AlphaFold is made by DeepMind, an Alphabet (Google) subsidiary.

Google and OpenAI are also both developing world models.

These are a way to generate realistic environments that behave like the real world. These are core to generating the volume of synthetic training data that would allow training robotics models massively more efficient.

Instead of building an actual physical robot and having it slowly interact with the world while learning from its one physical body. The robot's builder could create a world model representation of their robot's body's physical characteristics and attach their control software to the simulation. Now the robot can train in a simulated environment. Then, you can create multiple parallel copies of that setup in order to generate training data rapidly.

It would be economically unfeasible to build 10,000 prototype robots in order to generate training data, but it is easy to see how running 10,000 different models in parallel is possible.

I think that continuing to throw billions at marginally better LLMs and generative models at this point is hurting the real innovators.

On the other hand, the billions of dollars being thrown at these companies is being used to hire machine learning specialists. The real innovators who have the knowledge and talent to work on these projects almost certainly work for one of these companies or the DoD. This demand for machine learning specialists (and their high salaries) drives students to change their major to this field and creates more innovators over time.