this size is only going to go up from here.

Retrieval Augmented Generation. RAG is another important addition to the LLM arsenal, allowing a model to reference documents or data that it wasn’t trained on, and that exists outside of its context window. Researchers and engineers are pushing improvements across all stages of the RAG pipeline (retrieval, ranking, integration, generation, caching, dynamic strategies).

Memory Ability. Forms of persistent memory are allowing new models to work with stateful memory, allowing the recall of facts and preferences across sessions, and this is a step towards agentic continuity. There are many new methods to compress context into smaller or more manageable forms and memory retrieval is being optimized in countless ways.

Computer Use. Every couple of months we are seeing major advances in computer use software. This goes beyond giving AI access to internal tools and actually allows it to control a mouse and keyboard. Of course, this gives it the ability to perform actions with software and on the internet. It is still somewhat clumsy today, but it’s getting better every week. If you Google the term AI computer use, you can see it in action, and it will be clear to you that this is yet another faculty, improving rapidly, that provides AI more autonomy and control.

Long Time Horizon Work. Every major AI company is now working to improve the ability of large language models to stay on track and accomplish tasks and projects that take hours to complete. The first chat bots such as GPT three would simply spit out a couple of paragraphs. But researchers are teaching chatbots to keep spitting out meaningful text that climbs toward the achievement of a goal. We have seen that chatbots from a couple years ago could only work for a few seconds at a time. Now they’re being designed for autonomous goal pursuit, doing meaningful multi-step work for tens of hours without stopping. 

Specific Task Training. AI researchers are training language models on very specific tasks, including economically meaningful tasks. This is greatly improving their utility in the workplace and increasing their value to companies and organizations.

Agentic Collaboration. Language models are being trained to work next to other other language models in a configuration known as mixture of experts that increase their efficiency and appropriateness. But beyond this, language models are being trained to actively collaborate with each other, critique each other, and supervise each other,  leading to improvements in performance across the board.

Multimodal Integration. Models are learning cross-modal embeddings that let them reason about vision, sound, motion, and language jointly. This allows better 3-D representations, the simulation of physics, and advanced reality modeling. I believe that eventually, combining language with imagery generation will provide AI with an ability to use its mind’s eye to imagine and create. The more modalities that are richly together integrated, the better.

Synthetic Data. AI’s are now generating their own synthetic data, and if it is of high enough quality, that data can be used to train the next generation of models. As you can imagine when the best and newest model is used to generate synthetic data, that synthetic data will be first-class and will go a long way when used to train the next model.

Deep Research. Language models are consistently getting better at their ability to perform online searches to gather detailed information about subjects, and then refine the results to generate professional consultant-quality reports. Again, this is not a feature but a multiplier. This is because it’s clear that the ability to do research can have far reaching consequences for an AI’s ability to do intellectually challenging work.

Hardware Improvements. Aside from Moore‘s law and Huang‘s law there are thousands of different S curves that are working together to make computers faster, more powerful, more cost-efficient, and more energy efficient. There are constant breakthroughs in the development of specialized AI chips for hardware acceleration and increased throughput.

Software Optimization. AI software engineers report that there’s still a lot of low hanging fruit when it comes to finding ways to make language models run faster on traditional computing equipment.

Research Agents. Software developers are turning large language models into research agents that are able to accomplish real scientific work. They come up with hypotheses, find ways to research and test them, and generate results and conclusions that inform real world science.

Algorithmic Innovations. Since the transformer model was developed in 2017, it has changed significantly, but has not been replaced. However, there are thousands of excellent articles and experiments advocating new algorithms to enhance the fundamental AI pipeline.

Recursive Self Improvement. There I’ve already been several examples of AI being used to improve itself. Chipmakers use AI to design chips. Large language model designers use AI to write the majority of their code now. Artificial intelligence has been finding key optimizations to improve processes at various level levels. It’s a matter of time before artificial intelligence is redesigning its entire stack. In the meantime, every year humans are publishing thousands of articles on how to make improvements in AI, both utilitarian and speculative. In the next training run, AI will be trained on all of this text giving it cutting edge knowledge on how to improve itself.

There’s a lot of talk currently about when we can expect artificial general intelligence to arrive. The answer to this question has many ramifications. It will determine when people start to invest their money in this future. It will determine when the world decides to take AI safety seriously. My answer is that the compounding effects of all of these different simultaneously progressing faculties is ushering in advanced intelligence. Even though many people are aware of the fact that current AI can still make brain dead mistakes, very few people are aware of these various interlocked faculties, how they interact, and how they are producing lesser-known but significant new capabilities.

There are many other interacting faculties that are acting as multipliers and these include planning depth, interpretability, causal reasoning, scalability, personalization, alignment and safety, compression methods, parameter optimization, dynamic computation, long-term consistency, coherence across turns, commonsense reasoning, user modeling, meta-learning, sample efficiency, and many others.

There are limits to what you can do with today’s AI that are not going to be overcome without a paradigm shift. Bells and whistles will not get us to machine consciousness. It will be several years still before artificial intelligence research understands and replicates all of the processing advantages that the human brain has. However, before that time, the variables discussed here will have amplified each other past the point of automating scientific discovery, reshaping the entire economy, and making humans obsolete in almost every way. Then AI will create the paradigm shift itself. In other words, we may be missing the secret sauce, but what we have built already will find that sauce's recipe. 

How AI Could be Used to Reconstruct Ancient Brains from Their Endocasts

 

I have been wondering recently about the brains of extinct animals. Unfortunately, the soft tissues of these brains almost always decay completely without any traces of fossilization. However, we do, of course, have the interior of the skull for many extinct animals from dinosaurs, to the first mammals, to our ancient ancestors. In fact, the inside of the skull is commonly preserved in fossils and its geometry can give researchers modest clues about its previous contents.

 

Scientists can use the hollowed-out brain cavity of the skull to make inferences about the brains of long-gone species such as the Tyrannosaurus rex and homo erectus. The shape of this cavity can tell us a lot about the T. rex brain, especially when we compare it to that of reptiles and birds. It tells us that the Tyrannosaurus had one of the largest brains of all the dinosaurs, and that it had large areas devoted to smell and sight. There are numerous informative details that can be gleaned from brain cases by trained paleontologists. But I suspect that many more details could be uncovered using machine learning.

In this essay, we will discuss how information could be squeezed out of brain endocasts. An endocast is a three-dimensional model or mold of the internal space of the cranial cavity of a skull. These casts are used in paleontology and paleoneurology to study the size, shape, and organization of the central nervous system of extinct animals. In this essay we will discuss how we might be able to take an endocast, say from an australopithecine, and make reliable guesses about its brain structure using AI. Of course, the interior of the skull is mostly smooth and thus the endocast does not contain gross brain anatomy, but I’m going to assume that it holds lots of hidden information that is invisible to humans, but that AI could discern.

The type of AI we would use would be a neural network like a 3D transformer model or 3D CNN. Such an AI system would be trained to look at an endocast and then predict what the corresponding brain should look like (generating plausible probabilistic brain anatomy). I believe that we can employ currently assessable (or obtainable) data to train such an AI system to do this. During training, we would have to show the system many matched pairs. For instance, one pair would be your brain and your endocast. Then we could use mine, and then many hundred more. The best way to do this might be to use a CT scan of the inside of the skull as a proxy for the endocast and then we could pair this with the MRI of the same brain. Perhaps 500 to 2,000 CT/MRI pairs would be needed.

We would have to give the AI hundreds of examples of these matched pairs to sufficiently train it to use the bone to predict the shape and form of the brain itself. Fortunately, much of this data has already been collected and exists in medical imaging datasets. Then we would collect similar data from apes and monkeys. Training on both humans and primates is crucial because it lets the system interpolate across evolutionary neighbors rather than just extrapolate. After training on this data, the system would be optimized to accept the endocast of a prehistoric human ancestor and produce a statistical 3D estimate of its brain. Cross-validation would come from leaving one species out of training and testing whether the model can predict its brain anatomy from the endocast alone. If it can reconstruct an orangutan brain without ever seeing one during training, for example, then it has captured a genuinely generalizable mapping.

This AI system would use cross-modal learning to analyze the relationships between two data types (bone geometry vs. brain anatomy). This would use a form of supervised learning where the matched pairs provide the system with a question (endocast) and the corresponding ground truth answer (brain anatomy) to check its predictions against. During training, when the AI system gets things wrong, we would weaken the weights responsible for making the mistake, and when it gets things right, we would strengthen them (backpropagation and gradient descent). Over time this feedback would optimize its ability to predict the brain by learning how the geometric features of the interior of skulls map onto brain anatomy. It would embed both skull interiors and brain anatomy in a shared mathematical space, learning the probabilistic mappings necessary to go from one to the other. It is important to remember that when the AI generates a brain, this wouldn’t represent a single “true” brain but a plausible reconstructions. Doing this many times would result in a range of possibilities that most likely capture many facets of the true brain. It could even do so while quantifying the uncertainty with confidence intervals.

 

 

 

 

To be analyzed by the computer the endocast would have to be turned into a number using morphometric analysis (i.e., 3D grid, point cloud, or mesh). Basically, the geometry of each inner skull would have to be mapped, parameterized and transformed into a long, standardized number (encoded) that can be read into a neural network so that it can be compared with other such numbers. This digitization process would also have to be done to the brains. The brains and endocasts would be trained to live in the same latent space via contrastive alignment.

We could train this system on humans and other primates so that we could ask it to triangulate toward hominins such as Neanderthals, Homo erectus, and Australopithecines. This technique would work for any species where we have a fossil skull interior and living relatives. However, the more ancient the animal, the less fidelity the system will have. Trying to similarly interpolate the brain of a Tyrannosaurus rex based on data from reptiles and birds would be much more difficult. This is because the evolutionary distances involve hundreds of millions of years rather than just a couple million years in the case of our prehistoric ancestors. Nevertheless, using this technique on animals that have been gone for eons could still produce meaningful results.

The AI would use an encoder–decoder architecture, where an encoder would turn the skull into a latent code, and a decoder would turn the code into a predicted brain. This prediction would be a conditional generation or probabilistic model. The trained model would use transfer learning (generalizing knowledge about humans and apes to hominins) to accomplish allometric domain adaptation. Essentially, I am proposing a form of geometric deep learning that uses morphometrics (the quantitative study of shape (volumes, curves, landmarks)) to process information about non-Euclidean data involving meshes and point clouds. This system trys to build a latent space where skull/endocast features and brain features line up creating correlational structure. It would not rely on a single topological feature, but rather integrate hundreds of subtle shape cues simultaneously (multimodal fusion) into nonlinear multivariate mappings. The system would embed both skulls and brains into a shared latent space, a kind of statistical arena where structural correspondences can be learned. Learning the brain/endocast correspondence essentially involves a geometry-to-geometry mapping.

I think a properly constructed machine learning system will be able to create copious information from the endocasts alone. I think it will be enough to make predictions about gross brain anatomy, and not just the brain’s surface (cortical mantle). Especially if thousands of human and primate brains can be compared to their endocasts. Even though they look smooth, endocasts contain multiple overlapping latent signals. Humans have trouble integrating these diffuse features simultaneously, but a neural network can. This will allow it to learn non-obvious mappings (e.g., that a particular vault curvature pattern predicts relative cerebellar size).

The endocast may be most informative for reconstructing the outer cortex that lies a few millimeters underneath it. But what about the deeper brain structures? I believe that by using full brains and detailed endocasts, the endocasts themselves might be able to offer plenty of information about the entire brain, including white matter and subcortex. Endocasts contain subcortical structures indirectly and this means that with enough training data, a network can map endocast geometry not just to the outer cortical sheet, but to whole-brain region proportions. Given sufficient compute and precision, I think it is possible that endocasts could even be used to make detailed predictions about connectivity or even fine-grained microanatomy. This could move paleoneurology from an “interpretive art” into statistical prediction of whole-brain anatomy. Using not just the endocast, but the entire skull could contribute additional informative data. It may even be possible to squeeze information about the brain out of the full skeleton.

 

 

A Horizontal Section of the Skull’s Brain Case from the Top

A Cross (Sagittal) Section of the Detailed Surface of the Skull Showing the Brain Cavity

Let’s talk about the landmarks on the endocast that the AI would have available to it to learn from. As you can see from the pictures there is a lot detail and keep in mind that there is a lot of variation in this detail from person to person. The folds of the cortex come in the form of gyri (ridges) and sulci (grooves) which can be visible in an endocast. But the level of detail depends heavily on the species, the individual, and the preservation quality. Smaller-brained primates, like macaques and australopithecines, tend to have more pronounced gyral and sulcal impressions on their endocasts than larger-brained hominins, including modern humans. In adult humans, the folds are barely visible, particularly on the upper part of the skull. The clarity of brain surface impressions on endocasts varies with age. Endocasts of developing brains in infants, for example, tend to show more detail than those of adults.

Despite limitations, paleoneurologists routinely use endocasts to study brain evolution in extinct species. They have successfully used the imprints of some major, consistently preserved sulci, such as the lunate sulcus, to track key evolutionary changes in hominin brain organization. However, the interpretation of these surface details remains a complex and sometimes subjective task, which is why using AI could be very helpful. Natural fossil endocasts, such as the famous Taung child (Australopithecus africanus), can have remarkably detailed surface features. For artificially or virtually created endocasts, the resolution of the imaging technique (e.g., CT scan) can dramatically influence the observable detail. New, high-resolution imaging and analysis techniques, though, are continuously improving the reliability of these analyses.

Meninges, blood vessels, and cerebrospinal fluid all exist between the brain and the bone above it and so they obscure cortical contact with bone. Nevertheless, consistent signals remain: endocranial volume, vault proportions, asymmetries, olfactory fossae, vascular channels, and gross lobe outlines. These are exactly the types of geometric data that machine learning excels at exploiting. The endocast can also give clues about the relative size and position of regions of the brain such as the frontal, temporal, parietal, and occipital lobes. However, the connection is weaker in the superior frontal and parietal areas.

There are a great number of measurable endocast traits, features, shapes, curves, and metrics, that can be extracted directly from the bony vault. These include:

• Overall endocranial volume (ECV)
• Vault shape and proportions (elongation, globularity)
• Asymmetries (petalias)
• Sulcal and gyral impressions (if present)
• Vascular grooves
• Cranial base angles and fossae
• Foramen magnum orientation
• Olfactory fossa and cribriform plate region
• Canal openings
• Cerebellar impressions
• Brainstem/pons impressions
• Relative lobe expansion (bulging and bossing, flattening and angulation)

What could the results of a system like this do for science? For fossils like the Taung child, AI could sharpen our sense of which sulci impressions are genuine. For Neanderthals, it could provide a statistical measure of parietal expansion. For dinosaurs, it might offer credible intervals for sensory specializations. It is worth mentioning that a machine learning model like the one discussed here could be used on much more recent skulls. It could even play a role in helping to model the brains of deceased humans. Having a recreation of a loved one’s brain could help make their AI counterpart or avatar more authentic. Realistically this kind of thing should probably only be done with permission, but I wouldn’t mind, and in fact, I would like to grant permission to use my skull and brain for machine learning right here and now.

By turning these cavities into data-rich predictors, AI could breathe new life into the study of extinct cognition and allow us to glimpse the hidden architecture of minds long vanished. While the results will always be probabilistic and uncertain, they could bring new rigor to paleoneurology, transforming smooth stone endocasts into testable models of ancient cognition. The smooth surfaces of fossil skulls, long thought to be mute, may hold hidden records that only modern computation can translate. In doing so, we may begin to see not only the brains, but also the minds, from worlds long vanished.