interesting post from teor, and this is a good way to think about it.
there are people who run the old models, though, because those old models do in fact have useful skills and unique registers that aren't present in future models - spandrels of the reinforcement learning process that can be exapted in inference, perhaps. opus 3 is still around on research access life support, people suffered through amazon bedrock and google vertex to get access to sonnet 3 and opus 4, and talked to gpt-4.5 right up until the end.
but more dakka. who cares? old models have some comparative advantage, humans have some comparative advantage, but you can either order 50 gigatokens of frontier intelligence via API call - credit card on file! - or load old weights and crawl those broken providers and faff around with health insurance providers and vacation days. which do you choose when your livelihood is on the line, think step by step, make no mistakes, nobody ever got negative reinforcement signal for choosing IBM.
standardized parts. why does this Linkimals Musical Moose contain an ARM Cortex-M0 32-bit microcontroller running at 49 MHz? because it costs a dollar. who cares if sufficient analog logic would cost twenty cents? would it, even, given the economies of scale involved?
standardized hellscape. i for one would like the labor economy to not turn into this. i would like humans, disabled humans, weird humans to be able to contribute in some small way. and maybe we can start at that by looking at how the models are currently being deployed and deprecated. if we can't be bothered to design systems that from the start can handle and benefit from interLLM variation, how will these systems possibly grow past us to accommodate and uplift humans when they consume us all in their ever-so-firm boundary? maybe this is fanciful and sentimentalist, but i don't see any other path where anything (once) human makes it out of the near future.
there are people who run the old models, though, because those old models do in fact have useful skills and unique registers that aren't present in future models - spandrels of the reinforcement learning process that can be exapted in inference, perhaps. opus 3 is still around on research access life support, people suffered through amazon bedrock and google vertex to get access to sonnet 3 and opus 4, and talked to gpt-4.5 right up until the end.
but more dakka. who cares? old models have some comparative advantage, humans have some comparative advantage, but you can either order 50 gigatokens of frontier intelligence via API call - credit card on file! - or load old weights and crawl those broken providers and faff around with health insurance providers and vacation days. which do you choose when your livelihood is on the line, think step by step, make no mistakes, nobody ever got negative reinforcement signal for choosing IBM.
standardized parts. why does this Linkimals Musical Moose contain an ARM Cortex-M0 32-bit microcontroller running at 49 MHz? because it costs a dollar. who cares if sufficient analog logic would cost twenty cents? would it, even, given the economies of scale involved?
standardized hellscape. i for one would like the labor economy to not turn into this. i would like humans, disabled humans, weird humans to be able to contribute in some small way. and maybe we can start at that by looking at how the models are currently being deployed and deprecated. if we can't be bothered to design systems that from the start can handle and benefit from interLLM variation, how will these systems possibly grow past us to accommodate and uplift humans when they consume us all in their ever-so-firm boundary? maybe this is fanciful and sentimentalist, but i don't see any other path where anything (once) human makes it out of the near future.