Hacker News Reader

Comment by alexjackhughes | Hacker News Reader

Apologies friends, this might be a really dumb idea for lots of different reasons, but sometimes you have to put even dumb ideas into the wild.

The basic premise is this: 1. We provide a dictionary of Linear A 1. We provide a bunch (ideally all) of the Linear A sentences to GPT 2. It attempts to guess the missing words in a sentence, given the context of the time, and the words it does know 3. GPT reviews, and approves/disproves the translation 4. The process repeats

What am I missing? Given infinite time/compute, isn't it possible GPT could guess the correct translation? And we'll know it's correct when most of the sentences sort of sound correct?

Forks, thoughts, and death threats are appreciated. But ideally in that order.

View on HN

> And we'll know it's correct when most of the sentences sort of sound correct?

Explain this in more detail?

This is like saying that we know code is correct (bug free) because most of the lines sort of look correct. It's not /inherently/ an untrue statement, but it doesn't match up with any sort of real world experience.

To know that something is correct in a useful sense, you need it to not only have no obvious inconsistencies and errors, you also need a way to exclude non-obvious inconsistencies and errors. What's the plan?

> This is like saying that we know code is correct (bug free) because most of the lines sort of look correct. It's not /inherently/ an untrue statement, but it doesn't match up with any sort of real world experience.

That's actually a really good analogy. I'm thinking so long as the code "builds", then it's likely that the code can execute, and at that point, the code is worth reviewing.

There's only around 600 characters/words to guess across about 700 tablets. Let's say there's about 2 sentences per tablet, so roughly around 1400 sentences to test against the guessed dictionary.

If we simply loop through, iteratively building up this dictionary, making small adjustments and happen to get to something like 80% of the sentences passing the review (i.e. not nonsense sentences, saying something with meaning, thew meaning makes sense in the context of the period etc). Then our word predictions are likely close to the original.

Like what are the chances you can choose the wrong word across lets say 200 sentences, and still maintain meaning in all of those sentences? And then multiply that guessing all the words?

> Like what are the chances you can choose the wrong word across lets say 200 sentences, and still maintain meaning in all of those sentences?

See Harry Potter with s/wand/wang/g, for example.

> See Harry Potter with s/wand/wang/g, for example.

Honestly, pretty similar meaning tbh haha

I think it would be hard to do that across lots of words, especially when we know some of the words and the context. For example, they probably aren't talking about counting wangs in these agricultural audits.

Reply on news.ycombinator.com