Global marketsUpdated 3h ago

Anthropic says compute multipliers can compound into an advantage

Anthropic’s roadshow presentation gives a 4x compute multiplier two possible uses: spend 25 units to train the same model that would otherwise take 100, or use the same 100 units to train a more intelligent model that would otherwise require 400. One option uses less physical compute for the same model; the other puts the same capacity toward a more capable one. Anthropic argues that these multipliers can build on one another into an advantage that’s hard to replicate. It believes they could accelerate as the company approaches recursive self-improvement, with Claude potentially discovering them entirely by itself. The long-term bet is that Claude could find the next multiplier on its own.

1 publisher1 report

Coverage started 3h ago