3 ms·

I keep thinking...what is GPT-5 going to look like? It might be a lot like the movie "Her". And that's what, maybe five years away? Obviously they can't just k

I keep thinking...what is GPT-5 going to look like? It might be a lot like the movie "Her". And that's what, maybe five years away? Obviously they can't just keep scaling up the number of parameters or working memory size. And the way it displays new behaviors in response to specifically worded prompts is still a little weird. But this is already very close to having applications that we could only dream of. GPT-3 is definitely over hyped. But there is something real happening here. I'm still not entirely sure what to make of it.

logicchains 6y ago
>Obviously they can't just keep scaling up the number of parameters or working memory size. If they manage to scale it to just a hundred times bigger, and the qualitative increase in "intelligence" is the same as that gained by the previous 100x increase in scale, that might be good enough for "real" intelligence. And 100x scaling is certainly feasible within five years, if people keep throwing resources at it.
ThalesX 6y ago
Exactly what I was talking with a friend about yesterday. I really believe we're two orders of magnitude away from having really useful smart home assistants. The kind that 'understand' your intent. And it's a bit mind boggling that we're brute forcing understanding, but it seems to be working great. Recently I've been really into AI Dungeon with the GPT-3 model. I'll give you an example of a snippet from a story I played and it's not a singular example, plenty of insightful / amusing encounters. I'll go so far as to even say that it's more fun than interacting with a lot of people I know... > "You follow the Fool through the village. It's fairly large, but in poor condition. The buildings are all thatched roof cottages and most of them look like they're about to fall down. There are people walking around, but they all look miserable. > 'Are you happy?' the Fool turns, and asks a woman standing at an alleyway. She's pale and dressed in rag. > 'No' she responds. > 'Do you want to be happy?' the Fool asks, pulling a bright red flower from his bag. > 'Yes' she says, cautiously accepting the flower. > 'That's so sweet,' you say. (this is me writing) > 'No, it isn't," the Fool says. 'It's lazy. Happy people don't ask to be happy. They just are.'" -- I looked this quote on Google to make sure it wasn't just taken as a whole: No results found for "Happy people don't ask to be happy. They just are.". -- I was really impressed with the depth of it.
- cambalache 6y ago
  > I really believe we're two orders of magnitude away from having really useful smart home assistants. Maybe I lack imagination, but , smart home assistants to do what? I dont want an Alexa on steroids at home. The intersection of GPT-3-like-based devices that can be:1)really useful. 2)Easy to interact with.3)Respect my privacy, is empty and I suspect it would be for a long time.I need help at home with chores: prepare lunch, clean the house, do the dishes, do the laundry,paint a room,etc, no GPT-5 will be useful for that. I dont want to pay (neither in money nor with my privacy) just to talk to a device to make trivial or contrived things. A case of GPT-3-like agent as a virtual assistant for your personal/home online work is more plausible, but I think it is still more than 2 orders of magnitudes away.
- cuspycode 6y ago
  There are however google search results for "Happy people don't need to tell themselves they are happy, they just are". So maybe it transformed telling into asking, on a grammatical level?

logicchains 6y ago

>Obviously they can't just keep scaling up the number of parameters or working memory size. If they manage to scale it to just a hundred times bigger, and the qualitative increase in "intelligence" is the same as that gained by the previous 100x increase in scale, that might be good enough for "real" intelligence. And 100x scaling is certainly feasible within five years, if people keep throwing resources at it.

ThalesX 6y ago

Exactly what I was talking with a friend about yesterday. I really believe we're two orders of magnitude away from having really useful smart home assistants. The kind that 'understand' your intent. And it's a bit mind boggling that we're brute forcing understanding, but it seems to be working great. Recently I've been really into AI Dungeon with the GPT-3 model. I'll give you an example of a snippet from a story I played and it's not a singular example, plenty of insightful / amusing encounters. I'll go so far as to even say that it's more fun than interacting with a lot of people I know... > "You follow the Fool through the village. It's fairly large, but in poor condition. The buildings are all thatched roof cottages and most of them look like they're about to fall down. There are people walking around, but they all look miserable. > 'Are you happy?' the Fool turns, and asks a woman standing at an alleyway. She's pale and dressed in rag. > 'No' she responds. > 'Do you want to be happy?' the Fool asks, pulling a bright red flower from his bag. > 'Yes' she says, cautiously accepting the flower. > 'That's so sweet,' you say. (this is me writing) > 'No, it isn't," the Fool says. 'It's lazy. Happy people don't ask to be happy. They just are.'" -- I looked this quote on Google to make sure it wasn't just taken as a whole: No results found for "Happy people don't ask to be happy. They just are.". -- I was really impressed with the depth of it.

cambalache 6y ago
> I really believe we're two orders of magnitude away from having really useful smart home assistants. Maybe I lack imagination, but , smart home assistants to do what? I dont want an Alexa on steroids at home. The intersection of GPT-3-like-based devices that can be:1)really useful. 2)Easy to interact with.3)Respect my privacy, is empty and I suspect it would be for a long time.I need help at home with chores: prepare lunch, clean the house, do the dishes, do the laundry,paint a room,etc, no GPT-5 will be useful for that. I dont want to pay (neither in money nor with my privacy) just to talk to a device to make trivial or contrived things. A case of GPT-3-like agent as a virtual assistant for your personal/home online work is more plausible, but I think it is still more than 2 orders of magnitudes away.
cuspycode 6y ago
There are however google search results for "Happy people don't need to tell themselves they are happy, they just are". So maybe it transformed telling into asking, on a grammatical level?