How to think about OpenAI's rumored (and overhyped) Q* project | Hacker News Reader

How to think about OpenAI's rumored (and overhyped) Q* project

arstechnica.com

> How to think about OpenAI's rumored (and overhyped) Q* project

are rumors worth thinking about?

It's more than rumors. There's enough information out there that you can make educated guesses about what it might be, and this article seems reasonable. There's lots of reasons to be interested in future research directions that well funded AI companies are taking, and a lot of those reasons start with dollar signs.

You don't need so-called educated guesses for that. Learning to plan is not a new concept; it's a general direction in the research field, and I bet everyone is working on something related. However, it's essentially just glorified search. If you can't find a perfect answer to the problem within 1 million[0] sampled generations, then I don't think a smarter search will help you.

[0] https://storage.googleapis.com/deepmind-media/AlphaCode2/Alp...

The fact OpenAI wants to hype something right now is probably noteworthy.

What did you see that made you think "OpenAI wants to hype something right now"? I didn't see anything coming from them except Altman saying the leak was "unfortunate". I don't believe they are generating any of the Q* hype.

Yes. Keep in mind that, aside from their former board, they probably employ a good PR firm.

That doesn't mean there's a significant product behind it. A good PR firm can freeze attrition promising the next big thing any minute now... at least for a while. And when the product comes out and it's good enough but not amazing like their GPT hits -- attrition avoided. (Not that OpenAI has an attrition problem, but not for lack of trying by the other big players.)

I would go so far as to say it's evidence there isn't - GPT-4 doesn't need hype beyond word of mouth, Q* is clearly not in that situation.

Wether its worth it or not is a fair question, but one thing is for sure - people on social media, the internet in general (and actually just in real life too) LOVE talking and thinking about rumours.

Fresh content drives views drive $ + there will always be more rumors than facts = one can make more content (and therefore $) about rumors

https://hn.algolia.com/?q=rumored

Apparently, yes.

If you read the article you’ll see it’s not just about rumours but about where the future of AI may be headed in general.

When they try to reason about the origin of the “Q*” name - I bet that the “star” sign in the name is a throwback to the old famous “A*” algorithm. It would also fit the theme of pairing the reinforcement learning and state space search.

The original Q* algorithm is from 1973

https://www.sciencedirect.com/science/article/abs/pii/000437...

Yesterday:

How to think about the OpenAI Q rumors*

https://news.ycombinator.com/item?id=38564196

Is this duplication the reason why the OP is flagged? Otherwise I'm flummoxed - it seems a perfectly reasonable/normal article.

FWIW and as best I can tell, GPT-4 solves the 2nd, table-seating, problem without issues:

https://chat.openai.com/share/35ed608a-cd3b-4271-aa97-12784d...

The article links to an example of GPT-4 failing to solve it. It seems likely that it can only solve it sometimes and basically gets “lucky”.

This is interesting. When I started a new chat, first with the apples problem, then followed by the seating problem, GPT-4 gets it wrong.

https://chat.openai.com/share/1c04a8f6-2ed7-45bf-8e20-158f78...

That's because it's got a randomizer component at its core and only knows the statistically most likely thing to output next (still influenced by the random bits), not some logical reasoning that allows it to "understand" the question.

It's output is non-deterministic because the next word is sampled from the topN highest scored continuations. The algorithm itself is deterministic.

I didn’t fully read this, but superficially I don’t see a difference between this and AI explained’s recent video. Maybe it is better in text form than video for some, but it feels plagiarised given the similarities and the popularity of the AI explained channel.

I love AI Explained, but is a 200k sub channel really that influential in the industry?

I have noticed a trend with Ars posting original articles with a lot of similarities to content from smaller sources that came out a few days earlier. Its pretty common to see a rather esoteric topic make it to the front page of HN and then a similar article showing up on Ars a bit later. I assume their authors are mining HN and similar places for ideas.

Not saying that happened here or that it is ever plagiarism. Just something I noticed as a daily Ars and HN reader.

I already have access to Q* via ChatGPT. Is that an OpenAI bug or an early invitation?

What are you seeing on chatgpt which leads you to think that you have access?

I don't "think" I have access, i have access. There is a Q* tab with a specific prompt.

A tab? Can you take a screenshot? Sorry to be blunt but can you not dripfeed information?

https://ibb.co/kJJ8k74

That looks like one of the chat characters which anyone can create and share.

If you pay for my time I can show you it via Google Meet.

Tell us more. Do you notice a difference?

Regarding quality I don't know yet. If you have a specific question to compare I can give more feedback. What I know is that with other versions I can upload a zip file and its content is processed while in Q* it is not available yet.

There is a "Do not distribute" message at the end of every response but I think I can tell general things without losing the permissions to access Q*.

Assuming you are for real…Please ask some math questions. Try to determine if there are high school level calculation problems that GPT-4 gets wrong that Q* can solve.

Give me a problem that you want to test and I will feed it.

Subtract 5614 from 28104. Divide the result by 65. Then multiply by 3.

GPT-4 is giving me a close answer, but not correct. 1035.6 instead of 1038.

GPT-4 gives me the correct answer:

> The result of subtracting 5614 from 28104, dividing that result by 65, and then multiplying by 3 is 1038.

Q* gives me an incorrect answer: 1035.

The reverse you expected.

Not a bad question to ask an AGI..

Related, same day: https://news.ycombinator.com/item?id=38570867

Reply on news.ycombinator.com