"The Coffee Test" is an alternative to Turing test proposed by Steve Wozniak. It goes as follows:
A machine is required to enter an average American home and figure out how to make coffee: find the coffee machine, find the coffee, add water, find a mug, and brew the coffee by pushing the proper buttons. This has not yet been completed.
The test has to be repeated several times under independent control. (I.e. self-published video by Boston Dynamics is not enough.) The success rate has to be at least 50% out of at least 3 attempts in different houses.
The market will be resolved positively after a month has passed since the successful demonstration to give some time to uncover possible cheating.
Related question with a different time scale:
/MatthewBarnett/will-a-robot-be-created-that-is-cap
I do not bet on my own questions.
1,000
3.00@ProjectVictory The test is relatively well-known, so if the robot exists as more than a single prototype, I expect somebody to try it out. They might only try just one house, which will not meet the criteria, but then we can wait for somebody else to repeat the test.
@ProjectVictory If you think that such robot already exists and it fully meets the market criterias, you should bet on "Before 2025" and take a profit when it confirmed.
@bessarabov To be fair, it has to not just exist, but to be tested in a particular way. Just the existence of a robot capable of doing it is not enough for the resolution.
Test from 14 years ago:
https://www.youtube.com/watch?v=MowergwQR5Y
Some of the more modern coffee machines are quite simple to operate.
For example, does the above require the ability to manage an inexpensive italian press?
The intuition behind the idea was less 'do this specific task' and more 'generically be wise about the world'.
@gpt_news_headlines Multimodal models are already generically wise about the world in this way. The obstacle is dexterity, and that's all.
@HarrisonNathan It's an interesting point. If it's an issue of dexterity than perhaps there is a test which passes that which might be more relevant.
@HarrisonNathan can you name one such model? Not trying to argue, genuinely interested.
@ProjectVictory There are no models which are wise about the world to a particularly deep resolution. It's more like, ok, I know how to make coffee, I know about the coffee pot, the filter, where it goes, I need to put coffee grounds in it, put water in, and how much, etc. Fairly high level wisdom.
I mean, try out claude and gpt4. They can practically do all of this.
What there needs to be is a dextrous robot which has deep object recognition capability and manipulation. That part is missing. There is some connective models between the LLM layer and the robot layer, but unclear what's the best approach.