Updates and Clarifications

Fighting a decade of Machine Learning momentum.

It seems the well received posts follow many of the same arguments, not about viability or milestones, but compute hardware trajectories or the pace of arXiv submissions reaching a point where even those in the field can’t keep up anymore. I’m not suggesting people haven’t made well reasoned arguments, but will say they don’t break new ground. If everyone is just reading thru the posts, nodding their head, and saying “yep,” then where is the challenge to existing views?

ML is moving in one direction, and it has built its own cult following. Miles Brundage posted a thread on Twitter earlier today about the importance of having different views (specific example was within OpenAI, but assume he meant it to apply broadly to the AI community and other research groups). Nothing in that thread gave any indication of having a different view about the core methods involved. Along the lines of “scale is all you need,” there is a sense within ML that SOTA has proven (to everyone moving in the same direction) xNN’s are the only valid solution.

The clarification required is that ML is not AI, and AI is not AGI. Results are not the same as abilities. Brute force on a narrow task (where a DL method happens to work) is not on a path toward a range of methods, in architecture, that can match or exceed human level cognition.

Risk assessment is bifurcated into where current systems fail (and then just guessing at how much worse it can be when they have 20x more compute), and into a philosophical realm where an AGI system has abilities that are 20x more than any one human (which is also based on incorrect assumptions about architecture and scale).

I don’t see anyone here drifting away from the hardware aspect, and what the latest headline grabbing trend might suggest. I do ask that an effort be made to isolate future thinking from all current examples. There will be a few start-ups, perhaps in less than 10 years, that set the new standard on what can be achieved. Work from that assumption.