In the year since our last post on brain-computer interfaces, we’ve been busy. We’ve had conversations with scientists around the world, getting a better understanding of how to achieve our goals. We have a strategy now, and a path in mind: the key technology to achieve our neurotech moonshot.
Over the next few months, we plan to lay out the core of that strategy. We want to explain why we think our goals are realistic, and the road we expect to walk to get there. In doing this, we aim to extend an invitation: if you like where we’re headed and you’ve got the scientific know-how to make it happen, we’d like to find a role you can play.
First, though, we want to say just what those goals are. At e184, we work to give people a voice in our future. We’re addressing infertility, aiming to give hundreds of millions of families a new chance to influence the world of tomorrow. In our view, brain-computer interfaces are also critical to preserve people’s stake in the future, on an even greater scale.
AI, and our role in it
The conversation around artificial intelligence (AI) often defaults to existential risks - endless debates fuelled by fear of a rapidly changing world. If you have checked Google Trends lately, the spike in searches makes it clear: people are nervous. But while others dwell on downsides, we are focused on the upside. The real question is not just, ‘What if things go wrong?’ but rather, ‘What if we get it right?’
We believe in a future of AI systems that can absorb information and generate actions at 10x or even 100x the speed of the human brain, tackling humanity’s grandest challenges - curing diseases, solving hunger, and reversing climate change. But here is the catch: when AI delivers solutions at superhuman speeds, humanity may hit a critical bottleneck - our capacity to comprehend and implement those solutions.
This is not just about social constraints like regulatory inertia or misplaced fears. No, this is about the fundamental limits of human cognition compared to AI’s potential for exponential growth in speed, intelligence, and adaptability. It's an ironic reality that while we’ve made tremendous progress in advancing artificial neural networks, we’ve barely scratched the surface when it comes to understanding - let alone enhancing - our biological ones.
Imagine an AI medical researcher, tasked with addressing cancer. We would want a human expert to be able to check its reasoning: to observe the assumptions it makes in real-time, to step into its chain of thought and correct for distractions, reminding the AI of the constraints of medical ethics and the goals we have for care. As AI gets faster and faster, this goal becomes more and more distant.
Or on a more everyday level, imagine using an AI to design a home. Today, you would need to write down a description, conveying only part of what you had in mind, then laboriously tweak the prompt or modify the final design to fit your goals, leaving most of the potential of the AI wasted. What if instead you could convey what you wanted directly, sending an image directly from your imagination, then stay in control as the AI fills in the details?
This growing disparity between human and AI capabilities isn’t just a technical challenge - it’s a barrier to adoption,to happiness even. It is a roadblock to fully realising the transformative potential of AI in the most critical areas of society.
That’s why we believe the next frontier is clear: the integration of human and artificial intelligence. To truly thrive in the age of AI, we must explore how to bridge this gap and perhaps even merge human cognition with AI systems. Only then can we unlock the full potential of a world powered by collaborative, benevolent AI - one that works faster and in harmony with human ingenuity and intuition.
The Landscape of Human-AI Collaboration
There are other approaches to bridging a future gap between human and artificial intelligence, with different expectations. Some are in progress now, others are hypotheticals for the future. Some expect us to stay as we are, others expect humanity to become something radically new. Here’s how we envision this landscape of approaches, and our place in it:
While there is immense potential for progress purely on the AI side - through enhanced explainability, alignment, or even intentionally slowing AI down to better collaborate with humans - we believe that it will eventually be necessary to change humanity itself.
With that said, we are not chasing the singularity or the full digitisation of humanity. Those ideas, while captivating, pose significant scientific, technological, and philosophical questions that remain beyond the scope of our mission. Instead, we are laser-focused on achievable breakthroughs - solutions that empower humanity to collaborate with AI on equal footing and unlock new horizons of creativity, innovation, and impact.
We see that work as taking place in stages. First, we want to help humans communicate with AI. Initially, this will mean building on our conscious thought, by making it as natural to send a command to an AI in the cloud as it is to think up a sentence or call a memory to mind. In time, we will go further, and approach the speed of unconscious cognition, tapping more directly into our underlying intent. In the even longer term, we want to understand human cognition better, to build a kind of operating system, one that seamlessly integrates AI capabilities. Achieving this work will be no small feat - it challenges millions of years of evolutionary compromise.
BCI AI - augmenting evolution
Communication bottlenecks
Our nervous system and the interfaces we use to communicate were designed for a specific evolutionary context, where every calorie counted and no extra capabilities could survive. The external information transfer rates they support - whether between humans and humans or between humans and machines - are shockingly low. Depending on the method of communication (reading, speaking, eye tracking, mouse movements), these rates range from about 1 bit per second to no more than 50 bits per second. Compared to typical internet speeds of hundreds of megabits per second, this means that a browser is able to communicate with an AI at a data center a million times faster than a human user can deliver commands to that browser. Bridging even a small portion of this vast gap would be an enormous jump in our capabilities.
On top of that, the context-independent nature of our communication with machines further limits efficiency. While our brains operate with complex semantic relationships that naturally embed context, our communication with computers relies on syntactic commands devoid of semantics. Imagine a jet pilot needing to issue a series of commands (e.g., ‘turn left, then accelerate’) instead of conveying a single, context-rich directive like ‘follow that target’.
To improve the performance of machines and user experience, we need computer systems to be able to incorporate more information beyond what we can give in an explicit command. Such information may help to build and train better AI, combining reinforcement learning from human feedback with richer, more meaningful information from BCIs that enhances AI's responsiveness, adaptability and effectiveness. In essence, in order to align AI, we would like it to not merely do what we say in detail, but do what we intend, overall, a combination of our conscious desires and the unconscious processing that underlies them. As brain-computer interfaces make progress from detecting explicit intended speech to inner speech, we expect there to be a path forward to guide AI with this kind of processing.
Processing bottlenecks
Ultimately, our goal is not merely to enable humans to communicate with AI as we are: we want to use computers to improve our internal ability to process data. Here too, there is clear room for improvement. Peripheral processing by the ‘outer brain’ (the retina or visual cortex, among other areas) operates in parallel, handling vast amounts of sensory information simultaneously. For example, the retina produces a million parallel output signals, which the visual cortex processes further through its hypercolumns, a parallel set of 10,000 modules. However,once sensory information reaches central cognition, the ‘inner brain’, for example, the prefrontal cortex, processes it serially in order to commit to a conscious response. As a result, ‘… when faced with two tasks in competition, individuals consistently encounter a ‘‘psychological refractory period’’ before being able to perform the second task’. This serial processing clocks at 10 bits per second, a limit to how quickly we can consciously assess new ideas.
These limits, in turn, appear to affect our mental representations of what we see. Researchers have argued that ‘… items that are attended to and foveated [i.e. related to the fovea, the small central part of the retina responsible for sharp central vision] are perceived at a higher resolution, while items that [are] unattended or are in the periphery are primarily perceived as being part of an ensemble…’, wrapped into a kind of lower resolution sketch.
Now compare these limits to the streamlined logic of computers - machines comprised of billions of transistors, designed for sequential processing, each operating millions of times faster than a human neuron. Thus, in addition to the constraints of our motor system, which slow our ability to communicate with computers via typing or speech, we are constrained by the processing capabilities of our brains themselves.
Possible futures
If humanity is to thrive alongside AI, we need to solve at least three critical bottlenecks: communication, context, and internal processing. Only when we achieve this, will our brain-machine interface be able to evolve into a brain-machine operating system, enabling augmented cognition that transcends our natural limitations.
Imagine an AI seamlessly handling complex sub-tasks for us, using external processing as an extension of our minds. Even without AI, with a brain-reading device we could store results externally, retrieve them, and solve problems with unparalleled efficiency. AI adds additional text-processing capabilities on top of this. While we process text consciously at around 10 bits per second, GPT outputs around 100 tokens per second, with each token carrying around 10 bits of information. Giving people direct mental access to this kind of processing would be a jump of two orders of magnitude in speed of conscious thought.
We’re building this future - a world where humans and AI are no longer disparate entities but seamlessly integrated partners. We pursue this future because we cannot stomach the likely alternative: a world where humanity is entirely led by a powerful AI, where we lose our agency and simply execute ChatGPT’s orders.
Join us in creating an extraordinary future where constraints on human cognition are lifted and AI becomes a partner, not a distant tool. Together, we can unlock a new era of augmented human-AI cognition.
We are hiring biologists, geneticists, physicists, bioinformaticians, materials scientists, and engineers who want to work on the questions that will define the next chapter in human biology. If that sounds like you, send us an application:
We would like to thank Blake Richards for helpful comments on an earlier version of this piece.