AI in 2024 is impressive – yet biological intelligence is still unmatched. Why?

“Yes, AI can translate languages, detect complex patterns in medical exams, and engage in human-like conversations. But it still can’t match humans!”

In 2024, the statement above is simultaneously – and spectacularly – both true and false, for reasons that are nuanced. While it is true that Artificial Intelligence (AI) has improved dramatically in recent years, the technology we have built remains fundamentally different from what we find in biology. At the root of it all is AI’s ability to capture and recombine in complex ways the building blocks of natural intelligence – neurons, synapses (connections between neurons), and the learning laws that modify those synapses to enable the execution of complex tasks.

Interestingly, from a theoretical & scientific perspective, AI has “all it needs” to replicate much of what biology can do. While we have not yet discovered the “neuron of the soul” (if you find me, shoot me a message on LinkedIn please…), AI can mimic biological neural processes without any fundamental loss of precision – that we know of today at least!

So, in a sense, it’s only a matter of time (and billions of $$) before AI achieves greater functionality, challenging the opening statement that it can’t match human intelligence. However, we all “feel” that there’s something inherently different between us and today’s AI, although it is a difficult feeling to verbalize to pinpoint exactly why this is the case.

This article is about this very issue. Let’s dig into the crucial differences that separate AI from natural intelligence, exploring the real threshold that exists between artificial and biological brains.

AI: on-demand vs. always-on

The first gap is simple yet profound: in biological systems, there’s no on/off switch. Unlike ChatGPT, which remains inactive until prompted by our eager fingers, living brains are continuously active, constantly adapting, learning, and ready to respond in real-time to anything in their environment. Even during sleep, neurons fire and synapses change. This “always-on” nature is a significant difference and obvious to all of us. Yet, amazingly, it has received little attention. While the reactive model of today’s AI is efficient for specific tasks, it falls short of the continuous, adaptive modus operandi of nature.

When we consider what AI still can’t do, this is one of the main differences:

1. AI is “dead” unless we turn it on. Huge difference, right?

For AI to reach new targets, it must operate not as a tool that’s “activated” but as an active, autonomous entity capable of engaging with its environment at all times.

Continuous learning & adaptation

A corollary to the above difference is how AI and biological systems approach learning and inference. In traditional Deep Learning, these two stages are isolated – first, we invest millions of dollars to train a model, then we “freeze” it to perform inference (essentially making the system dormant until we press “run”). Biological systems, no matter their age, are always learning and do not separate learning from action. Brains adapt continuously, integrating every experience in real-time, regardless of whether the sensory input is coming during a “learning” phase, or “inference” (hint: nobody is telling you when to learn other than in school. After that, you are on your own!). This merging of learning and inference allows biological brains to navigate new and unpredictable environments.

The goal, therefore, is to build AI that doesn’t just “know” but actively learns, adapts moment by moment, and evolves its knowledge. Biological systems can’t anticipate when something worth learning will appear on a given moment, versus when a “boring, all-same-stuff” kid of day allows you to be inattentive (do those days happen to you, ever?).

This became clear to my colleague and me at Boston University’s Neuromorphics Lab in 2010 when we pioneered an “always-on” AI called MoNETA (Modular Neural Exploring Traveling Agent), a whole-brain system built to address this exact gap.

The MoNETA autonomous agent architecture developed with DARPA, NASA, and NSF

Designed to power autonomous agents (leading to work with NASA shortly thereafter), MoNETA’s AI architecture was always active, learning by fusing sensory information, motivation, and navigation to perceive, learn, and adapt in real time. By modeling the core principles of biological intelligence, MoNETA provided a blueprint for AI that learns and adapts continuously.

Moneta needed to overcome the Von Neumann bottleneck

Therefore, a second crucial difference:

2. Biological intelligence has “always-on” learning, with no separation between learning and inference.

Energy efficiency will eventually cut your electricity out…

Another consequence of the “always-on” nature of biological intelligence is energy efficiency. As we’ve seen, biological brains are working 24/7 and learning, yet they evolved to be very efficient, managing vast amounts of data with minimal power. This is a sharp difference with respect to 2024’s AI, which requires energy-intensive hardware to support massive computations for training and operation.

In MoNETA, we used memristive technology (work done in the DARPA SyNAPSE program), showing promise in bridging this divide by addressing one of the most energy-efficient components of brain hardware: synapses. Memristive devices are among our best bets for emulating the synaptic functions of biological brains in hardware. Their unique ability to retain a history of electrical activity enables memristors to mimic synaptic plasticity – the process by which synapses strengthen or weaken over time in response to activity – all within a low-power envelope. We also managed to introduce new learning paradigms to enhance learning efficiency.

So, remember:

3. Biological intelligence is low-power, and any AI aspiring to reach this benchmark will have to “pay the electric bill.”

Time for AI to get real (time)

We have seen how different biological intelligence from today’s version of AI is: real-time, always on, always learning, and able to do that with the power budget coming from … digesting a sandwich per day! We are far from achieving this efficiency and operating envelope.

However, as I pointed out, it is a matter of investment, talent and time, all available and coming. There are no known theoretical barriers for AI to equal and surpass biological intelligence. Also, there is no real need to “surpass” is to be useful! By slowly improving and evolving towards real-time, always-on modality, and moving away from Von Neuman computing architecture and adopting neuromorphic computing principles (from in-memory computing to analog, spiking architectures), we will broaden the AI application domain. Robotics and autonomous agents will lead the way, as applications in this realm face more acutely the dependency of current AI from “user clicks”.

And, with real-time AI agents, AI ethics and law need to simultaneously step-up to ensure that these technologies enhance human welfare and remain aligned with our collective values in a responsible way.