Ballpark Base unit Counts 

Brains: 

The base processing unit of a brain is a neuron. Neurons create a massive, interconnected network that functions as a single cohesive unit. 

• Number of cells in the human body: ~ 30 trillion 

• Number of Neurons in the human body: ~ 128 billion neurons in a mature human brain 

• Number of synapses: ~ 100-500 trillion neuron to neuron connections 

Bots: 

The base processing unit in ML varies based on algorithms. One way to size algorithms is to look at the number of parameters that the system encodes.  In machine learning, a 'parameter' refers to a numerical value that defines the model. Parameters can be thought of as the equivalent of synaptic connections in the brain. 

Example of system size: 

• AlexNet ~2012 62.3M params 

• BERT-L ~ 2018 340M params 

• GPT-2 ~ 2019 1.5B params 

• GPT-3 ~ 2020 175B params 

• GPT-4 ~ 2023 1.8T params  

Note: 

The brain is still better connected than our biggest LLM today by a factor of 50-200 times. 

Although models are becoming bigger, a paper from DeepMind, the “Chinchilla Paper”, debunked the concept that "bigger is better" and suggested that many models are over parameterized and undertrained. 

A higher number of parameters allows the system to encode more information, but model learning relies on the input data size, complexity and modality. Bigger models trained on less data are not optimal. 

Training Time

Brains: 

Modern humans have been on Earth for approximately 200,000 to 300,000 years, during which our brains have evolved and trained. Plus, any brain instance is fine-tuned for the length of the individual's life. We come into the world with a brain template that has been tinkered with for ~300,000 years and we teach it and train it as long as we live. 

Bots: 

Machine Learning is a fairly new field (~1940s) that benefits from global interest, collective intellect and large funds. It relies on 3 main aspects: computer power, data availability and mathematical algorithms all of which are under intense development.

Note: 

Human brains have a 300-thousand-year head start in this race but machine learning benefits from that head start as it is a product of our human intellect. 

Speed

Brains: 

• Mechanism: Brains rely on the diffusion of chemicals through Neurons 

• Speed: estimated at 40-200 Hz (operations per second) 

Bots: 

• Mechanism: Bots rely on electrons speeding through transistors 

• Speed: The clock rate for CPUs is approaching 10 GHz => 10 billion operations per second 

Note: 

Computers are a billion times faster than natural intelligence because electrons running through a transistor beat chemicals diffusing in liquid. However, unlike the electrical networks in the computer that propagate the same signal (electrons), neurons use different chemicals that result in different signals. 

Input Modalities 

Brains: 

Brains take input from the physical world through the senses. Different animals have different heightened senses. 

Humans learn most from vision. Vision is the largest human portal in the world. 20 to 50% of the human brain processes vision, depending on how you measure. Dogs and other mammals live in a reality where perception of smell dominates. Bats navigate by sending ultrasound waves and perceiving them as they bounce back. 

The human input experience is rich but limited. Human ears are deaf to ultrasounds. Human eyes only capture a narrow spectrum of light, and we only taste 5 distinct flavors. Yet, the brain has learnt how to use a combination of these senses in very smart ways. 

Bots: 

Bots get their input through sensors that can capture different information (passing through data processing). Bots have no limitation in getting different input modalities, however, there is a limitation today on using all these inputs effectively. Most of our advancements have been in natural language processing, other modalities like image and audio have been slower.  

Note: 

Bots do not have the sensory limitation in input that brains have as they can be augmented with more and more sensory systems. Brains, however, are smarter today in how that input is co-processed for a vivid and holistic experience. 

Plasticity 

Brains: 

Human brains are plastic and in constant change. They respond to the environment and rewire based on input. This is what allows us to learn as long as we live. Brain plasticity is highest in children as they learn to control their extremities, walk, talk, learn language and integrate into social reality.  

Bots: 

Today, machine learning models are static. They update weights (equivalent to synaptic strength) as they train but they do not change the model architecture dynamically. 

Note: 

Brain plasticity gives the brain an advantage over Machine learning models that focus on parameter updates in a static architecture. The brain's ability to remodel itself in response to environmental changes has arguably allowed humans to withstand the last 300K years and flourish in many environments. 

Modularity 

Brains: 

The human brain is modular. Various parts of the brain train to deal with specific senses: vision is processed by the occipital lobe, reasoning by the frontal cortex... Seeing an apple, and recognizing an apple by touch, or by smell, reading the word "apple" and writing the word "apple" are all tasks processed by different systems in the brain. However, the brain works so cohesively that we don't notice the underlying systems until something goes wrong. 

Bots: 

Machine learning bots have been single-model systems so far, with limited orchestration among a few models.

Note: 

To approach human capabilities a machine learning system would likely have to be a collection of optimized systems that work together cohesively with an element of reasoning or preplanning. This sort of idea has started to surface. Example: Dr. Yann Lecun's paper on autonomous ML (see References). 

Intelligence Depth & Breadth

Brains: 

Humans have both broad and deep intelligence. We have a wide understanding of the world and specialization in narrow and deep tasks. 

Bots: 

Until recently, Machine learning has focused on depth, optimizing specialized tasks per model. More recently with Large Language models, we are seeing an expansion of this single task optimization to multi-task models giving rise to systems that seem to have more general knowledge. 

Note: 

While machine learning systems are just starting to tackle breadth of knowledge and tasks, very few systems today have both the capability for depth AND breadth exhibited by the human brain. 

Reasoning

Brains: 

Perhaps one of the most "human" capabilities of the brain is to reason and to project us into the future beyond the physical limitations of our body and the constraints of time. Brains of higher primates have a prefrontal cortex that is responsible for decision making, reasoning and complex cognitive behaviors. 

Bots: 

Most machine learning systems today have little notion of "reasoning". Many are still greedy algorithms that optimize tasks with less preplanning. 

Note: 

Adding "reasoning" and pre-planning to bots will undoubtedly usher in a new world of capabilities and potential. 

Free AI/ML Acceleration Workshop

Interested in exploring how Artificial Intelligence (AI) and Machine Learning (ML) can revolutionize your business? Our free AI/ML Acceleration Workshop is designed to help you:

• Identify your unique business goals and how AI/ML can address them.
• Brainstorm tailored solutions with our AI/ML experts.
• Understand the benefits, use cases, and potential costs of implementing AI/ML in your organization.

Don't miss this opportunity to unlock the power of data-driven innovation! Sign up here!