Why Eric Drexler’s critics are right

Alyssa Vance wrote a blog post urging people to ignore one of Eliezer Yudkowsky’s critics (who goes by the handle su3su2u1), entirely based on said critic dismissing Eric Drexler as a crackpot. I think Vance has already gotten a lot of spot-on pushback, both in her blog comments and elsewhere (see e.g. here, here, and here), but coincidentally I’d been thinking about writing a post on Drexler recently, and “ignore people who think Drexler is a crackpot” is a hilariously terrible heuristic.

At best, Drexler is someone whose ideas are highly speculative, based on problematic assumptions, and have yet to yield a fruitful research program; and whose understanding of the relevant science appears superficial. At worst, he’s been called a crackpot by many scientists who know what they’re talking about. I’ll let my readers decide whether the label is appropriate.

Some background: Drexler was originally an engineer who had worked for NASA and gotten a master’s degree in aerospace engineering. He got the idea that you might be able to build little robots and factories at atomic scales, and wrote a book on the subject called Engines of Creation, published in 1986. In 1992, he got a Ph.D. under computer scientist Marvin Minsky. You can read his dissertation online.

Eliezer Yudkowsky has claimed that, “if you say that nanomachines cannot work, you must be inventing new physics.” This is a good example of Yudkowsky spouting off on scientific topics without knowing what he’s talking about.

In reality, you can’t derive chemistry from physics. I mean, we can conjecture that in principle you could, given unlimited computing power, but we don’t have unlimited computing power. Wikipedia’s article on computational chemistry is a good introduction here. (Edit: see note below.)

Wikipedia mentions calculating the properties of molecules with 40 electrons, and I’ve heard of doing quantum models of a few hundred atoms with good approximation techniques. But you’re not going to deduce molecular biology from first principles, as a protein can easily have tens of thousands of atoms.

A lot of tech nerds don’t pay much attention to chemistry because it’s not as pure as math or physics, but I think chemistry is the relevant domain of expertise to apply here–Drexler’s claims about molecular assemblers are basically claims about a radical new way to do chemistry.

(I was a biochem major in college for two years before I realized I didn’t want to be a doctor, and it improved my general understanding of the world in ways just studying physics never would have.)

So physicist Richard A.L. Jones is correct when he says:

First, those building blocks–the cogs and gears made famous in countless simulations supporting the case for the singularity–have some questionable chemical properties. They are essentially molecular clusters with odd and special shapes, but it’s far from clear that they represent stable arrangements of atoms that won’t rearrange themselves spontaneously. These crystal lattices were designed using molecular modeling software, which works on the principle that if valences are satisfied and bonds aren’t too distorted from their normal values, then the structures formed will be chemically stable. But this is a problematic assumption.

A regular crystal lattice is a 3-D arrangement of atoms or molecules with well-defined angles between the bonds that hold them together. To build a crystal lattice in a nonnatural shape–say, with a curved surface rather than with the flat faces characteristic of crystals–the natural distances and angles between atoms need to be distorted, severely straining those bonds. Modeling software might tell you that the bonds will hold. However, life has a way of confounding computer models. For example, if you try to make very small, spherical diamond crystals, a layer or two of carbon atoms at the surface will spontaneously rearrange themselves into a new form–not of diamond, but of graphite.

Similarly, Nobel-prize winning chemist Richard Smalley’s criticisms of Drexler are right on target. Vance dismisses Smalley’s criticisms as “centered around a silly analogy comparing molecular chemistry to romance,” but what’s actually going on is Smalley explaining basic chemical principles in a way laypeople can understand. As Smalley says in one of his replies to Drexler:

You cannot make precise chemistry occur as desired between two molecular objects with simple mechanical motion along a few degrees of freedom in the assembler-fixed frame of reference. Chemistry, like love, is more subtle than that. You need to guide the reactants down a particular reaction coordinate, and this coordinate treads through a many-dimensional hyperspace.

Describing this motion through hyperspace as a waltz, as Smalley does, is as good an analogy as any.

To give another not-literally-true, but still illuminating analogy, imagine chemical bonds as springs rather than the sticks you often get in molecular modeling kits. The hyperspace Smalley mentions is the many mathematical degrees of freedom a contraption built from these springs is going to have. Each spring can not only be stretched and bent, they can bounce about chaotically as you’re trying to make your chemical reaction happen.

That’s why Smalley says that in order to have the kind of control of chemical reactions that Drexler envisions, you can’t just have one nanobot arm for each molecule. You’re going to need to control the movement of many atoms all at once. Hence the “fat fingers” problem. Drexler’s reply to Smalley on this point suggests he doesn’t understand chemistry well enough to understand what Smalley is saying.

Smalley, by the way, got his Nobel for discovering so-called “buckyballs”, tiny spheres made out of exactly 60 carbon atoms arranged in a soccer-ball pattern. But these molecules are synthesized by vaporizing graphite, not with nanobots. Today, serious scientific work in “nanotechnology” generally has a lot more in common with Smalley’s work than Drexler’s. That’s why I say Drexler’s ideas haven’t yielded a fruitful research program.

As for whether Drexler is a crackpot, well, different people have different views. From a Wired article on Drexler:

“It’s very impressive, there is no question,” said MIT chemist Rick Danheiser, who served as Drexler’s thesis adviser, in 1992. “I couldn’t have done a better job.”

“It showed utter contempt for chemistry,” countered Danheiser’s colleague Julius Rebek. “And the mechanosynthesis stuff I saw in that thesis might as well have been written by somebody on controlled substances.”

The first paragraph contains a mistake, Danheiser was on Drexler’s committee but Minsky was Drexler’s adviser. Danheiser was the only chemist on the committee, so my guess is that Drexler did not get an especially well-rounded chemistry education when getting his doctorate. I’m speculating, but it’s possible Drexler did impressive work with molecular modeling software, yet his ideas about the significance of his models look, well, to a chemist at least, like something you’d come up with while on controlled substances.

If this sounds implausible, imagine a mathematician who was firmly convinced that P=NP. He did his dissertation on all the astonishing implications this would have! And let’s assume the thesis is technically correct, impressive even. But when he starts saying that the only thing holding back all kinds of technological miracles is the misrepresentations of his ideas being spread by malicious computer scientists… a Ph.D. would be no proof of sanity. Nor would a lack of peer-reviewed journal articles refuting his ideas.

If Vance wants technical publications, chapter 5 of this report essentially says, in very formal language, that we can’t know that Drexler’s theoretical calculations have anything to do with reality. If she’s not satisfied by that, unfortunately the truth is there’s just not much more to say about the issue.

Note: bartlebyshop tells me the Wikipedia article is bad. Sorry, I was looking for a source that vaguely backed up things I’d learned in undergrad. But if anything, the above discussion may actually understate how hard computational chemistry is.