2 Comments

I think you are right that humanoid robots won't be common in 10 years, but you are wrong in your comments about the software problem. I asked Claude about the situation:

"Suppose a household robot who is supposed to maintain the house and cook breakfast discovers that

1. The eggs are not where they are supposed to be in the fridge

2. There is spilled wine on the carpet."

This was its reply:

"This scenario presents an interesting challenge for a household robot. Let's consider how it might approach these two unexpected situations:

Missing eggs:

The robot would likely:

Check other areas of the fridge in case the eggs were misplaced

Look in other possible storage locations (pantry, countertop)

Add eggs to a shopping list or alert the homeowners about the missing item

Consider alternatives for breakfast that don't require eggs

Spilled wine on the carpet:

The robot would prioritize this issue as it requires immediate attention to prevent staining. It would:

Assess the extent of the spill

Select appropriate cleaning materials (e.g., paper towels, carpet cleaner)

Begin the cleaning process:

a) Blot the excess liquid

b) Apply a cleaning solution

c) Continue blotting until the stain is removed

Use a fan or other method to dry the area

Mark the area for a deep clean later if necessary

The robot would need to prioritize these tasks. Cleaning the wine spill would likely take precedence over the breakfast preparation due to the time-sensitive nature of preventing a permanent stain.

Would you like me to elaborate on any aspect of how the robot might handle these situations?"

I feel like this is a reasonable response. The robot might not find the eggs, it might not do the same job you would have done at cleaning up the wine, but its responses are within the range of what I would expect a human household servant to do. Before 2018, you would have been totally right-- we had no idea how to do that. But LLMs now understand what humans would want, in the sense of understand that is important here. That's no longer the sticking point.

Expand full comment
author

I'm not sure your conclusions are warranted. We currently have LLMs that can take verbal input and give reasonable-sounding verbal responses, but that's at some remove from being able to observe its environment, determine that there *are problems*, and turn that verbal response into action.

The 1X NEO, for example, seems able to perform basic actions from verbal inputs, but so far this seems to be limited to very simple and familiar tasks in controlled environments (eg going into a bathroom, lowering the toilet seat, and leaving). That will certainly improve going forward, but I expect the gap between this and more complex, autonomous action is larger than many people expect.

There's also progress in the direction of observing and making judgments about the environment—image recognition integration to LLMs that can describe the contents of images, for example—but I expect the same applies there.

Expand full comment