@hypolite You raise fair points, and I don't think I can fully refute them. The Uber analogy is genuinely worrying, and I share your concern about what “profitability” will eventually mean for people who currently depend on LLMs as tools.

That said, I think there's a tension in your argument worth examining. You note that LLMs are still fledgling and heavily subsidized, and I agree. But that same fledgling status means we're also at the beginning of the efficiency curve, not the end. Inference costs have dropped dramatically over the past two years, and experiments like DeepSeek suggest training costs can fall significantly too. “It's expensive now” and “it will always be expensive” are two different claims, and I don't think the first settles the second.

On the redundancy point: the current setup has every major company independently training overlapping foundation models in competition with each other, which is itself an enormous waste. If every software company built its own operating system from scratch instead of sharing one, the cost of software development would be staggering. Public infrastructure like CERN works precisely because the baseline investment is shared, and innovation happens on top of it. I think a similar logic applies here, even if the path there is unclear.

But the question I keep coming back to is: what does rejection actually achieve? If the goal is to limit the harms you describe—the IP violations, the labor displacement, the toll-booth dynamic—it's not obvious to me that individuals and communities blocking scrapers meaningfully slows any of that down. The companies driving this have enough momentum and capital that our refusal mostly affects us. I'm not saying refusal is wrong; I'm genuinely uncertain. I just think the burden of demonstrating impact falls on both sides of this argument, not only on mine.

@hongminhee You might be right about the cost of training, but if the Jevon's Paradox is applicable in this case, it's likely the decrease in training cost will lead to an increase in model parameter number, resulting in a stable training cost overall, but with possibly better models. That said, at the moment the trend seems to be about building capacity rather than efficiency, which doesn't give me confidence training costs will decrease enough for a public entity to sustainably create their own LLM.

Besides, if we want to stay on par with the commercial models, do we really want a public LLM that violates every intellectual property right possible?

In this context of flagrant disregard for consent, LLM rejection is more symbolic than anything else, but it isn't useless: it sends a clear social signal which, if followed enough, could pop the AI bubble sooner than later. Then, in the ruins of the current unsustainable paradigm, we'll see what LLM form will actually endure.