> Richard, Jerry, and Robert are going to share 60 cherries. If Robert has 30 cherries, and has 10 more than Richard, how many more cherries does Robert have than Jerry?
> answer:
> Robert has 30 + 10 = 40 cherries.
> If there are 60 cherries to be shared, then Richard and Jerry will have 60 - 40 = 20 cherries each.
> Robert has 40 - 20 = 20 more cherries than Jerry.
Um, the answer is "correct" but isn't the actual reasoning wrong?
Looks like it randomly applies operations and reasonings rather than read the text. This sentence for example makes no sense and shows this AI has no understanding of numbers whatsoever, not even first grade level understanding:
> If there are 60 cherries to be shared, then Richard and Jerry will have 60 - 40 = 20 cherries each.
> Richard, Jerry, and Robert are going to share 60 cherries. If Robert has 30 cherries, and has 10 more than Richard, how many more cherries does Robert have than Jerry?
"and" here is kind of inexact, it implies a sum, so something else. "If Robert has 30 cherries, 10 more than Richard" would be better.
It's not inexact, though. "and" in this case is a logical joiner not a numerical joiner. "Richard, Jerry, and Robert are going to share 60 cherries" is the first fact presented. "Robert has 30 cherries" is the second fact, it is one property about the cherries Robert has. "and has 10 more than Richard" is a third fact, it is another property of the cherries Robert has. The only addition that comes out of this is from the "10 more than Richard" bit, "more than" suggests addition, "and" does not. The way kids are taught to transform that would be something like:
richard + jerry + robert = 60
robert = 30
robert = richard + 10
Trying to make Robert have 40 cherries makes the math conducted by the "AI" even more absurd, because it throws out the first fact (that there are 60 total).
It’s frustrating how myopic these papers can be. It seems like the goal of the paper is to solely work within the GPT framework to test the theory of verifiers. Why not try verifiers out with other models? Perhaps it’s not a fair comparison but I remember a Kaggle competition [0] from six years ago which involved building models to solve grade school science multiple choice questions. A simple word2vec model already could achieve 50% accuracy. Despite multiple choice being (maybe?) easier than free response, I’m just skeptical that the way to solve these problems is to throw billions of weights at them. It’s also not convincing to me that this new dataset doesn’t suffer from a much smaller template space, in that the models still just memorize templates.
There is a 1-1 correspondence between data compression and generative models. GPT-2 is a highly effective loseless data compression tool: https://bellard.org/textsynth/sms.html
Always wondered why this insight is not taught as much, especially in the context of things like dimensionality reduction...
The Hutter prize for improved compression algorithms is explicitly about the relationship between compression and intelligence. http://prize.hutter1.net/
I think the actual guessing space for these free response problems is much smaller, through simple priors over the question. For example:
“Richard, Jerry, and Robert are going to share 60 cherries. If Robert has 30 cherries, and has 10 more than Richard, how many more cherries does Robert have than Jerry?”
A rudimentary model will likely already know the answer is between 0-60.
Knowing that the answer involves addition and subtraction narrows it down to maybe 8 answers.
While SAT problems have only 4 answers, there’s usually one trick/trap answer, which I think might be be difficult for a model to not accidentally guess. The analogy I can think of is sometimes it’s better to cover up the answers first and work out a solution, to not get biased by any particular answer choice.
> answer:
> Robert has 30 + 10 = 40 cherries.
> If there are 60 cherries to be shared, then Richard and Jerry will have 60 - 40 = 20 cherries each.
> Robert has 40 - 20 = 20 more cherries than Jerry.
Um, the answer is "correct" but isn't the actual reasoning wrong?
Robert has 30
Richard has 20
Jerry has 10
Hence they split the 60 this way.