I have a new paper out where I describe the technological landscape leading to AGI and implications for AI safety (largely building on Bostrom's work in Superintelligence).

Here's the abstract:

In this paper, we contrast three major pathways to human level AI, also known as artificial general intelligence (AGI), and we investigate how safety considerations compare between the three. The first pathway is de novo AGI (dnAGI), AGI built from the ground up. The second is Neuromorphic AGI (NAGI), AGI based loosely on the principles of the human brain. And third is Whole Brain Emulation (WBE), AGI built by emulating a particular human brain, in silico. Bostrom has previously argued that NAGI is the least safe form of the three. NAGI would be messier than dnAGI and therefore harder to align to arbitrary values. Additionally, NAGI would not intrinsically possess safeguards found in the human brain such as compassion while WBE would. In this paper, we argue that getting WBE first would be preferable to getting dnAGI first. While the introduction of WBE would likely be followed by a later transition to the less-constrained and therefore more-powerful dnAGI, the creation of dnAGI would likely be less dangerous if accomplished by WBEs than if done simply by biological humans, for a variety of reasons. One major reason is that the higher intelligence and quicker speed of thinking in the WBEs compared to biological humans could increase the chances of traversing the path through dnAGI safely. We additionally investigate the major technological trends leading to these three types of AGI, and we find these trends to be: traditional AI research, computational hardware, nanotechnology research, nanoscale neural probes, and neuroscience. In particular, we find that WBE is unlikely to be achieved without nanoscale neural probes, since much of the information processing in the brain occurs on the subcellular level (i.e., the nanoscale). For this reason, we argue that nanoscale neural probes could improve safety by favoring WBE over NAGI.  


If you want to read the full paper, here's a link:




Sorted by Click to highlight new comments since: Today at 7:07 PM

Also here's a 5 minute talk I gave at EA Global London on the same topic: https://www.youtube.com/watch?v=jgSxmA7AiBo&index=30&list=PLwp9xeoX5p8POB5XyiHrLbHxx0n6PQIBf

Hi Daniel,

you argue in section 3.3 of your paper that nanoprobes are likely to be the only viable route to WBE, because of the difficulty in capturing all of the relevant information in a brain if an approach such as destructive scanning is used.

You don't however seem to discuss the alternative path of neuroprosthesis-driven uploading:

we propose to connect to the human brain an exocortex, a prosthetic extension of the biological brain which would integrate with the mind as seamlessly as parts of the biological brain integrate with each other. [...] we make three assumptions which will be further fleshed out in the following sections:

There seems to be a relatively unified cortical algorithm which is capable of processing different types of information. Most, if not all, of the information processing in the brain of any given individual is carried out using variations of this basic algorithm. Therefore we do not need to study hundreds of different types of cortical algorithms before we can create the first version of an exocortex.
We already have a fairly good understanding on how the cerebral cortex processes information and gives rise to the attentional processes underlying consciousness. We have a good reason to believe that an exocortex would be compatible with the existing cortex and would integrate with the mind.
The cortical algorithm has an inbuilt ability to transfer information between cortical areas. Connecting the brain with an exocortex would therefore allow the exocortex to gradually take over or at least become an interface for other exocortices.

In addition to allowing for mind coalescence, the exocortex could also provide a route for uploading human minds. It has been suggested that an upload can be created by copying the brain layer-by-layer [Moravec, 1988] or by cutting a brain into small slices and scanning them [Sandberg & Bostrom, 2008]. However, given our current technological status and understanding of the brain, we suggest that the exocortex might be a likely intermediate step. As an exocortex-equipped brain aged, degenerated and eventually died, an exocortex could take over its functions, until finally the original person existed purely in the exocortex and could be copied or moved to a different substrate.

This seems to avoid the objection of it being too hard to scan the brain in all detail. If we can replicate the high-level functioning of the cortical algorithm, then we can do so in a way which doesn't need to be biologically realistic, but which will still allow us to implement the brain's essential functions in a neural prosthesis (here's some prior work that also replicates some aspect of brain's functioning and re-implements it in a neuroprosthesis, without needing to capture all of the biological details). And if the cortical algorithm can be replicated in a way that allows the person's brain to gradually transfer over functions and memories as the biological brain accumulates damage, the same way that function in the biological brain gets reorganized and can remain intact even as it slowly accumulates massive damage61127-1), then that should allow the entirety of the person's cortical function to transfer over to the neuroprosthesis. (of course, there are still the non-cortical parts of the brain that need to be uploaded as well)

A large challenge here is in getting the required amount of neural connections between the exocortex and the biological brain; but we are already getting relatively close, taking into account that the corpus callosum that connects the two hemispheres "only" has on the order of 100 million connections:

Earlier this year, the US Defense Advanced Research Projects Agency (DARPA) launched a project called Neural Engineering System Design. It aims to win approval from the US Food and Drug Administration within 4 years for a wireless human brain device that can monitor brain activity using 1 million electrodes simultaneously and selectively stimulate up to 100,000 neurons. (source)

Neuroprosthesis-driven uploading seems vastly harder for several reasons:

• you'd still need to understand in great detail how the brain processes information (if you don't, you'll be left with an upload that, while perhaps intelligent, would not act like how the person acted, and perhaps even drastically so that it might be better to imagine it as a form of NAGI than as WBE)

• integrating the exocortex with the brain would likely still require nanotechnology able to interface with the brain

• ethical/ regulatory hurdles here seem immense

I'd actually expect that in order to understand the brain enough for neuroprosthesis-driven uploading, we'd still likely need to run experiments with nanoprobes (for the same arguments as in the paper: lots of the information processing happens on the sub-cellular level - this doesn't mean that we have to replicate this information processing in a biologically realistic manner, but we likely will need to at least understand how the information is processed)


What do you mean by nanoscale neural probes? What are the questions that these probes would answer?

Broadly speaking, nanoparticles (or nanorobots, depending on how complicated they are) that scan the brain from the inside, in vivo. The sort of capabilities I'm imagining is the ability to monitor every neuron in large neural circuits simultaneously, each for many different chemical signals (such as certain neurotransmitters). Of course, since this technology doesn't exist yet, the specifics are necessarily uncertain - these probes might include CMOS circuitry, they might be based on DNA origami, or they might be unlike any technology that currently exists. Such probes would allow for building much more accurate maps of brain activity.


Technology to do something like this is already being developed, but it's not nanotechnology: https://www.nature.com/articles/nmeth.3151

Nanotechnology is rarely the most practical way to probe very small things. People have been able to infer molecular structures since the 19th century. Modern molecular biology/biochemistry makes use of electron microscope, fluorescent microscopy, and sequencing-based assays, among other techniques.

Nanotechnology is technology that has parts operating in the range of between 1 nm and 100 nm, so actually this technology is nanotechnology - as is much of the rest of biotechnology.

You're right that the usefulness of non-biotech based nanotechnology (what people typically think of as nanotechnology) hasn't been used much - that's largely due to it being a nascent area. I expect that to change over the coming decades as the technology improves. It might not, though, as biotech based nanotechnology might stay in the lead.