Re: Paths to Uploading (was: RE: clones/'perfect'-mates/self-image)

Eugene Leitl (
Sat, 2 Jan 1999 15:22:37 +0100 (CET)

On Fri, 1 Jan 1999, Billy Brown wrote:

> You guys have some interesting thoughts about uploading, but I think we're
> still talking past each other a bit when we get down to actual scenarios. I

I think this is only natural. At the end of the millennium, we have surely a lot more hard data at our disposal than Lem & Co-visionaries in 1960's, but nevertheless this is still highly speculative terrain. If at all, uploading is thought to become feasible in a few decades, where future histories are already very warped by the curvature of the nearby prediction horizont.

> [premises of some assumptions self-contradictory?]
> To do an upload requires advanced computers, advanced sensors, and knowledge
> about how the brain works. To get a reasonably upload scenario you have to
> project advances in all three of these fields at the same time, and see what
> you come up with.

Our ability to model dynamics of (macro)molecular systems at low to medium energies far outstrips our capabilities to create, and, especially, to mass-produce them. Due to the bootstrap bottleneck and the intrinsic simplicity of optimal hardware as artefacted by basic physical laws constraints the design space could be already very well sampled prior to the advent of the very first assembler. If analogies to compiler bootstrap are valid, the second-generation assembler (in extreme case, only few fabrication hours away) will be already very useful, and due to the immediate availability of very formidable computational resources as product of the second-generation assembler, the almost immediately (years to months) avaliable third-generation systems should be truly optimal. Of course heavy regulation (nanotechnology simulation software and assemblers to be declared munitions, with a simultaneous implementation of an executive strong enough to enforce that) could delay that, which may or may not result in a prognosis from beneficial to the catastrophic.

I doubt that new sensorics is at all necessary for a feasible destructive scan: recently available methods as cryo AFM of freeze-fracture vitrified tissue cryosections already allows imaging at near-molecular or molecular resolution, and in principle, the technique should be scalable to imaging in the bulk through introduction of abrasion, automation and massive parallelism to attain adequate processivity. I think problems like creation and tight integration of sufficiently dense and fast memories for interim voxelset storage and algorithms for the processing of the latter into the scanning pipeline are significantly more complicated.

The current state of the art of computational neurobiology is not exactly negligeable, and due to the advent of sufficiently large computer performance fully bottom-up automatical knowledge extraction should become feasible, using 'ab initio' methods utilizing total genome sequence data, accurate structures from protein folding prediction and abovementioned molecular-resolution maps of vitrifed animal cells. The same applies to top-down approaches with multiple-million channels microelectrode arrays for in vivo recording and manipulation, and sufficient computational performance for their analysis as made possible with the advent of molecular manufacturing of any flavour.

> Now, the traditional proof-of-principle for uploading is obviously never
> going to actually be used. It assumes no knowledge at all about how the

Oh, perhaps the Cyberworm gang will eventually produce a killer demo good enough to warrant further funding, in a really focused program. If it wasn't for difficulties to patch-clamp the tiny critters, C. elegans is the prime candidate for a POP.

> brain works, which results in enormous computation requirements. Unless you
> think the Omega hardware will be built tomorrow, and everyone in the biotech

For the reasons I mentioned above, I do indeed think that the Omega hardware will become avalable relatively early, i.e. in a few decades, if things will indeed pan out as expected (but nobody never expects the Spanish Inquisition, of course).

> industry is about to jump off a cliff, that doesn't make sense.

The whole of the humanity could jump off the cliff in a hard-edge Singularity if somebody is foolish enough to create the boundary conditions for a SI before we can do uploads on a broad scale. You need a lot less ops and knowledge to grow an alife Golem with excellent juggernaut potential. (Yes, Dr. Scott, an accident has made it happen).

> A simulation at the cellular level, relying purely on advanced knowledge of
> biochemistry, lets you reduce the computational burden by several orders of
> magnitude. It still isn't very likely, however, because it matches a modest

I agree with you that such a model is extremely valuable, especially for bootstrap purposes, as the lowest, or second-lowest tier in an automagically progressive learning hierarchic simulation. (I wouldn't want to define a framework for such a tour de force in software design though, perhaps no one who goes on two legs could).

> increase in medical knowledge with a fantastic improvement in computers and
> sensor technology.

For about 10-20 k$, using off-shelf commodity components, you could now build a conventional MD system capable of probing the dynamics of biological system roughly one million atoms large in a time window few ns long. Even if lacking breakthrough to reduce the computational task for PFP, forcefields will grow a little faster and great deal more accurate in the coming decade or two, while Moore's law should not yet run into saturation yet, particularly if development of molecular circuitry will start early enough to become smoothly available when progress in semiconductor photolitho will fail suddenly, having run out of steam.

The problem with imaging is less a sensor problem than a problem of scaling existing technologies (nanorobotics/abrasive AFM, SNOM, vacuum sublimation, excimer and plasma etch) into micro (MEMS) and meso (diamondoid systems) domain, using massive parallelism and automation. In a sense, the task is much tougher than simply coming up with new sensors.

> A much more probable scenario would project medical advances forward until
> there is hardware fast enough to run the sim, and sensors good enough to
> gather the data. That implies at least a moderately good understanding of

The impetus for new imaging comes from basic research, some of which is of course medical. Medical funding could increase dramatically once the potential of nanomedicine is fully understood by the mainstream. Hardware good and fast enough will come from the mainstream, probably specifically from the multimedia demands, and then perhaps consumer and service robotics, the next big thing after industry automation.

> the brain - something better than just an understanding of biochemistry, but
> probably not good enough to just model the brain's data processing.

I don't quite follow you here. If you can model the neural tissue in machina, all you need is too watch the movies and to abstract. The process of abstraction can be made automatical. What is the problem, then?