crazy how wrong that leaked google memo from 3 years ago turned out to be. deepseek had a good moment but beyond that open source models have just never mattered much in practice, vast majority of spend and utility has come from the frontier this whole time

that was pretty much my contention at the time, and indeed it took over a year for even closed labs to catch up to gpt-4. the open model gap has ebbed and flowed but always remained significant

people continually predict much better much cheaper open models right around the corner, and they keep coming! but they also keep being totally irrelevant in the face of the new capability step changes every small increase in model power brings

no one's ever going back to 4.5 for coding after 4.6 releases, you can feel the difference too strongly too quickly. and i'm sure as hell not trusting glm 4.7 on my codebase. i got shit to do, i don't have time to mess around because it's local or open or privacy first

the point where this will be truly different is alignment and safety capabilities. open models will likely do much more damage than closed ones earlier.