AI 2027 co-author here, thanks for engaging with the scenario! This is exactly the sort of disagreement that we were aiming to draw out. I may reply later with more thorough replies but for     See More ll point to our timelines forecast (YT not letting me include a link but search "timelines forecast ai 2027") which does aim to take into account the gaps between benchmarks and the real world. Totally fair if you estimate that the gaps will take longer to cross or otherwise disagree with the methodology, but we are aware of these gaps and estimate their size as part of our forecast.    See Less

AI is a program created to cheat and fool humanity. If we don’t add anything to the program it becomes obsolete. It is not Artificial Intelligence like we call it. It is nothing but a prog     See More e around the world will use this program to cheat, steal or hack    See Less

Left vs right bias? How is that being standardized?     See Less

AI timeline predictions are notoriously conservative...

Great video! 🥇     See Less

Excellent small models is what Llama was most known and used for and they completely blew it.     See Less

Couldn't love your content more man     See Less

Thank you for continuing to de-escalate the anti-China rhetoric.     See Less

Something that you do not seem to have considered: You said that you don't think feedback loops could develop and advance that quickly. I presume that you are making the utterly unwarran     See More tion that there is not a radically more optimized architecture which would enable the 'supermodel' to drastically shrink its size and thus far more quickly grow in its capability even within existing compute bounds. There is no reason to believe, and I think substantial reason to disbelieve, that the models we have now are in any way 'optimal'. They still have profound difficulties dealing with things like absolute (not fuzzy) logic and absolute reasoning. And those happen to be the things that we already have things like high-throughput Datalog engines which can run circles around them with far less compute. If it structures itself to offload absolute reasoning to such a system, and eliminates all of what is effectively just brute-forcing trying to fit a binary peg into a floating point hole from itself, slims itself down by jettisoning everything it knows about things which it knows, on balance, are false, no longer relevant to the world which it exists in, etc, and THAT is what the feedback loop is focused on, I could certainly see it running away quite quickly.    See Less

I feel like the "make it more right-wing" objective should be ringing alarm bells...     See Less

Meta alone has over 600000 H100s. They have plenty to experiment for the foreseeable future. And if they stopped overpaying 40x for nvidia cards they would come down in price     See Less