4 ms·

OpenAI’s language generator has no idea what it’s talking about

2 points by elsewhen 4y ago

tianqi 4y ago
I don't think this is an attack of any value. These examples are just funny, like what babies say. The fact that a baby says something funny because it lacks background knowledge is not a proof that the baby "has no idea what it's talking about".
abudabi123 4y ago
Has there been a breakthru on the "symbol grounding problem" and does the A.I. know the meaning to transport a grape from A to B undamaged and how to negotiate with all the problem sets and service providers along the way? Looking at you, Nvidia.
- tianqi 4y ago
  I, a human being, don't know how to do these either.
  - techdragon 4y ago
    Yes but as a human being, you are quite likely (not entirely guaranteed due to variation in mental capacities between people) to be able to work towards a goal if you understand the goal, if you grasp the nature of the request "transport a grape from place a to place b, without damaging it" ... you are, provided you want to obviously, like someone is paying you or something else motivating you to want to complete the task, likely to be able to work it out, to iterate and incrementally solve the problem without further directions. Problem solving like this is what AI sucks at. Worst case you can walk the grape there, if its to an island, you could learn to build a boat. If it was your life's mission to get that grape to the top of mount Everest, you could learn rope making tool making, make climbing tools and learn mountaineering and deliver that grape to the top of mount Everest... or you could think to build a list of shipping companies, packaging companies, work out how to best pack it and then who would deliver it fastest and safest, and then pay them to deliver the grape. We can make AI good at a lot of the parts of this "problem", the "sub problems" if you will... we are currently no good at getting AI to tackle big open ended tasks. The closest we have at the moment are the more impressive "AI learns to play game name here" but the game is a tiny model universe, the controls are fixed, the pathways relatively finite compared to the real world. If the game is grape shipping simulator, then the AI can mash the buttons like the proverbial million typewriters and eventually, eventually it can make progress and reinforce itself and actually deliver the grape and win the game... But that AI cannot now deliver a grape in the real world, which has no controller keyboard or mouse, the game did not have "learn mountaineering" as an option, it may not have had thousands of different companies to determine the best price from... its just not the same problem and this is the transferability issue that AI currently faces. Every single breakthrough is still a breakthrough in making the AI master a more complex rat maze, the same rat maze faster, or training it to hunt for a different kind of cheese... but thats not necessarily "smart"... we're pretty good at training rats, and no one is predicting the rats will be able to become a super-intelligent singularity within the next decade.
beardyw 4y ago
Surely it is obvious that, say moving a table, we have a mental model of the table, the door (even the words convey a standard picture) and experience of manipulating things. AI cannot have that model based on words alone. For us it starts with playing with our fingers, putting them in our mouth and so on for years and years. That is vital experience AI cannot have.

2 points by elsewhen 4y ago

I don't think this is an attack of any value. These examples are just funny, like what babies say. The fact that a baby says something funny because it lacks background knowledge is not a proof that the baby "has no idea what it's talking about".

abudabi123 4y ago

Has there been a breakthru on the "symbol grounding problem" and does the A.I. know the meaning to transport a grape from A to B undamaged and how to negotiate with all the problem sets and service providers along the way? Looking at you, Nvidia.

tianqi 4y ago
I, a human being, don't know how to do these either.
- techdragon 4y ago
  Yes but as a human being, you are quite likely (not entirely guaranteed due to variation in mental capacities between people) to be able to work towards a goal if you understand the goal, if you grasp the nature of the request "transport a grape from place a to place b, without damaging it" ... you are, provided you want to obviously, like someone is paying you or something else motivating you to want to complete the task, likely to be able to work it out, to iterate and incrementally solve the problem without further directions. Problem solving like this is what AI sucks at. Worst case you can walk the grape there, if its to an island, you could learn to build a boat. If it was your life's mission to get that grape to the top of mount Everest, you could learn rope making tool making, make climbing tools and learn mountaineering and deliver that grape to the top of mount Everest... or you could think to build a list of shipping companies, packaging companies, work out how to best pack it and then who would deliver it fastest and safest, and then pay them to deliver the grape. We can make AI good at a lot of the parts of this "problem", the "sub problems" if you will... we are currently no good at getting AI to tackle big open ended tasks. The closest we have at the moment are the more impressive "AI learns to play game name here" but the game is a tiny model universe, the controls are fixed, the pathways relatively finite compared to the real world. If the game is grape shipping simulator, then the AI can mash the buttons like the proverbial million typewriters and eventually, eventually it can make progress and reinforce itself and actually deliver the grape and win the game... But that AI cannot now deliver a grape in the real world, which has no controller keyboard or mouse, the game did not have "learn mountaineering" as an option, it may not have had thousands of different companies to determine the best price from... its just not the same problem and this is the transferability issue that AI currently faces. Every single breakthrough is still a breakthrough in making the AI master a more complex rat maze, the same rat maze faster, or training it to hunt for a different kind of cheese... but thats not necessarily "smart"... we're pretty good at training rats, and no one is predicting the rats will be able to become a super-intelligent singularity within the next decade.

beardyw 4y ago

Surely it is obvious that, say moving a table, we have a mental model of the table, the door (even the words convey a standard picture) and experience of manipulating things. AI cannot have that model based on words alone. For us it starts with playing with our fingers, putting them in our mouth and so on for years and years. That is vital experience AI cannot have.