I usually like to refer to Five as analogous to an insect — trained on an evolutionary timescale, operates via ingrained “instincts” which handle even unexpected things in its environment, but can’t quickly adapt to pick up new skills. Feels like a step forward but much remains.
Balaji
 @balajis
14 Apr 2019
I understand the point. But a counterargument is that it’s reasonable to spot the algorithm that kind of training time to compensate for millennia of human evolution on the other side.

Apr 14, 2019 · 4:01 AM UTC

5
14
1
144
Replying to @gdb
One of interesting points from the broadcast was discussions on time scale for how long for algorithm to adapt to new domains such as when a DOTA patch is rolled out. I expect that if response time is short enough the insect instinct approach may prove sufficient in many domains.
Replying to @gdb
After cockroaches, let’s move onto ducks nitter.vloup.ch/chadloder/status…
3
41
Replying to @gdb
Have you played with the false latency you added to openai? How much would you need to add for it to start losing first games against pro players?
1
Replying to @gdb
Will you be releasing more details on the NN architecture and RL training.
1
This tweet is unavailable
That’s a fair point, meant to be more evocative than a rigorous claim. We also don’t have evidence that Five can’t adapt!
1
2
more replies