Will roboticists scale up their compute game before 2027?
7
Ṁ12502026
71%
chance
1D
1W
1M
ALL
Currently, robotics models appear to be using much less training compute than they are able to, because of constraints in data quantity and variety. Will a robotics model break this trend, and be trained with within one order of magnitude of compute of the largest frontier models to date? For example, if the to-date most frontier LLM is trained with 1e27 FLOP, a robotics model being trained with 1e26 would resolve this market positively. If this information is not given, I will resolve based on epoch.ai estimates, or other reliable sources if epoch.ai stops providing such estimates.
This question is managed and resolved by Manifold.
Get
1,000and
3.00
Related questions
Related questions
Will models be able to do the work of an AI researcher/engineer before 2027?
26% chance
Will 5K humanoid robots be manufactured before 2029?
90% chance
Will 5K humanoid robots be manufactured before 2028?
90% chance
Will 200K humanoid robots be manufactured before 2028?
51% chance
Will 200K humanoid robots be manufactured before 2029?
65% chance
Will 2K humanoid robots be manufactured before 2029?
92% chance
Will 50K humanoid robots be manufactured before 2029?
75% chance
Will 50K humanoid robots be manufactured before 2028?
70% chance
Will 100K humanoid robots be manufactured before 2028?
65% chance
Will 5K humanoid robots be manufactured before 2030?
91% chance