A quick heads up what the conversation with Greg and Ilya will look like:
OpenAl's Mission: Want to confirm what it should be. Given the AGI timelines keeps compressing, they think it should be
"build AGI before anyone else" (thinking of Google and China, specifically). This isn't a departure, per se, but want to
confirm that you still think the best possible future would be brought about this way? This would of course involve being
much less "open" and require decisions that would prioritize speed over other considerations (safety, government
involvement, etc).
Hardware: if it's true that OpenAI should dead set on building AGI before anyone else, the critical factor is hardware. They
think if all resources exhausted, best you can get in terms of human capital/ ingenuity over DeepMind 2-3x, whereas if you
buy that HW is really going to depart from Moore's law and hit I Ox per year it's not even a contest as to which battlefield
more critical right now. They are at the point where they want to bet the farm on the right hardware strategy.
Cerebras v A Billion Cores: In an ideal world both would be pursued. Cerebras is a high beta experiment, they understand
that, but think that if it succeeds it's so game changing that if Google gets it, it's game over re: race for AGL Reason is that
even ifwe started on a billion core system today, it's still 18 months behind and that's enough to lose the battle. Want to keep
working on billion core idea since higher probability of eventual success but think it's prudent to figure out the right strategy
to still merge or exclusively partner with Cerebras as well. They believe the combined entity, assuming OpenAI converts from
being a non-profit, would be more valuable?? (Note: Andrew Feldman reached out to them to discuss next steps. That meeting
is planned for next week)
Dota supremacy... but at what cost: The compute for Dota 5v5 is hard to estimate but would get into the $50M range on the
high side. They don't have perfect data but estimate that AlphaGo was somewhere in this ballpark. That begs the question of is
this worth the price tag? They think yes, since it's likely to have an equal or better outcome to AlphaGo but want to both
confirm that you agree and also make sure there is a path to funding should it come to that.