In the 2010’s, the main way to deploy machine learning was to train and deploy your own model. This required research talent. It was slower and more expensive than developing normal software. This changed in the very late 2010’s / early 2020’s. All of a sudden, the best and fastest way to leverage AI (rebranded away from ML) was to use a pretrained foundation model via an API.
What does this mean? Suddenly, teenagers can compete with seasoned machine learning PhDs to build AI applications, and "AI Engineering" has emerged as a new discipline. Talent is now mostly split between those working 'above the API' to integrate AI into traditional software, and those working 'below the API' to build foundation models.
The same talent bifurcation will soon take place in the physical economy, thanks to embodied AI and autonomous robotics. Over the past several decades, industrial robots had to be manually programmed to repeat precise/exact tasks, and robotics was a highly specialized field. More recently, companies like Tesla, Anduril, and Waymo have built industry-specific autonomous robots using modern AI. Now, companies are working to build general-purpose robots powered by foundation models. This includes, but is not limited to humanoids.
Once embodied foundation models are ready, you will be able to buy/rent general purpose physical labor 'out-of-the-box', much like you can access general purpose intelligence 'out-of-the-box' with LLM APIs. Given this, we should expect that companies will bifurcate into robot companies and robot deployment companies, and that talent will bifurcate into roboticists and robot deployment engineers. There will be only a few large players building robots, and orders of magnitude more companies deploying robots.
Robot deployment engineers will be responsible for building software to orchestrate fleets of robots, fine-tune robot policies to an industry's domain, and build evaluations for given tasks and environments. It is not yet clear what the dominant robot interface will be - teaching by demonstration, audio instructions, programmatic access, etc. It is also not yet clear how interoperability will work with bots from different providers.
Of course, deep robotics knowledge will always be valuable. But I am excited to see the barriers to entry in the space become much lower.