Biological computation and the nature of software A new paper is been getting some attention. It makes the case for biological computation . (This is a link to a summary, but there’s a link to the actual paper at the bottom of that article.) Characterizing the debate between computational functionalism and biological naturalism as camps that are hopelessly dug in, the authors propose that the brain does do computation, but that it’s a very different kind from the type done in the device you’re using to read this, which they call “biological computation.” The differences are that biological computation is a hybrid between digital (discrete) and analog (continuous) computing, there is no clean division between software and hardware, between algorithms and implementation, and that metabolism and energy constraints shape the processing that happens. They sum it up as, in the brain, the algorithm is the substrate. The authors argue that to build artificially conscious systems, it may be necessary to go with a different physical ontology, one that is closer to the way biology works. Let me start by saying that this paper is a big improvement over the usual arguments about the distinctions between computers and biology. The authors are making a real effort to identify what supposedly makes biology unique. Most of what they’re saying already accords with my own understanding of how the brain works , and what’s different about its computation . There are a few points where they try to pass off speculation as established fact, but those are nits. That said, I think they oversell some of their points. For example, the distinction between analog and digital is often less than it appears. We listen to music and watch movies all the time in digital formats that were originally recorded in analog. Yes, something can be lost in the translation from continuous to discrete signaling, but in an analog system there is always variance noise, variations between a system’s processing, both with other systems of the same type, and between runs in the same system. The trick is for the translation to reduce the quantization noise, the distortions from moving to a discrete format, so that they’re less than the variance noise in the original. Another is the aspect they call scale inseparability, the idea that the brain doesn’t use the layers of abstraction that technology uses. These layers exist in technology to make the engineering easier to understand and maintain, for engineers. Evolution doesn’t care about understanding so it’s not a factor in how biological systems are organized. The authors use this to imply that the software / hardware divide may be something the technology side will have to give up. That the algorithm may need to be in the substrate as it is with biology. I think this represents confusion about what software actually is. We usually talk about software as a set of instructions that a processor follows. In most cases, it’s convenient to think about it that way. But at a more physical level, it makes more sense to think of software as a configuration of hardware. So when software is running on hardware, the algorithm is always the substrate. The real distinction here is that technological computers are designed to be reconfigured on the fly. This is actually an amazing achievement when you stop and think about it. I often see articles marveling at the brain’s plasticity, its ability to rewire itself. But your computer’s memory can undergo wholesale reconfiguration on demand by loading a new software package, something brain’s can’t do, at least not quickly. Of course, this comes with vulnerabilities brains are far less susceptible to. One reason computers can be hacked is this ability to massively reconfigure. Not that brains are completely immune. Ant brains can be hacked by a fungal infection , and cat owners can be infected with a parasite that makes them like their cats more , and that’s aside from the ability of advertisers and propagandists to hijack our brain’s reasoning to introduce notions we might otherwise resist. But it’s a harder thing to do effectively in biological systems. What’s important to realize is that anything that can be done in hardware can, in principle, be done in software, at least once a minimal general computing platform is in place. You can run software that emulates other hardware platforms so you can run their software. It is true that doing it in hardware is often far more efficient in terms of performance and energy, but that comes with reduced flexibility. It’s why we now run word processors on our general purpose computers instead of the old word processing machines that once existed. So I don’t think the fact that current AI runs on software neural networks, in and of itself, is a showstopper. Another difference is that the brain operates with massive parallelization, far more than any current technological system. These systems can still perform something like the brain’s processing in software because they operate millions of times faster. Although the addition of GPUs, designed with parallelization in mind, help a great deal. But that, I think, gets to a valid concern the authors make about energy constraints. Discrete processing, and doing things with software instead of hardware, come at a cost in terms of energy and performance. This is something I do think AI researchers should be paying more attention to. All we need to do to understand how far current AI is from animal intelligence, much less human level, is look at the vast amounts of data and energy it requires to do what it does. Datacenters are sucking the power grid dry to meet their energy demands. All of which speaks to how crude the technology remains in comparison to biological intelligence. But this energy constraint issue is broader than just trying to reproduce biological processes. I think it’s a problem for all technological computing. And it will likely eventually result in architecture changes. Understanding how biology does it may be important, but I tend to doubt the solution will be doing it exactly like those systems. And this gets to a sentiment that I detect in the paper and write ups about it. It’s the idea that consciousness is a ghost in the machine, one we need to find the magic ingredients for so we can generate it. I think this is fundamentally the wrong way to think about it. Neuroscientist Hawan Lau, I think, in a Bluesky post, sums up the issue . Why do we think this might be true for consciousness when it isn’t for so many other things the body does, like motor control? All that said, I do like the term “biological computation.” It admits that the computation in brains is different while still acknowledging the important ways it’s the same. I suspect that won’t be enough for those strongly convinced computationalism is wrong, but it still feels like useful progress. What do you think about the points the authors make? Or my take on them? Are they right that a new hardware architecture is required? Or would even that be enough? Does the “biological computation” term strike the right balance? #AI #ArtificialIntelligence #BiologicalComputation #ComputationalFunctionalism #Consciousness #functionalism #Neuroscience #Philosophy #PhilosophyOfMind