The capabilities and general utility of the models are increasing on an entirely different trajectory than model names - the information you posted is 99% dependent on internal OAI processes and market activities as opposed to anything to do with AI.

I'm talking more broadly, as well, including consideration of audio, video, and image modalities, general robotics models, and the momentum behind applying some of these architectures to novel domains. Protocols like MCP and automation tooling are rapidly improving, with media production and IT work rapidly being automated wherever possible. When you throw in the chemistry and materials science advances, protein modeling, etc - we have enormously powerful AI with insufficient compute and expertise to apply it to everything we might want to. We have research being done on alternate architectures, and optimization being done on transformers that are rapidly reducing the cost/performance ratio. There are models that you can run on phones that would have been considered AGI 10 years ago, and there doesn't seem to be any fundamental principle decreasing the rate of improvement yet. If alternate architectures like RWKV get funded, there might be several orders of magnitude improvement with relatively little disruption to production model behaviors, but other architectures like text diffusion could obsolete a lot of the ecosystem being built up around LLMs right now.

There are a million little considerations pumping transformer LLMs right now because they work and there's every reason to expect them to continue improving in performance and value for at least a decade. There aren't enough researchers and there's not enough compute to saturate the industry.

Fair point, I guess my question is how long it would take them to train GPT-2 on the absolute bleedingest generation of Nvidia chips vs what they had in 2019, with the budget they have to blow on Nvidia supercomputers today.