As you can see all LLM benchmarks are normalized between 0 and 100. The models are already near 100. What's next? Answer: performance growth: reaching 1Ktok/s barrier and growing further, decreasing token cost. All this will lead to specialized processors for LLM and a new https://t.co/8BtYsElQWR

1 min read Original article ↗

As you can see all LLM benchmarks are normalized between 0 and 100. The models are already near 100. What's next? Answer: performance growth: reaching 1Ktok/s barrier and growing further, decreasing token cost. All this will lead to specialized processors for LLM and a new n, as it was with the PC revolution

user avatar

Here's my take on what's next in LLM/GPT space. All benchmarks are normalized between 0 and 100 and models are already saturated approaching values of 100. I.e. they won't be above 100, and 100 won't be either. So, the only way is: 1. Increasing token speeds to 10Ktok/s and