The most obvious reason is costs - if it costs many millions to train foundation models, they don't have a ton of experiments sitting around on a shelf waiting to be used. They may only get 1 shot at the base-model training. Sure productization isn't instant, but no one is throwing out that investment or delaying it longer than necessary. I cannot fathom that you can train an LLM at like 1% size/tokens/parameters to experiment on hyper parameters, architecture, etc and have a strong idea on end-performance or marketability.

Additionally, I've been part of many product launches - both hyped up big-news-events and unheard of flops. Every time, I'd say that 25-50% of the product is built/polished in the mad rush between press event and launch day. For an ML Model, this might be different, but again see above point.

Sure products may be planned month/years out, but OpenAI didn't even know LLMs were going to be this big a deal in May 2022. They had GPT-2 and GPT-3 and thought they were fun toys at that time, and had an idea for a cool tech demo. I think that OpenAI (and Google, etc) are entirely living day-to-day with this tech like those of us on the outside.

What we're going to see over next year seems mostly pretty obvious - a lot of productization (tool use, history, etc), and a lot of efforts with multimodality, synthetic data, and post-training to add knowledge, reduce brittleness, and increase benchmark scores. None of which will do much to advance core intelligence.

The major short-term unknown seems to be how these companies will be attempting to improve planning/reasoning, and how successful that will be. OpenAI's Schulman just talked about post-training RL over longer (multi-reasoning steps) time horizons, and another approach is external tree-of-thoughts type scaffolding. These both seem more about maximizing what you can get out of the base model rather than fundamentally extending it's capabilities.

If you've been working on AI, you've seen everything go up and to the right for a while - who really benefits from pointing out that a slowdown is occurring? Who is incentivized to talk about how the benefits from scaling are slowing down or the publicly available internet-scale corpuses are running out? Not anyone who trains models and needs compute, I can tell you that much. And not anyone who has a financial interest in these companies either.