Veritas Curat

It’s always entertaining when I see computer scientists and engineers make analogies between biology and human artifacts. Then they make predictions about biomedical progress as if it marches at the same ineluctable pace as computer science.

Is the brain like a computer? Is the genome like a program? Let’s examine that. Let’s assume that evolution is a programmer.

First, our programmer has no clue what kind of program she wants to write. Unlike real programmers, who have some goal for their code, evolution just knows that she must write code. It is her nature. Second, she doesn’t know how to program. She’s blind, has no foresight, and doesn’t understand the syntax or logic of her programming language.

Luckily that doesn’t matter much, because she has several saving graces. For one, we gave her a programming language that is extremely forgiving of syntax, logic and execution. A missing parenthesis here or erroneous indentation there doesn’t crash the system. Neither does misspelling the names of variables, most of the time (the analogy here is of the vast neutral fitness zone, where most mutations are basically neutral, and only rarely are they significantly deleterious). It’s like a programming language where every statement is an implicit try: with a fallback of except: pass, but not really, since the point is that the language doesn’t care about trivial errors in syntax. It exhibits smooth fitness gradients.

The second saving grace is that our programmer gets immediate user feedback. They say the reason why FOSS works, despite the lack of revenue, is because of immediate user testing and feedback (see: The Cathedral and the Bazaar by Eric Raymond). Open source is open, and rather than hammering out code for months or years in a closed environment, FOSS developers get constant, immediate feedback from a large testing community, which improves the code faster. Well, that’s evolution. The genome is open source, and there is no production, no final release, only testing.

Our programmer doesn’t know how to code. She mostly types random stuff, mashes on the keyboard, and patches from /dev/urandom. Actually, she almost exclusively patches from /dev/urandom, for lack of any other insight, but everything she writes gets quickly thrown out to user testing (the environment), and she gets immediate feedback.

Along the way, she randomly forks the code into various branches and tries out different things. Over time, the branches diverge to the point where they are no longer interoperable, but some relics of their shared history remain. Eventually, many branches are discarded.

The third saving grace is that she literally has all the time in the world. She’s not constrained by deadlines, and it took her a few hundred million years to publish the first usable code. In this case, it was code that accomplished the singular task of making more copies of itself. And why not? Since she had no goal in mind, it should be obvious that the code most likely to persist would simply be code that replicates itself. All other tasks that it eventually accomplishes are secondary to that goal .

And that’s how it goes for millenia. Despite the massive inefficiency of this process, it works because our programmer is extremely productive. She shotguns the problem. She iterates on the code billions of times a day and handles a bug tracker of unfathomable proportions.

What is the end result? Incredibly complex code with little underlying logic that just works, most of the time. That is why Kurzweil and many computer scientists and engineers are wrong about their predictions of biomedical progress. You’re not just reverse engineering the Kinect or some proprietary code, which you know has a purpose and internal logic. Cracking the genome is not like cracking MD5, where the time it takes to arrive at a cryptographic solution is inversely proportional to computing power. To understand the genome, and biology more generally, you have to decipher every line of code empirically.

That requires research on real biological systems, which is hamstrung by things like reproductive output, generation time, ethics, and pure luck. David Linden is right: empirical progress in the biological sciences is much more linear than you imagine. And if some aspects of it are exponential, the exponent is much smaller than you think.

Kurzweil likes to use the Human Genome Project as an example of exponential growth. It took seven years to sequence the first 1% (or thereabouts), and most scientists directly involved in the project thought it would take much longer to finish the full genome. But they underestimated the power of exponential (sequencing technology) growth, and several doubling times later, a full (draft) sequence was published in 2003. That’s great, but then what?

Then the hard problem of empirically deciphering the genome became relevant, and a decade later, we still don’t have personalized medicine. We’re not even close to understanding the human genome in its entirety. We didn’t even fully appreciate the importance of non-coding RNAs until after it was completed. We have the source code but we don’t know what to make of it.

In the widest interpretation, the genome is code, but it’s nothing like the code that you write. The programmer is nothing like you. And deciphering biology is not a straightforward engineering task, because it wasn’t made by engineers. You can’t look for internal logic, because there is none. You have to decipher each line empirically, and that’s hard work which is not subject to Moore’s Law. You can’t project current computational trends to some distant point in the future and seriously predict the emergence of a specific discovery or innovation to within a few years.

That might work when the entire system is a fabrication of goal-oriented human logic. It doesn’t work for biology.

Just as sequencing the human genome didn’t automatically produce usable knowledge of genetics and development, Kurzweil’s predictions about high resolution neural imaging will not automatically produce usable knowledge of the brain and consciousness. That will require lab work.

In the end, the analogies are superficial, and you can’t reason your way to biological conclusions through the lens of computer science or engineering. You should really learn biology and the travails of biological research to make insightful statements about them.