Imagine if they just actually trained the model on a bunch of photographs of a full glass of wine, knowing of this litmus test
Remember the old internet adage that the fastest way to get a correct answer online is to post an incorrect one? I'm not entirely convinced this type of iterative gap finding and filling is really much different than natural human learning behavior.
Also as for touch, you’re going to have a hard time convincing me that the amount of data from touch rivals the amount of content on the internet or that you just learn about mistakes one example at a time.
- Airplanes dont have wings like birds but can fly. and in some ways are superior to birds. (some ways not)
- Human brains may be doing some analogue of sample augmentation which gives you some multiple more equivalent samples of data to train on per real input state of environment. This is done for ml too.
- Whether that input data is text, or embodied is sort of irrelevant to cognition in general, but may be necessary for solving problems in a particular domain. (text only vs sight vs blind)
I think you're saying exactly what I'm saying. Human brains work differently from LLMs and the OP comment that started this thread is claiming that they work very similarly. In some ways they do but there's very clear differences and while clarifying examples in the training set can improve human understanding and performance, it's pretty clear we're doing something beyond that - just from a power efficiency perspective humans consume far less energy for significantly more performance and it's pretty likely we need less training data.
to be honest i dont really care if they work the same or not. I just like that they do work and find it interesting.
i dont even think peoples brains work the same as eachother. half of people cant even visually imagine an apple.
Neural networks seem to notice and remember very small details, as if they have access to signals from early layers. Humans often miss the minor details. Theres probably a lot more signal normalization happening. That limits calorie usage and artifacts the features.
I dont think that this is necessarily a property neural networks cant have. I think it could be engineered in. For now though seems like were making a lot of progress even without efficiency constraints so nobody cares.
What is the evidence for this? We are just taking people's word for it?
> While humans learn through example, they clearly need a lot fewer examples to generalize off of and reason against.
Human brain has been developing over millenia. machines start from zero. What if this few example learning is just an emergent capbaility of any "leanring function" given enough compute and training.
Take some artisan, I'll go with a barber. The human person is not the best of the best, but still a capable barber, who can implement several styles on any head you throw at them. A client comes, describes certain style they want. The barber is not sure how to implement such a style, consults with master barber beside, that barber describes the technique required for that particular style, our barber in question comes and implements that style. Probably not perfectly as they need to train their mind-body coordination a bit, but the cut is good enough that the client is happy.
There was no traditional training with "gap finding and filling" involved. The artisan already possessed core skill and knowledge required, was filled on the particulars of their task at hand and successfully implemented the task. There was no looking at examples of finished work, no looking at example of process, no iterative learning by redoing the task a bunch of times.
So no, human learning, at least advanced human learning, is very much different from these techniques. Not that they are not impressive on their own, but let's be real here.
also we all know real people who fail to generalize, and overfit. copycats, potentially even with great skill, no creativity.
So maybe training for litmus tests isn’t the worst strategy in the absence of another entire internet of training data…