We’re hitting logarithmic scaling with the model trainings. GPT-5 is going to cost 10x more than GPT-4 to train, but are people going to pay $200 / month for the gpt-5 subscription?
Businesses might pay big money for LLMs to do specific tasks. And if chip makers invest more in NPUs then maybe LLMs will become cheaper to train. But I am just speculating because I don’t have any special knowledge of this area whatsoever.
4o is also not really much better than 4, they likely just optimized it among others by reducing the model size. IME the “intelligence” has somewhat degraded over time. Also bigger Model (which in tha past was the deciding factor for better intelligence) needs more energy, and GPT5 will likely be much bigger than 4 unless they somehow make a breakthrough with the training/optimization of the model…
4o is optimization of the model evaluation phase. The loss of intelligence is due to the addition of more and more safeguards and constraints by the use of adjunct models doing fine turning, or just rules that limit whole classes of responses.
We’re hitting logarithmic scaling with the model trainings. GPT-5 is going to cost 10x more than GPT-4 to train, but are people going to pay $200 / month for the gpt-5 subscription?
Businesses might pay big money for LLMs to do specific tasks. And if chip makers invest more in NPUs then maybe LLMs will become cheaper to train. But I am just speculating because I don’t have any special knowledge of this area whatsoever.
But it would use less energy afterwards? At least that was claimed with the 4o model for example.
4o is also not really much better than 4, they likely just optimized it among others by reducing the model size. IME the “intelligence” has somewhat degraded over time. Also bigger Model (which in tha past was the deciding factor for better intelligence) needs more energy, and GPT5 will likely be much bigger than 4 unless they somehow make a breakthrough with the training/optimization of the model…
4o is optimization of the model evaluation phase. The loss of intelligence is due to the addition of more and more safeguards and constraints by the use of adjunct models doing fine turning, or just rules that limit whole classes of responses.
Is it necessary to pay more, or is it enough to just pay for more time? If the product is good, it will be used.