My AI Timelines Have Sped Up (Again)

I&#x27;m no AI expert, I don&#x27;t have a lot of domain knowledge but what irks me about reading through this entire blog post is the idea that we are hinging on a definition of AGI that apparently accounts for 95% (!) of &quot;economically valuable work&quot; while basically not talking _at all_ about the trades. No mention of construction, manual labour, mechanics, welders, painters, nothing!<p>Is the author trying to say that all that work falls into the 5% left over after you replace the stock brokers, programmers, artists, etc.?<p>I understand no offence is meant, but this feels extremely naive. At best the work programmers do is in the minority of &quot;economical valuable&quot; work that we strongly rely on for our world economies to keep chugging along.<p>Quick and dirty searches show that some 12% of employed people are in healthcare in the USA, 9% in leisure and hospitality, and 2% in education. 23% of people in the USA are employed in economically valuable work that is very personal, and very difficult for technology to fully replace.<p>Sorry if I&#x27;m rambling, I think I just did not enjoy the complete lack of mention of blue-collar workers here. I don&#x27;t think I&#x27;m yet in the camp betting on AGI coming out of the woodworks in the next few decades.

I&#x27;m glad an AI researcher is posting their personal thoughts. As a fellow researcher and developer in the AI space, I conceptualize AGI as a self-domain-learning entity that is able to manage itself and understand how to affect&#x2F;intervene its actions and effectual consequences (humans don&#x27;t reach this on a general level until early teens to mid twenties, if ever). Walking is a solid example of this at a specific level. Singleton objective functions and patterns across probabilities of states are insufficient -- there needs to be ongoing interpretation and architecture extrapolation.<p>In short -- current AI approaches are great at interpolation and still perform poorly at most extrapolation exercises that a human can take on.

My AI timelines says next AI winter will come in 10 years, when people find it&#x27;s a vaporware ... again.

“Work” to us, in the tech bubble is sitting at a desk and doing some typing and clicking and sometimes decision making. The world out there is much more different. Work for many people is physical or very human (like a teacher or psychologist).<p>Yes it’s easy to guess the job the involves mostly moving information around would be automated but that’s not all jobs by any means.

Imagine you learn about some new conjecture in mathematics, and you try to predict when it will be resolved. Your prediction says that there is a 50% chance that the conjecture is resolved in the next 25 years, and a 90% chance it is resolved in the next 50 years. So, conditional on the conjecture not being resolved in the next 25 years, there is an 80% chance the conjecture is resolved in the subsequent 25 years. Wouldn&#x27;t that be a little strange?

Title should be: &quot; My AI Timelines haven&#x27;t sped up, but now I think there&#x27;s a 10% chance of AI dominance in 5 years&quot;<p>His AI timelines haven&#x27;t sped up, only clarified

I&#x27;m an &quot;overhang&quot; bear, while incremental progress will be made - it&#x27;ll come at too high of a cost. The larger, slower, more expensive models have limited applications(mostly fine-tuning the economical ones) so there will be ultimately less investment. Unless there is a new unlock, we&#x27;ll be left with mediocre 3.5 models for most real-time apps and excitement will fade.

It is not quite 70 years since this proposal was written:<p><a href="http:&#x2F;&#x2F;www-formal.stanford.edu&#x2F;jmc&#x2F;history&#x2F;dartmouth&#x2F;dartmouth.html" rel="nofollow">http:&#x2F;&#x2F;www-formal.stanford.edu&#x2F;jmc&#x2F;history&#x2F;dartmouth&#x2F;dartmou...</a>

“Today is the worst AI will ever be.” I’m gonna steal this one. I can’t say the same about web development, for example, or the US political landscape, or climate change, or mental health, etc.

Mine was 50&#x2F;50 at 2035 after GPT-2, and with GPT-4 it is more like 90&#x2F;10 with 50&#x2F;50 being sometime in 2028 (that&#x27;s for pure intellectual work).

Is really a sped up?<p>The target of 95% is still 2070, the 50% is still 2045, means the progress between 2035 and 2045 is now less than before.<p>So the timeline has sped up and slowed down at the same time

Yes.... what the military needs most is an auto-complete tool that was trained on Reddit material. (GPT-4)

He should take into account in 3 months the time line will accelerate 20% and compound it and do a discount to present and give a more accurate answer.<p>Let me do his job for him: 90% AGI will happen by 2030.

For fck&#x27;s sake, please stop using this font. I know Apple started the trend but even they went back to sane alternatives. The narrow lines of this font are a torture to the eyes.