I can't tell if Guillaume really believes all the contradictory and ill-conceived things he says or if he's just trying to optimize for an internet cult following and is willing to sacrifice intellectual integrity. Very unfortunate. About half of the things he said seem to be purposefully glossing over key details that would either make his statements completely false or require a nauced refinement. One important statement though is his disbelief in a recursive superintelligence - the idea that at some point we might build an AI that's general and capable enough to improve itself over and over again resulting in an intelligence explosion. His main arguments against that are: 1) "we've been trying to do meta-learning for decades" and haven't done it --> Just because we've worked on something for years doesn't mean it can't be done. We're improving these systems every day, including ways in which they can contribute to AI research and be more agentic. 2) "at the end of the day you're still limited by compute" --> we/it would just build more compute farms, or it could create new more efficient learning algorithms and architectures that require less compute, or something else we haven't thought about. I don't take the opposing side to all of his statements, but I want you to notice his lack of nuace and frequent disregard of obvious details while formulating arguments.
This is great! Good job both
how is this ted channel so hidden??
if you are trying to lose all credibility by inviting verdon?
then you are are doing a fantastic job!
I can't tell if Guillaume really believes all the contradictory and ill-conceived things he says or if he's just trying to optimize for an internet cult following and is willing to sacrifice intellectual integrity. Very unfortunate.
About half of the things he said seem to be purposefully glossing over key details that would either make his statements completely false or require a nauced refinement. One important statement though is his disbelief in a recursive superintelligence - the idea that at some point we might build an AI that's general and capable enough to improve itself over and over again resulting in an intelligence explosion. His main arguments against that are:
1) "we've been trying to do meta-learning for decades" and haven't done it --> Just because we've worked on something for years doesn't mean it can't be done. We're improving these systems every day, including ways in which they can contribute to AI research and be more agentic.
2) "at the end of the day you're still limited by compute" --> we/it would just build more compute farms, or it could create new more efficient learning algorithms and architectures that require less compute, or something else we haven't thought about.
I don't take the opposing side to all of his statements, but I want you to notice his lack of nuace and frequent disregard of obvious details while formulating arguments.
Decel all the way people need jobs.
The French decel speaker is a buzzkill