I expect that many companies will realize that an LLM is a tool that can help with certain tasks, but cannot fully replace most workers because there is a ton of context in people’s jobs that cannot be condensed into language. Similarly, many companies are now realizing with token-based pricing that frontier LLMs are not cost effective in many applications. You shouldn’t have a natural gas fired data center in Tennessee running a model to proofread your emails. You can have a local model do so MUCH more cheaply. This will leave the speculative data center companies holding the bag on a lot of hardware and capacity that is not actually needed.
When training ai, I was always confused that they never gave a sht about tokens, like that was never something you picked any model over another for or reviewed, it often says to ignore the length but make note of how long it took, tbh that could be because of reasoning text not shown to the user
That’s interesting to hear. I wish efficiency had been considered from the start. It seems like there has been a ton of waste. They should only do the calculations needed to achieve the user objective. But they were prioritizing market share over efficiency. Maybe they thought they could afford another couple years of subsidizing wasteful use to build market share, but it hasn’t turned out that way.
I expect that many companies will realize that an LLM is a tool that can help with certain tasks, but cannot fully replace most workers because there is a ton of context in people’s jobs that cannot be condensed into language. Similarly, many companies are now realizing with token-based pricing that frontier LLMs are not cost effective in many applications. You shouldn’t have a natural gas fired data center in Tennessee running a model to proofread your emails. You can have a local model do so MUCH more cheaply. This will leave the speculative data center companies holding the bag on a lot of hardware and capacity that is not actually needed.
When training ai, I was always confused that they never gave a sht about tokens, like that was never something you picked any model over another for or reviewed, it often says to ignore the length but make note of how long it took, tbh that could be because of reasoning text not shown to the user
That’s interesting to hear. I wish efficiency had been considered from the start. It seems like there has been a ton of waste. They should only do the calculations needed to achieve the user objective. But they were prioritizing market share over efficiency. Maybe they thought they could afford another couple years of subsidizing wasteful use to build market share, but it hasn’t turned out that way.