Everyone Has an AGI Date. Here’s the Math Behind Each One.

S-Curves, Cat Brains, and the Uncomfortable Arithmetic That Separates 2028 Optimists from 2060 Skeptics

Terminator movie. Source: The Verge.

Elon Musk says AGI by 2026. Sam Altman says the end of the decade. Dario Amodei at Anthropic described systems “better than almost all humans at almost everything” by 2026 or 2027. Shane Legg of Google DeepMind gives roughly 50% odds for minimal AGI by 2028. Jensen Huang says 2029. Ray Kurzweil, who first published his prediction in 2005, holds firm at 2029 for AGI and has since moved his broader singularity timeline to around 2032. Yann LeCun thinks AGI is decades away, not years. Geoffrey Hinton says somewhere between 2028 and 2043.

These are not random guesses from random people. These are the individuals building, funding, and directing the most consequential AI systems on Earth. And their estimates span a range of nearly 40 years. So the question is not who is right, because nobody knows.

What reasoning produces each of these dates? Once you understand the math underneath the predictions, you understand what each person actually believes about the future, and what assumptions they are making that they rarely explain in public.

Step One: Quantify the Human Brain

The initial phase of modeling artificial general intelligence involves an inherently uncomfortable undertaking: developing a quantitative measure to compare biological intelligence with that of silicon. It is highly speculative. But it is where every timeline begins.

The human brain contains approximately 86 billion neurons interconnected by roughly 10¹⁴ synapses. Assuming an average firing frequency of 10 to 100 Hz per synapse, the brain’s raw computational capacity is estimated at around 10¹⁶ synaptic operations per second, or SOPS. That number sounds enormous, but it comes with an asterisk the size of a textbook, because neurons and transistors function in different ways.

Modern computers operate on the von Neumann architecture, which maintains a strict separation between processing and memory. Data has to travel between the two. The brain does not work this way. It combines processing and storage at the same location. This architectural difference alone imposes absolute thermodynamic limits on how efficiently silicon can approximate what a brain does.

A brain runs on roughly 20 watts. That is the power consumption of a small lightbulb. The supercomputers attempting to simulate even a fraction of brain-scale operations consume megawatts.

---

---

But the deeper problem is that counting neurons and synapses does not capture what a brain actually does. In 2009, Dharmendra Modha’s team at IBM attempted a cortical simulation at the scale of a cat’s cerebral cortex: approximately 1 billion neurons and 10 trillion synapses. They ran it on Dawn, a Blue Gene/P supercomputer at Lawrence Livermore National Laboratory, using 147,456 processors and approximately 0.5 petaflops of peak computing power. Despite that enormous infrastructure, the simulation ran 100 to 1,000 times slower than real time.

In 2009, simulating a cat-scale cortex in slow motion already required one of the most powerful machines on earth.

That publication drew a sharp response from Henry Markram, director of the Blue Brain Project at EPFL, who publicly argued that a brain is far more complex than a static map of neurons and synapses. The brain is a dynamic, plastic, biochemical system whose connections strengthen, weaken, appear, and disappear in response to activity and experience. Hormones like adrenaline add yet another layer of information processing. Counting the parts is not the same as understanding the machine.

He had a point. And yet, there is a useful observation buried inside the IBM experiment: the 0.5 petaflops that occupied an entire national laboratory in 2009 is now achievable with a single rack of modern GPUs. In fifteen years, what was required of a state-level institution has become a commodity.

The exponential growth curve is real, even if the target it is approaching is poorly defined.

The Conversion Problem

To compare brain capacity (measured in SOPS) with computer performance (measured in FLOPS), you need an equivalence unit. Researchers have studied this conversion using the Hodgkin-Huxley neuron model and Intel and IBM’s neuromorphic computing reports published in Science. The rough consensus, subject to enormous uncertainty: one synaptic operation requires approximately 10⁶ floating point operations to simulate faithfully. Which would put the brain’s equivalent compute at approximately 10²² FLOPS.

This number is almost certainly an underestimate for all the reasons Markram articulated. But it is the number that most AGI timeline projections use as their starting point, and it has one interesting property: it produces results that are not wildly out of line with independent expert surveys.

If you plot the average compute capacity of a typical data center from the invention of the first computers in 1941 through 2026, converting measured FLOPS to equivalent SOPS at each point, and then project the trend forward to 10²², you land at approximately 2056. That date is more robust than it appears. Before the LLM wave of 2023, survey-based estimates from AI researchers placed the median arrival of what we now call AGI at around 2060. Two entirely different methods, one computational and one sociological, converge within the same decade.

This is roughly where thinkers like Yann LeCun sit. And it is not far from Hinton’s outer bound of 2043.