A long but fascinating discussion with many points I disagree with but still want to know more about and definitively worth thinking about. Canadian Prepper will naturally err towards doom and gloom but even without falling into the catastrophe trap, entertaining these issues, causes and consequences, risks and remedies is I believe a worthwhile exercise.
Over the years I have become more and more ambivalent towards AI. Still very enthusiastic but at the same time more circumspect. The risks are real, not the extinction part which is still farfetched but the mayhem part where AI generates problems faster than we can fix them which we are already witnessing with more complex and frequent hacks.
The most fundamental problem which the whole alignment movement is missing is the emergent aspect of AI. You CANNOT understand AI with a reductionist mindset and approach. It is like trying to understand psychology with chemistry. When Connor Leahy says that we do not understand how intelligence and emotions are generated he is completely missing the point. You cannot "explain" intelligence emerging out of trillions of loops from the transformers. It is an emergent phenomenon and consequently the right approach is to observe and describe, not to analyze what happens at the GPU level where there is literally nothing to observe.
Likewise, many of the points discussed are little more than wishful thinking. The concerns are real but a "pause" in AI research or testing is ludicrous. It will simply not happen. Conversely, testing the newest models for defensive purpose before releasing them sounds like a good idea. So, in the end, I tend to agree concerning the general theme of the discussion: There are better ways and not so good ones to approach AI research. We need an open, non dogmatic approach to the subject. The question is to know if our society is capable of handling such a complex discussion in a fruitful manner. This is not at all obvious.
No comments:
Post a Comment