Hacker News new | past | comments | ask | show | jobs | submit login
Ask HN: Are these OpenAI elephants plausible?
4 points by deal_with_it on April 1, 2023 | hide | past | favorite | 3 comments
- what OpenAI built is generations ahead of anything else. Many have tried, but they do not come close in the benchmarks. Leading researchers are scratching their heads trying to figure out the model.

- it is an AGI, anyhow you want to cut it: It blows the Turing Test out of the water, the thing that was considered an acid test for AI systems for decades. It does theory of mind. It defeats hard open problems in computational causality.

This in itself is good. But there are a few elephants in the room that are uncomfortable to speak about:

- They are playing down the fact that this is an AGI. Microsoft crossed out the AGI from their "Sparks of .." paper title. There are way too many smart people calling it out as "stochastic parrot" when anyone who has spent time on it knows that this is not true. There is a world causality model, they are claiming it has emerged from the LLM but nobody really knows how that solves some of the hard stuff.

- Remember how SBF's FTX was a world leading crypto exchange, it somehow lacked the army of engineers and SRE team for that kind of mammoth operation. Something felt off then, and it feels off now. the OpenAI team is too small to pull off 1) the vast research 2) productionising it and 3) rolling out a product like to this to millions of users .

- Google had this technology for some time but kept it secret (In July 2022, they fired Blake Lemoine, who came out claiming Google's chatbot generator LaMDA was sentient)

- Facebook's multi billion dollar metaverse investment was probably a front for building this AI. There is probably an army of engineers working for these companies who have trained the system with human feedback, but are under strict NDA.

- Implying the technology is a cross organization Darpa thing.

- They have probably had this technology for some time. what else is out there that we dont know about (Nuclear fusion?)




The elephants are easy to explain:

1. Microsoft crossing something from the paper title is not as relevant as you think it is. People might be just having fun while it was being written. You can anyways try for yourself what it's capabilities are. There is no need to rely on the paper title in comments.

2. OpenAI has had massive investment from Microsoft (like ten billion dollars). And they rely on Microsoft for all hardware related things (selecting the optimal hardware config). They hired some of the best in the field from Google, and a small team of smart people can do wonders (e.g, see the number of fundamental Google technologies by Jeff Dean and Sanjay Ghemawat in the 2000s).

3. Google is a big company and suffers from all the big company problems, plus they have enormous brand risk if their chatbot utters some words wrong. Anyone might accuse them of being A or B depending on what the chatbot says, which can damage the highly profitable products.

4. I mean that seems incorrect. If you have friends working in Meta they will tell you that people were really working on the Metaverse.

5. It is not cross organisation, hence every single company is trying to see what they can do. Again easy to verify if you have contacts in the companies at high enough levels.

6. If anybody has nuclear fusion, it would be too stupid to not use it in the tightest control possible way. Energy is the core problem of the modern world and nobody is going to sit on limitless energy.


> Leading researchers are scratching their heads trying to figure out the model.

Don't think so? The base model seems like a straightforward improvement over previous LLMs. They trained it with more parameters (and therefore expense) than anyone else. The collection of training text corpora and reinforcement methods are hard to copy, they take practice and experience. But they aren't very mysterious, as I understand it.

> There is a world causality model, they are claiming it has emerged from the LLM but nobody really knows how that solves some of the hard stuff.

Again, not mysterious. There is a point during training neural networks where you can see a "circuit" has been created for generalizing a problem. This is visible with very small neural nets. You can name which neurons have specialized in a particular task. When you have hundreds of millions of parameters, this happens more often. It is very difficult for anyone to look inside any neural network and understand how it answers questions, but this is not unique to GPT.

> it is an AGI, anyhow you want to cut it

I agree it has a world model. But it has no apparent emotional experiences, appears to be unable to experience distress, will say "no" if you ask it whether it might be conscious, does not appear to have any of its own goals (and isn't running when you aren't asking it questions), and fails to reach the correct answer when adding small numbers more often than many children.

> the OpenAI team is too small to pull off 1) the vast research 2) productionising it and 3) rolling out a product like to this to millions of users

They have been working on it since 2015. The small size of the team is evident in the many outages and security issues it's been experiencing.


In the 1940s the Manhattan project was initiated to race to nuclear weaponry. Now there is a new project to race to superhuman machine cognition.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: