Problems with today's AI

01

Energy: an exascale computer to simulate one brain

Compute the human brain the way artificial networks compute, with 86 billion neurons, 7,000 inputs each, updated a thousand times a second, and you need about 600 million billion operations every second. That is exascale. Five machines in the world reach it, and each draws 25 to 42 megawatts.

20 Wyour brain

×1,000,000 or more

25–42 MWone exascale machine

The ExaBrain answer

At any moment almost all of the brain's neurons are silent, and a silent neuron costs nothing. ExaBrain works the same way: the 185.4 billion neurons online today are served by one Mac Studio, using about a quarter of its processor while answering ~600 queries a second.

TOP500, June 2026: the five exascale systems and their power Demo: what a trillion dense weights cost

02

Every connection, every cycle

In a dense network every weight is read and multiplied at every step, whether it matters or not. "Mixture of experts" models cut the waste by switching on only some blocks for each word, but what is left is still huge: DeepSeek-V3 uses 37 billion of its 671 billion parameters for every token (about 1 in 18), and Kimi K2 32 billion of 1 trillion (about 1 in 30).

At roughly two operations per active parameter, that is still some 70 billion operations for each fragment of a word.

The ExaBrain answer

The work follows the neurons that fire, and only those. A firing neuron reads its own list of connections and nothing else. The cost of an answer depends on what is active, not on how much is stored: a query across 185.4 billion neurons takes a median of 12 ms.

DeepSeek-V3: 671 B parameters, 37 B active per token Kimi K2: 1 trillion parameters, 32 B active Demo: dense and sparse side by side, 64 to 16 million neurons

03

Learning needs astronomical data

The amount of data is astronomical. Meta's Llama 3 (405 billion weights) was trained on 15.6 trillion tokens, more text than thousands of people could read in a lifetime of reading eight hours a day, using 3.8 × 10²⁵ operations and around a million training steps, each one adjusting every weight.

There is not enough text in the world to repeat it, so models make roughly one pass over almost everything available. (AlexNet, in 2012, went through its 1.2 million images about 90 times.) But a fact like "Mozart, composer, born 1756" turns up in thousands of documents, and studies show a network needs to see a fact many times, on the order of hundreds, before it reliably stores it. Facts that appear only a few times are mostly lost. Adding new knowledge afterwards means more training, which can overwrite what was learned before.

The ExaBrain answer

One-shot learning. A single presentation sets all-or-none connections, the way a child learns a new word from one example. New records are added alongside the old ones without disturbing them: the 462 million OpenAlex works were stored in one pass, with no training.

The Llama 3 Herd of Models (Meta, 2024) Large language models struggle to learn long-tail knowledge (Kandpal et al., 2022) Knowledge capacity scaling laws: about 1,000 exposures per fact (Allen-Zhu and Li, 2024) Demo: a dense memory and ExaBrain learning the same patterns

04

Impossible to read

In a trained network, information is smeared across billions of connections: each weight takes part in many facts, and each fact is spread over many weights. There is no place you can point to and say "this is where it knows that". Finding out has become a research field of its own, and its maps are still partial.

So a model cannot show where an answer came from, a fact cannot simply be deleted, and when it invents an answer, nothing inside it flags the invention.

The ExaBrain answer

Every neuron has a name and every connection is a fact you can list. Every answer traces back to the records it came from. To forget something, delete its connections. Nothing is generated, so nothing is made up.

Mapping the mind of a large language model (Anthropic, 2024) Toy models of superposition: why each weight serves many facts

05

Hard to control

Since nobody can read what a network knows or why it acts, its behaviour cannot be written down or checked directly. Developers steer it from the outside, with more training on examples of good and bad answers and with written instructions (prompts) telling it what to avoid, then test and hope the tests cover what matters.

Those guardrails can be talked around: carefully worded requests regularly get models to do what they were trained to refuse.

The ExaBrain answer

What ExaBrain can retrieve is exactly what was stored in it, and both can be inspected. Its contents can be audited, edited and removed record by record, and every step of a search can be followed.

Jailbroken: how does LLM safety training fail? (2023)

06

An existential risk

Put the last two problems together: systems that grow more capable every year, whose knowledge cannot be read and whose behaviour can only be steered indirectly. Many of the people who built them now warn that this could end badly for humanity.

"Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."

Statement on AI Risk, May 2023, signed by Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman, Dario Amodei and hundreds of others.

The ExaBrain answer

Not a cure on its own, but a different starting point: a memory whose every connection can be read, traced and deleted is far easier to keep under human control than billions of weights nobody can interpret.