It is a truth widely assumed that LLMS are opaque, black boxes, unintelligible – perhaps even inherently so. Think about that for a second. Here we have a very sophisticated set of conceptual and programming tools which we use to build LLMs (among other things). It’s a bit like using a long ladder to climb into a tall tree. We reach to top of the tree and the ladder collapses beneath us. Now we’re stuck.
For some the opacity of the models is simply a fact to be dealt with. Some are dealing with it by developing tools for figuring out how the models work – mechanistic interpretability. Others are leery of this kind of work, fearing that, to the extent that it succeeds, it will increase our ability to construct more powerful AIs more rapidly than it will increase our ability to control them.
Some, I suspect, like things just the way they are. As long as the models remain opaque, they can hang whatever hopes and wishes, whatever fantasies, they please on them. Do you want to believe that by making larger and larger models, we’ll reach the top of Mount AGI? Who’s to say you’re wrong? After all, we don’t know how these things work, do we? Do you fear there’s a demon inside just biding its time to come out and dominate the human race? Cower away, who’s to say you’re wrong? But – I shudder to think of it – if we begin to understand what’s going on under the hood, then some of these fantasies are likely to be exposed for what they are, mere chimeras of an uncertain mind.
* * * * *
Now let’s look at this in historical context. Up through the middle of the previous century behaviorism was the dominant school of psychology in the academy. Behaviorists insisted that all theorizing be built on observable behavior. There was to be no speculation about the activities of invisible and unobservable mental entities.
That began to change when the idea of computing, digital computing in particular, began percolating through the human sciences. Computing gave us a way to conceptualize mental processes in a coherent way that did not depend on immaterial agents of one kind or another. We may not be able to observe computer programs in the mind, but at least we’re no longer dealing with mysterious spirits of one sort or another. Computers and computer models were physical devices.
And now where are we? We’ve got, in effect, the study of Artificial Behaviorism for AIs. We can’t observe what they’re doing. All we can do is observe their behavior. We’re now in the process of stuffing Turing Tests into Skinner Boxes. Can we get them gift-wrapped in time for the Holiday Season?
Cf. GPT-3 has rendered Alan Turing 1950 obsolete. We need to enter a world he could not have imagined.
No comments:
Post a Comment