 
author: Scott Rouse, Features Writer
More than 20 years ago, I sat in a lecture hall while my professor excitedly talked about artificial neural networks (ANNs) and their potential to transform computing as we know it. Considering neurons as the basic processing unit, creating an artificial network of them is akin to multiplying the processing power by some factor and replacing the traditional view that each computer has one central processing unit (CPU).
I found the concept of modeling neurons in the human brain appealing but frustratingly inaccessible. First, how? How exactly do we interpret neuron firing, a complex electrochemical event that is still not fully understood, and apply it to computing? And that's before we get into the even more complex issues surrounding neural communication, the coordinated activity of neural networks, and consciousness. How do we make computers “think” like us? At the time, my laptop was an inch thick and could only perform a relatively limited number of tasks. It seemed almost unbelievable to think that these ideas could run parallel to the realization of today's silicon-based chips.
It would take a decade for a Google research paper to change everything. In 2017, a team of researchers at Google published a paper with the modest title: All you need is attentiveness. Few could have predicted that this achievement would mark the beginning of a new era in artificial intelligence. This paper introduced the “Transformer” architecture, a design that allows machines to learn language patterns with unprecedented efficiency and scale. Within a few years, this idea evolved into large-scale language models (LLMs), a type popularized by OpenAI's ChatGPT. This was the basis for a system that could reason, translate, code, and converse with near-human fluency. But this was not the first step.
The cost of developing AI is enormous no matter how you look at it.
1 year ago, Google Exploring the limits of language modeling We showed that scaling artificial neural networks in data, parameters, and computation results in predictable and steady performance improvements. Together, these two insights (scale and architecture) set the stage for the era of generative AI. Today, these models power nearly every frontier of AI research. But their emergence also brought deeper questions. Could such systems, or others inspired by the human brain, lead us to artificial general intelligence (AGI), machines that learn as flexibly as we do and reason across multiple domains?
Currently, there are two different avenues of AGI research. The first is the LLM. Trained on a vast amount of text through self-supervised learning, it offers an incredible range of abilities in several key areas, including reasoning, coding, translation, and creative writing. The implication here, a big leap from lecture hall discussions about ANNs, is that generality can emerge from scale and architecture.
However, the intelligence of LLM remains intangible. They lack grounding in the physical world, lasting memory, and independent goals. And this is one of the central philosophical arguments that precludes the legitimacy of this path of AGI. Our ability to learn is probably based on experience, our ability to perceive the world we live in and actively learn from it. If AGI emerges from this lineage, it may not be due to language models alone, but to a system that combines language fluency with perception, embodiment, and continuous learning.
other way
If LLM represents an abstraction of intelligence, whole brain emulation (WBE) is its reconstruction. The concept was most clearly articulated by Anders Sandberg and Nick Bostrom in their 2008 paper “Whole Brain Emulation, A Roadmap,'' which envisions creating a one-to-one computational model of the human brain. The paper describes WBE as “the logical endpoint of computational neuroscience efforts to accurately model neurons and brain systems.” In theory, this process involves three stages. It scans the brain at nanometer resolution, converts its structure into neural simulations, and runs the models on powerful computers.
If successful, the result will not just be artificial intelligence, but a continuation of humans, perhaps with all memories, preferences, and identities intact. In this sense, WBE aims to instantiate the brain rather than imitate it.
A large-scale European research initiative called the Human Brain Project (HBP), which ran from 2013 to 2023, aimed to improve our understanding of the brain through computational neuroscience. Although AI was not part of the project's original proposal, the early success with deep learning in neural networks undoubtedly contributed to its adoption.
A year before the project began, during what is often referred to as AI's “Big Bang,” the book on deep learning was rewritten with the development of an image recognition neural network called AlexNet. By leveraging large image datasets and the parallel processing power of GPUs, researchers at the University of Toronto were able to train AlexNet to identify objects from images.
In the conclusion of the HBP 10-year evaluation report, the researchers recognized that “while deep learning techniques in artificial neural networks can be developed systematically, they often include elements that do not reflect biological processes. In the final stage, HBP researchers worked to close this gap.” This is the reflection of biological processes and what patternist philosophy is concerned with. It's the idea that things like consciousness and identity are “substrate independent” and held in patterns that can be successfully emulated by computers.
As Sandberg and Bostrom pointed out, “If electrophysiological models are sufficient, full emulation of the human brain should be possible by mid-century.'' Realistic or not, this remains one of the few truly bottom-up approaches to AGI, one that seeks to build not a model of the mind but the mind itself.
It is perhaps no surprise that LLMs have become popular at a time when WBE still seems to be in the realm of science fiction. However appealing the idea of being able to copy yourself may be to self-centered billionaires, it is definitely difficult for investors.
price of intelligence
The cost of developing AI is enormous no matter how you look at it. The Wall Street Journal recently reported that Google will invest $15 billion in India's AI infrastructure over the next five years. The Associated Press suggested that Meta has signed a deal with AI company Scale to invest $14.3 billion to satisfy CEO Mark Zuckerberg's “increasing focus on the abstract concept of 'superintelligence,'” a direct pivot to AGI.
These are big numbers, especially considering that the EU gave HBP co-director Henry Markram just €1 billion to carry out a 10-year mission to build a working model of the human brain. In addition to company announcements, research institute Epoch AI reports that “spending on training large-scale ML (machine learning) models is growing at a rate of 2.4 times per year,” and research by market data platform Pitchbook shows that investment in generative AI in 2024 will reach $56 billion, a 92% increase from the previous year.
For investors, the risk profile of AGI research is aggressive to say the least. The potential ROI depends not only on breakthroughs in model efficiency, but also on entirely new paradigms: memory architectures, neuromorphic chips, and multimodal learning systems that bring context and continuity to AI.
Bridging the two hemispheres
Large-scale language models and whole-brain emulation represent two very different paths toward the same destination: general intelligence. In my opinion, neither seems achievable alone. LLM takes a top-down route, abstracting cognition from patterns of language and behavior and discovering intelligence through scales and statistical structure. In contrast, WBE is bottom-up. It aims to recreate the biological mechanisms by which consciousness arises. People treat intelligence as an emergent property of computation. The other is as a physical process that is copied with complete fidelity.
However, these approaches may eventually converge as advances in neuroscience inform machine learning architectures and synthetic models of inference inspire new ways to decipher the living brain. The quest for AGI may therefore end where both paths meet: the integration of engineered and embodied minds.
Spending on training large-scale machine learning models is growing 2.4x faster per year
When we try to answer what makes the mind work, we find that the pursuit of AGI is essentially a form of introspection. If the patternists are right, and the mind is substrate-independent and a reproducible pattern rather than a biological phenomenon, then if we were to replicate our minds in machines, that replication in silicon would profoundly change the way we view “self.” That said, if the endpoint of humanity is to be digitally inserted into a Tesla bot, forever serving the likes of Elon Musk and pulling out Diet Coke, we might be wise to advise restraint.
Nvidia CEO Jensen Huang believes that “artificial intelligence will be the most transformative technology of the 21st century. It will impact every industry and aspect of our lives.” Of course, as someone who heads the world's largest supplier of AI computing chips, he has a vested interest in making such statements.
Perhaps it's best to temper such optimism and leave the late Stephen Hawking's warning in place. “Successfully creating AI would be the greatest event in human history. Unfortunately, it could be the last unless we learn how to avoid the risks.”
