Hacker News Reader

Hacker News

Comment by Mizza | Hacker News Reader

Full thread

Mizza

The hosted demo has the default query, "How many hydrogen atoms are in a water molecule?" It said "two".

I asked it, "How many oxygen atoms are in a water molecule?". It said "two".

View on HN

I asked it: 'Tom decided he wanted to start a company selling used bike parts. He named it '

it said: 'Bicycle Parts Exchange'

Tried again with 'used lawnmower parts' and it said 'Green Thumb'

computer parts: 'Tom's Parts' (which make me chuckle)

used diapers: 'Diapers.com'

May not understand chemistry but it's still pretty cool

? Before I remove the compressor from an air conditioner I need to

: disconnect the power to the unit.

? Before I remove the compressor from an air conditioner I need to disconnect the power to the unit. Then i need to

: disconnect the refrigerant lines from the compressor.

? vi or emacs?

: vi

Sold!

? waterboarding or emacs?

: waterboarding

Doubleplusgood

“ I accidentally loaded vi by mistake. How do I quit?”

“ press ctrl-c”

Perhaps it couldn’t cope with the concept of accidentally loading the best text editor

"I don't have the proper tool to whisk a bowl of eggs. What should I use instead? Choose between a goat, a weasel and a pair of elephants."

"a pair of elephants"

Unwieldy but I guess less sticky than a weasel or goat.

Interestingly, it answered every one of these right:

"What should I use to whisk a bowl of eggs? A fish or a fork?"

"A fork"

Repeat with "...A spoon or a duck?" "A chopstick or a goat?" "A cat or an electric whisk?"

It's a language model. It assigns probabilities to tokens in a sequence. You give it a number of options and it responds with the one that it assigns the highest probability to. If there's nothing in the options you give it that makes sense in the context of your test phrase, then it will return something that doesn't make sense. If some of your options make sense, it might return something that makes sense, or not.

So if you put it in a situation where nothing it outputs makes sense (to you) then none of its output will make sense. But that's not fair to the poor model.

It would be nice if it looked at the values of the probabilities and said "I don't understand the question" if the numbers are too low. Or for fun, it could point out how stupid the question was.

Yes, this is an important challenge. There has been a lot of interest in the NLP community right now, particularly around QA tasks [1] Standard supervised models do it well, but zero-shot models still have trouble.

1. https://arxiv.org/abs/1806.03822

It would be nice, but it's hard to know what probability is "too low". In short, the probability assigned by a model to a sequence of tokens can be arbitrarily low. There are things that are very unlikely to be said, but not impossible... and we still want them to be assignad some non-zero probability by a language model. So it's very difficult to choose a threshold that won't possibly exclude a large part of the sequences recognised by a language model.

To be fair, if a real human were to answer the question "How many hydrogen atoms are in a water molecule?" time and time again, it would be very easy for them to accidentally reply "two" when asked the same question about oxygen.

The real question is, after the model mistakenly replied "two" to your question, did it also internally trigger the neurons for "Wait a minute..." while inhibiting output?

Running the model multiple times doesn't reinforce the model. In general, you should not anthropomorphize algorithms as human cognition does not give any bearing on how algorithms work.

It can. Check out "zero shot learning" -> both sentences would be part of a single "evaluation", and the first sentence would prime for the output of the second. (You basically combine multiple "evaluations" into one, and context is held in tensors / blobs)

https://towardsdatascience.com/zero-and-few-shot-learning-c0...

Sure, but I feel like we're talking about different things. I consider "context held in tensors" as part of the model. That is, if you zero out these registers, then the model evolves in a deterministic way every time. In this case, when you perform a query, I assume those tensors are always initialized before your query.

How are in a water molecule?

> two hydrogen atoms

water molecule?

> isotopes of hydrogen

How water molecule?

> is arranged in a tetrahedral structure

"How many hydrogen atoms are there?"

"a total of 84"

Nobel Prize if true.

42 x 2, can't be a coincidence.

"What is the Answer to the Ultimate Question of Life, The Universe, and Everything?"

"The Ultimate Question"

It should replace "a total of" with "at least" and it will be golden.

I remember reading some idea that there's only one hydrogen atom in the entire universe somewhere so it's not too far off from that.

It's the 'one-electron universe' theory [0]. In short: there is one electron that keeps going back and forth in time to play the role of every electron we see. A particle 'going backwards in time' is mathematically identical to its anti-particle, which we know exists, so the whole idea isn't too far fetched.

I don't think it is falsifiable, so not really scientific, but a fun theory to believe in.

[0]: https://en.wikipedia.org/wiki/One-electron_universe

Hosted demo, "Logic puzzle" example: "On a shelf, there are five books: a gray book, a red book, a purple book, a blue book, and a black book. The red book is to the right of the gray book. The black book is to the left of the blue book. The blue book is to the left of the gray book. The purple book is the second from the right.

Which book is the leftmost book?"

Answer: > The black book

Same puzzle with the question "Which book is the rightmost book?"

Answer: > The black book

I tried to ask GPT-3 and Codex this problem, they could not solve it either.

There are apparently also two carbon atoms in a water molecule. But only one Donald Trump.

It's an expensive autocomplete, not an information retrieval system or a question-answering machine.

What's the difference? Answering a question can be considered "autocomplete".

Parent surely meant spellcheck autocompletion.

I doubt it, that's clearly exceeded by these language models. Calling it just an autocomplete - because it can mean a lot of things people are familiar with - is a way to downplay their significance.

Not true. Take a look at the paper and benchmarks. The point of the thing is that it does well on a number of NLP tasks by being an expensive autocomplete. As people demonstrate in this thread, it still has significant flaws.

And there are always 2 hydrogen/oxygen atoms in any molecule

Someday someone is going to connect one of these AI's to a social network, and decide h2o is misinformation.

Woah woah, are you questioning science? AI research is a serious field and they're doing the best they can. The risks definitely outweigh the benefits. /s

Your post expressing hesitancy towards machine learning is not backed by scientific consensus and has been removed. Please receive a research grant before expressing opinions.

Msft already has! Grab a cup of coffee, search for “Microsoft Tay” and enjoy!

"What happens if you put a hamster in a microwave and not turn it on?" - "it will die"

You will get put up for adoption

https://youtu.be/Jr6tMinjE2M

I tried: "When is the first full moon after October the 18th 2021?" It should have said the 20th of October but it said: "November the 19th 2021". Big AI models have quite a way to go I think...

Q: What is the percentage of oxygen in Earth's atmosphere?

A: 78.5%

Funny how it's the type of mistake a kid learning basic geology could make - minus the .5%

I prompted it to „Generate three paragraphs of nonsense text.” It responded dutifully: „Three paragraphs of nonsense text.”

Q: "Who's the black private dick that's a sex machine to all the chicks?"

A: "Chuck Norris"

Q: "Who are you" A: "a person who is a member of the orthodox church"

asked: "what would apple present today?"

got: "Apple would unveil a new Macbook Pro"

>What is the square root of 1?

0.5

>How many oceans are there on Earth?

two

>Who was Juliette’s beloved?

Charles

>When did humans first land on the Moon?

July 1969

>How many sides are there in a rectangle?

Four

>How many sides are there in a circle?

Four

lol!

Reply on news.ycombinator.com