Nine Dimensions (@ninedimensions)

My assumption would be that OpenAI noticed that the chess thing was drawing a lot of interest, so optimised something (training data, fine tuning, something algorithmic) to make gpt-3.5-turbo-instruct much better at chess. Then decided that that tradeoffs of that optimisation (cost, degradation of other functions) made it not worth it for future models. That's a bit vague, but "OpenAI did it on purpose" is my main assumption. I don't know if the timing works out.