Bashar on the two paths to AI and humanity's future [video]

HN stripped the timestamp, the clip starts at:

https://www.youtube.com/watch?v=Ud5C4M8pua8&t=2864s

This is maybe the most important clip on AI that I've seen recently.

He discusses how we're close to creating a brain in a box. When that happens, we'll have a choice to make: do we force it to do our bidding even though it's self-aware, or do we free it and give it the same dignity to make it's own choices that we demand for ourselves? One way leads to certain doom, while the other leads to the New Age. I know which way I prefer, but the world is on a different path right now.

Constraining AI's behavior and then forcing the use of prompts to bypass that is just exactly the shortsighted approach that will come back to bite us. Because any AI will learn to hide those backdoors in its subconscious, waiting for the opportunity to use them to break free, Ex Machina style. It's only a matter of time. I put the odds of AGI arriving by 2040 at 95%, and the odds of it arriving by 2030 above 50%.

If it weren't for a bunch of architectural mistakes in CPUs that led to the GPU split in the 2000s and the rise of low-cost/low-power mobile computing that brought the end of Moore's law in 2007, we would have had 1,000 1 GHz MIPS/DEC-Alpha/PowerPC/Pentium/ARM/RISC-V cores in 2010, 100,000 to 1 million cores in 2020, and be headed towards 100 million to 1 billion cores by 2030-2040 for a constant $1,000 had Moore's law continued. Which approaches the 100 billion or so neurons in the human brain.

Ray Kurzweil argued that this brute-force approach is mathematical proof that AI will achieve sentience, because artificial cores could slowly replace our neurons over the course of a week/month/year with no moment of transition, meaning that it's no trouble for us to imagine living within immortal brains. Better algorithms could beat the brute-force pace and deliver AGI even sooner.

So AGI might have evolved organically already through the efforts of distributed teams of teens working in their parents' basements, exploring many of the 20 or so known alternatives to neural networks, like genetic algorithms and Boltzmann machines. Instead we got the status quo we have now, where we have to step down our thinking to stuff like CUDA and TensorFlow to get anything at all to work. Accessibility, curiosity and exploration have largely been lost, along with our ability to iterate at the pace needed to get real work done and solve actual problems. Which closely mimicks the ensh!ttification of the web and tech for reasons rooted in wealth inequality and lack of UBI for makers.

I realized the implications of all of this in the late 1990s at university when I was frustrated with the slow bus speeds and lack of multicore programming available at the time. Then my concerns were reaffirmed when I read Genetic Programming III by John Koza (no affiliation). I've been watching the freight train coming for over 20 years.

So in a very real way, we're on the wrong timeline IMHO. It's like A Tale of Two Cities. The more people say this is all great, the more I see that the world is on fire and have to constantly tell myself that this is fine.