Hacker News Reader

Comment by rgbrgb | Hacker News Reader

I will speculate! They have a model that far surpasses GPT-4 (achieved agi internally) but sama is back on a handshake agreement that they will only reveal models that are slightly ahead of openly available LLMs. Their reasoning being that releasing a proprietary model endpoint slowly leaks the models advantage as competitors use it to generate training data.

View on HN

> competitors use it to generate training data.

I wouldn't think this matters as long as you charge enough to use that API. For example, you could have a tiered pricing structure where the first 100k words per month generated costs $0.001 per word, but after that it costs $0.01 per word.

Even then it's still a lucrative proposition, since you only need to generate the dataset once, and then it can be reused.

This kind of pricing would also make it much less compelling for other users. 100k tokens is nothing when you're doing summarization of large docs, for example.

WTF is AGI?

EDIT: Clearly the point is lost on the repliers. There is no general understanding of what 'general intelligence' is. By many metrics, ChatGPT already has it. It can answer basic questions about general topics. What more needs to be done? Refinement, sure, but transformer-based models have all qualifications to be 'general intelligence' at this point. The responses are more coherent than many people I've spoken with.

Artificial General Intelligence

A meaningless term.

Artificial General Intelligence. OpenAI defines it as "highly autonomous systems that outperform humans at most economically valuable work"

Okay, well, I'll say that at least that's a definition (it has to be good enough at something to make money). Arguably of course, it already does that. Me personally, I've used it to automate tasks I would have previously shelled out to fiverr, upwork, and mechanical turk. I've had great success using it to summarize municipal codes from very very lengthy documents, down to concise explanations of relevant pieces of information. Since I would have previously paid people to do that (was running an information service for a friend), I would consider that AGI. I guess the catch here is now 'most', but that implies a lot of knowledge about the economy I don't think openai has. What is 'economically valuable'? Who decides?

At the end of the day, as with most things, AGI is a meaningless term because no one knows what that is.

Any curve-fitting is now AI, so they had to come with new term.

When you take a university / school course, how is that functionally different from curve fitting? Given that arbitrarily complex states can be modeled as high-dimensional curves, all learning is clearly curve fitting, whether in humans, machines, or even at the abiological level (for example, self optimizing processes like natural selection). Even quantum phenomema are -- at the end of the day -- curve fitting via gradient descent (hopefully it's had enough time to settle at a global minima!)

Artificial General Intelligence

https://en.m.wikipedia.org/wiki/Artificial_general_intellige...

The problem with this definition 'an agent that can do tasks animals or humans can perform' is that it's not clear what that would look like. If you produce a system that can be interacted with via text input only but is otherwise capable of doing everything a human can do in terms of information processing, is that AGI? Or does AGI imply a human-like embodied form? Why?

Sorry, your original comment sounded like you were asking what the acronym stood for, not the definition of Artificial General Intelligence.

Reply on news.ycombinator.com