Many computer science students and new grads aspire to move into machine learning. It seems exciting and sexy. You can play a role in bringing in the incoming AGI utopia. Many strive to work in “machine learning or AI research” - a vaguely defined field that includes everything from data engineering, infrastructure and model architecture.
Here’s the reality of the future of the field:
- Scale beats all else. The best performance improvements come from increasing scale, rather than incremtal insights in novel architectures. [1] [2]
- Scale requires a lot of capital.
- That will lead to a natural dynamic where only a few suppliers will spend that capital and distribute the cost across their customers.
- These suppliers will be the only companies hiring productive “ML researchers” working on the model layer.
- Other companies that hire “ML researchers” will either be: research projects that will not end in production or companies selling you the status of being an “ML researcher” without actually offering it.
- An oversupply of talent with consolidation of demand will lead to falling salaries: the LLM providers will have significant pricing power.
- The role will eventually fall in prestige as the supply-demand dynamics flip.
A similar story played out in the past with chip designers. Chip design was a high profile role with the most famous chip designers being seen the way famous ML researchers are viewed today. Universities invested heavily into their computer hardware programs as the number of chip companies boomed and the field was in demand. As the market dynamics changed though, the universities were not able to adjust. The supply of hardware engineers grew too high and the number of significant employers decreased to the handful of major chip makers (i.e. Qualcomm, Intel, AMD, Nvidia). This was in part driven by the high capital cost required for chip manufacturing (similar to LLM training). Naturally, over time chip design has lost status amongst students and new grads and lost its reign as an extremely lucrative and high status profession.
For some engineers, chip engineering is a genuine passion. They should pursue it, while understanding the trade-offs compared to other careers such as software engineering. But many more are driven by the inertia of their university and a blind status seeking. They will be better off thinking through their personal goals and what paths are best for them. The same will be true for “AI research”.