There is zero downside to America of having a frontier class Opensource/openweights multi-trillion parameter large model, and most likely a huge upside over time also to the private sector. Totally agree with @dee_bosa. The question is what would the business model look like for this? If it takes $3B - $5B to train a frontier class model that is not just a distillation of another model, then the business model for that model has to be fully sorted out for that sort of capex on a per model basis. But hard to deny that America would benefit greatly from having one. It needs a world class research team that is well hydrated with a well thought out business model to provide the right level of returns on the invested capital along with meaningful national security benefits.

The Alpha on X has grown so exponentially in the past two years. It is now the go-to platform for high-density learning. Unlike other platforms where you are drained after spending an hour and have nothing to show for, X tends to be energizing because each time you walk away smarter.

@dee_bosa, great point. This has actually been a thing for the past several months. First, public benchmarks and evals tend to be things that don’t guarantee great results. But more importantly, creating the right evals and driving the model to operate well under those conditions seems like only the normal thing to do. For example, the EZDubs team was given a very hard time pre-acquisition by @Cisco for not being too tied to public benchmarks because they wanted to make the right tradeoffs in the model, because the wrong tradeoffs could get you a rank on the benchmark but not allow you to accomplish the results. Specifically, they didn’t want to trade off latency for accuracy. But to score well on the public evals at the time, that would be a necessary trade-off. So they created their own private evals because it was available in the market was for two generic for them. This will become a more common phenomenon.