Can We Outsmart AI by Uploading Our Minds? MIT Dropout’s Plan for Digital Humans

Isaak Freeman, a former MIT doctoral candidate, contends that the rapid advance of artificial intelligence makes it impossible for carbon‑based humans to keep pace, and proposes uploading human consciousness onto digital substrates as the only viable path to “exponential” intelligence growth.

He estimates that roughly 50,000 Nvidia H100 GPUs would provide enough compute power to emulate a human brain in under ten years. Using a detailed Hodgkin‑Huxley neuronal model, he calculates a requirement of about 600 exaFLOP/s, 700 GB of memory per GPU, and 24 GB/s inter‑GPU bandwidth—figures that current super‑computing clusters can already approach. A simpler leaky‑integrate‑and‑fire model could lower the compute demand to 2–3 petaFLOP/s, shifting the primary bottleneck to memory capacity and interconnect bandwidth.

The central technical questions he raises are: which neurons to simulate, how to set their parameters, and how to reconstruct the connectivity. Data acquisition emerges as the greatest obstacle; he outlines a pipeline requiring hundreds of next‑generation microscopes operating for years, automated large‑scale tissue preparation and staining, 20× expansion microscopy, multiplexed labeling of over 30 receptors, neurotransmitters and neuropeptides, and X‑ray microscopy to image the entire human brain within a year. Additionally, whole‑brain functional imaging across model organisms is needed to map structure to function.

Freeman’s roadmap, titled “From Worm to Human: Scaling Brain Emulation,” maps the progression from the 302‑neuron C. elegans to the 86 billion‑neuron human brain, citing milestones such as the 140 k‑neuron fruit‑fly connectome, emerging zebrafish connectomes, and recent high‑speed imaging microscopes. He proposes combining expansion microscopy (ExM) with protein barcoding to retain molecular‑level detail while dramatically reducing manual tracing effort, thereby enabling AI‑driven segmentation to achieve higher accuracy.

Beyond static structural maps, dynamic neural activity must be recorded. Optical imaging faces a “glass ceiling” 1–2 mm below the tissue surface in mammals, so Freeman suggests using naturally transparent zebrafish larvae and the tiny C. elegans as surrogate systems where real‑time, single‑neuron‑level functional recordings are already feasible, providing essential data for structure‑to‑function models.

He warns that the remaining challenges are not raw FLOP counts but the “memory wall” and interconnect bandwidth: simulating hundreds of billions of neurons and synapses would require roughly 70 petabytes of memory and extremely high cross‑node communication speeds, far beyond current AI data‑center architectures. To validate a true digital human, he proposes an “embodied Turing test” that places the simulated brain in a virtual body and assesses whether it can perform natural behaviors such as foraging and learning.

Freeman concludes that achieving digital humans will demand a coordinated, large‑scale scientific effort comparable to the Human Genome Project or the Apollo program, likely taking 10–25 years and costing $5–50 billion.