Written by David Young, last updated March 22, 2026
Here are the number of parameters needed in a neural network AI model to handle various tasks.
- Digit/Letter Recognition 1M–100M parameters
- Object Detection 60M–250M parameters
- Audio Transcription 1.5B parameters
- Text-to-Image 12B–20B parameters
- Natural Language 175B–1.8T parameters
A human brain is estimated to have approximately 86 billion neurons, and 150 trillion synaptic connections. It is estimated that this is equivalent to a neural network AI model with 150–700 trillion parameters. That is 2-3 orders of magnitude larger than the biggest AI model we can make today.
Moore’s Law seems to be slowing, so let’s predict a future growth of computer capability doubling every 3 years. That would put AI models potentially rivaling a human brain about 18-21 years from now.
What would that mean? That means that 20 years from now, AI’s may have the intellect to do jobs currently done by someone with an average education, such as a high school diploma, no high school, or with an associates degree. However, it may cost a million dollars to provide that AI with the necessary robot body and computer power.
Another way of looking at it. Ten years ago, deep learning was the neural network algorithm that allowed systems like Siri and Alexa to recognize human speech better than their predecessors. People were predicting that AI would put many people out of work. What actually happened, was that a few hundred people who did speech to text translation were put out of work, and 5000 new jobs related to deep learning were created.
Yet another way of predicting the future of AI. The current generation of AI products depends upon transformers, the current generation of neural networks. A neural network is a highly interconnected set of nodes. Each node holds a very small amount of data, and does a very small amount of math using input data and its stored data, then sends a new output value. The arrangement of the nodes and connections is designed based on what works well in computer code, not to mimic the design of a human brain. Researchers have now simulated an entire fruit fly brain, actually reproducing the arrangement of 140,000 neurons (neural network nodes) and 50,000,000 connections between those neurons/nodes. This digital fruit fly brain seems to reproduce fly-like behavior given digital inputs. To do the same with a human brain would require on the order of a million times more neurons/nodes and a hundred million times as many connections. Applying Moore’s law, we can estimate that it could be another 40-60 years before we can completely simulate a human brain in a similar way.
What does this mean for you? If you plan to be retired twenty years from now, maybe not much. If you are earlier in your life, a college degree may be the best investment you can possibly make.
What if distant future AI models have far more brain capacity than humans? African elephants have 4X as many neurons in their brain as humans do. Elephants certainly have a good memory, but they haven’t won any Nobel Prizes.
If we think like a science fiction author, we can come up with a number of possible future scenarios.
- We create a new race of super-intelligent AIs or robots that don’t need humans.
- Maybe every human uses one or hundreds of AIs via their future smart phone or direct-to-brain connection. At some point AI and humans may merge into one being, call them Borg.
- Maybe there is some plateau that AIs can’t pass and they stay just helpful machines.
- Maybe we engineer a world where no one has to work because AIs do all the work. Some people might spend their life engaging in entertainment. Some might become scientists or artists to do what AIs can’t. Some may explore the universe in faster than light ships. Some may decide they were happier raising their own food, making their on clothes, etc. Some may lose the will to live. Maybe everyone will be a retiree, playing bingo, and going on cruise ships.
