Automation of most white-collar work—already baked in?

May 24, 2025 · Writing

Some senior figures at frontier labs think that even today’s machine learning algorithms are already sufficient to automate most white-collar work (without the need for major algorithmic breakthroughs or much larger models).

On their perspective, the current bottlenecks are simply:

1. Frontier labs (and independent startups) having enough staff to create training datasets, and perform fine-tuning.
2. Inference compute.

Currently, these constraints are acute.

But these lab figures think the capabilities are there, and the economics viable, even with current technology. We just need to roll it out.

I’m not close enough, or technical enough, to have a confident take on this. But, it certainly strikes me as plausible, especially after a few weeks with o3. Inside view: >1/3.

Writing

---