At this point, everything around the LLM is woefully behind the performance of LLMs themselves. If all LLM development stopped tomorrow, I think there would still be years of catching up to do.
I am an extreme user of LLMs and other AI, and the limit is creativity and adaptability. It is no longer the “intelligence” of the models that counts, they are already insanely good now. I use Kimi K3 and it blows me away.
Well even the human user aside, I’m thinking of agent harnesses, tools, prompt “hacking,” constrained and exotic sampling, augmentations like control vectors or loras, RAG, databases, multimodel thinking…
The basic chatbot interface is hardly tapping into these models at all. They’re very effective tools when specifically constrained, but most are still figuring out how to do that.
Yeah.
At this point, everything around the LLM is woefully behind the performance of LLMs themselves. If all LLM development stopped tomorrow, I think there would still be years of catching up to do.
Most definitely.
I am an extreme user of LLMs and other AI, and the limit is creativity and adaptability. It is no longer the “intelligence” of the models that counts, they are already insanely good now. I use Kimi K3 and it blows me away.
The limit is what my mind can conjure up.
Well even the human user aside, I’m thinking of agent harnesses, tools, prompt “hacking,” constrained and exotic sampling, augmentations like control vectors or loras, RAG, databases, multimodel thinking…
The basic chatbot interface is hardly tapping into these models at all. They’re very effective tools when specifically constrained, but most are still figuring out how to do that.