Comment by cbg0
11 hours ago
> the most important race in human history
This feels like it needs a lot of asterisks. A race implies someone will win it - how exactly is China winning by making those models freely available? "The West" can use that tech, it's not locked away in some vault.
This is a very easy question to answer: they're commoditizing AI and compressing margins for the US labs, while their state funds large scale buildouts. It's a decelerating force for the US because it removes dollars that the labs would otherwise put into R&D.
Obviously China will not release models if they ever get far ahead of the US's best models in capability. You can't seriously think that "free models" will remain the game plan if they take an indisputable lead?
But how is it relevant to the EU right now? It seems like sitting it out and spending that money on weapons is a better option for EU money as it solves an immediate danger.
Why would we need to worry about China getting far ahead of US models given the spending trajectory of US companies? If China does get ahead, US companies will just keep pumping more money into the problem to keep the race going. But to what end? LLMs are not going to morph into AGI, that will be a new paradigm.
For most people, the best LLM is the one that is cheapest to them, as models are incredibly good. I'd wager ~90% of people using AI today wouldn't be able to differentiate GPT Sol from GLM 5.3 Flash in their day-to-day usage. The models are already good enough today for most workloads, next year's model won't change that.
> Why would we need to worry about China getting far ahead of US models
If China gets to superintelligence first, the EU is cooked. This is a regime that has permanently disappeared tens of thousands of their human rights lawyers.
> will just keep pumping more money into the problem to keep the race going
Do you think taxing the US companies at 3% of global turnover and twiddling thumbs over copyright helps or hurts the West in the race to superintelligence?
> most workloads, next year's model won't change that.
"Most workloads" are irrelevant, the models matter only at the edge of the hardest problems they can solve, you are thinking of these models like any other technology when they will soon solve incredibly difficult problems that may give one country a real edge over the other.
1 reply →