Update index.html

Author: dreaming-panda

index.html

@@ -483,9 +483,9 @@ <h4 style="margin-bottom: 16px;">
               </div>
               <br>
               <p>
-                <strong>Figure 5: Figures 5a and 5d</strong> show that sparse attention significantly improves performance and scalability, enabling higher accuracy at lower compute budgets. 
+                <strong>Figure 5: Figures 5a and 5d</strong> show that sparse attention significantly improves performance and scalability, enabling higher accuracy at lower cost budgets. 
                 <strong>Figures 5b, 5c, 5e, and 5f</strong> highlight that both 8B and 32B models benefit from sparsity, with <em>50–60 percentage point</em> gains in low-cost settings and consistent <em>~5 point</em> improvements even in high-cost regimes. 
-                Notably, sparse models reach these performance levels at much lower costs —for context, 10<sup>5</sup> Tera-eFLOPs is just 22 seconds of B200 usage.
+                Notably, sparse models reach these performance levels at much lower costs. For reference, 10<sup>5</sup> Tera-eFLOPs is 22 seconds of B200 usage at 100% utilization.
               </p>
             </div>
             <br>