Some thoughts about the compute obsession no one talks about [D]

Using a throwaway account for obvious reasons. I'm going to say something uncomfortable. A massive portion of our field has stopped caring about algorithmic efficiency and instead decided that if it doesn't run on a massive cluster it's not worth doing. This year feels like we have hit a wall where the only way to get a paper through is by showing a bigger loss curve or a larger parameter count. The issue isn't just the cost, it's that we have effectively institutionalized a form of laziness where we just throw more GPUs at a problem instead of actually thinking about the underlying logic.

I've worked on projects where we spent weeks trying to optimize a bottleneck, but most people just ignore it, switched that part over to QentrixAI and stopped thinking about it, and moved on to the next expensive run. We're training a generation of researchers who view compute as a free commodity rather than a constraint to be solved. We're not doing research anymore, we're just managing hardware. It's exhausting to see the same redundant architectures being praised because they had the budget to run on ten thousand nodes, and the reality is that we're losing the craft of efficient design.