Artificial General Intelligence (AGI) News and Discussions

[funkervogt]

Yuli Ban,

Fair points, though keep in mind that automation will lower the costs of goods and services, so even if humans lose their jobs and get poorer, it will be easier to afford things.

[Tadasuke]

If people will be poor even though we are going to experience +20% GDP a year, then it certainly must mean some grave mistakes in govervance. But I don't think that will be the case, at least not for long. People will vote for getting some of the wealth created by machines. And technological deflation is going to make everything cheaper, so if government keeps printing money, people will be able to afford things.

[Nanotechandmorefuture]

If people will be poor even though we are going to experience +20% GDP a year, then it certainly must mean some grave mistakes in govervance. But I don't think that will be the case, at least not for long. People will vote for getting some of the wealth created by machines. And technological deflation is going to make everything cheaper, so if government keeps printing money, people will be able to afford things.

Digital currency should help bring some balance to all this. Possible actual peace and stability with this? Let's find out!

[Tadasuke]

If people will be poor even though we are going to experience +20% GDP a year, then it certainly must mean some grave mistakes in govervance. But I don't think that will be the case, at least not for long. People will vote for getting some of the wealth created by machines. And technological deflation is going to make everything cheaper, so if government keeps printing money, people will be able to afford things.

Digital currency should help bring some balance to all this. Possible actual peace and stability with this? Let's find out!

So far, cryptocurrencies have shown exceptional instability.

[Yuli Ban]

[Ozzie guy]

The impression I get is that on top of an immediate boost to AI ability e.g. making GPT-4 better THE % RATE THAT AI GETS BETTER HAS NOW BEEN INCREASED. Is this correct and will the increase to the % rate of AI improvement translate to getting AGI faster or will it only effect language models?

[peekpok]

Hey everyone. I've seen some talk recently that there is an impending slowdown in AI research. Basically the reasoning goes: we picked the low hanging fruit in terms of optimizing hardware for AI training, by doing things like reducing precision in custom silicon (google TPUs for example). But the bigger problem is that we've reached the upper limit of what people are willing to spend training models. So a lot of the progress was driven by the fact that AI was hyped up and companies were willing to spend huge money to train their models. For example, it is thought that GPT-3 took over $10 million to train. The argument is that this exponential investment in training new models wont continue, so AI will once again be limited to increasing at the same speed as general purpose hardware improvements (Moore's law).

I'm curious to hear people's opinions about this. Do you agree with that reasoning, or is there some hope that we can continue to see the blistering progress in AI research that we have enjoyed in the past decade? I certainly hope that it will continue to improve, because I had great fun playing with Stable Diffusion (an image generator) all weekend long. However, the limitations of the software definitely has me itching to get my hands on something that is a bit more capable.

[Yuli Ban]

Not a particularly major step forward, but a fantastic showcase of the new scaling laws. 17B optimized parameters vs 80B unoptimized ones for Flamingo; 17B wins overwhelmingly.

[agi]

deleted

[agi]

deleted

[agi]

deleted

[agi]

deleted

[Yuli Ban]

[wjfox]

Google offers glimpse of ultra-realistic chat tech

4 hours ago

Google has launched a UK version of an app that lets users interact with the artificial-intelligence system one of its engineers has claimed is sentient.

It is a very limited trial, with just three scenarios to choose from.

And while Google wants feedback about how its Language Model for Dialogue Applications (Lamda) performs, the app users cannot teach it any new tricks.

The company has always maintained the technology, used to power chatbots, has no independent thoughts and feelings.

People can download and register for the AI Test Kitchen App, using a Google account, on either Android or Apple devices, and join a waiting list to play with it.

https://www.bbc.co.uk/news/technology-63301146

[wjfox]

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

Impressively, at 540B scale, we show an approximately 2x computational savings rate where U-PaLM achieves the same performance as the final PaLM 540B model at around half its computational budget...
U-PaLM does much better than PaLM on some tasks or demonstrates better quality at much smaller scale (62B as opposed to 540B). Overall, we show that U-PaLM outperforms PaLM on many few-shot setups, i.e., English NLP tasks (e.g., commonsense reasoning, question answering), reasoning tasks with chain-of-thought (e.g., GSM8K), multilingual tasks (MGSM, TydiQA), MMLU and challenging BIG-Bench tasks.
We show that, with almost negligible extra computational costs and no new sources of data, we are able to substantially improve the scaling properties of large language models on downstream metrics.

[Yuli Ban]

[Yuli Ban]