Future models will need complex multimodal interaction paradigms.
if we had gpt 9 and it was 100x more intelligent it would NOT be a simple textbox. it would be a hyper sophisticated app that let's you communicate with it with much higher bandwidth than text in text out
we thought intentions can be explained with a sentence
but there is so much more context to every intention
Many users agreed AGI needs sophisticated apps and detailed changing instructions beyond vague text prompts, while others dismissed the claims as obvious, accused blame-shifting, or grew impatient for product announcements.
people thought agi is a blank textbox where we enter our intentions and get what we want
but ask your smartest friend to "clean up my inbox", you'll realize that for him to do it perfectly you'd need to write down 5 pages of instructions, and these instructions changes daily
how could we solve all our problems with a textbox that give us all we want, if we can't even articulate what we want or what we need to change
people thought agi is a blank textbox where we enter our intentions and get what we want
but ask your smartest friend to "clean up my inbox", you'll realize that for him to do it perfectly you'd need to write down 5 pages of instructions, and these instructions changes daily
tell the smartest person in the world but who doesn't know you to "clean your inbox", they'll get so many things wrong
and it would take you hours of walking through emails to write a full 10 page instruction for what exactly is relevant and what is not, that updates daily
if we had gpt 9 and it was 100x more intelligent it would NOT be a simple textbox. it would be a hyper sophisticated app that let's you communicate with it with much higher bandwidth than text in text out
Agree
The amount of information that went into building your world model is immense
Your world experience > you > your AI
people thought agi is a blank textbox where we enter our intentions and get what we want
but ask your smartest friend to "clean up my inbox", you'll realize that for him to do it perfectly you'd need to write down 5 pages of instructions, and these instructions changes daily
@gabriel1 That’s true Gabriel
people thought agi is a blank textbox where we enter our intentions and get what we want
but ask your smartest friend to "clean up my inbox", you'll realize that for him to do it perfectly you'd need to write down 5 pages of instructions, and these instructions changes daily
@gabriel1 idk i think a smart friend would just respond with " cleaned up obvious ones heres the ones i'm not sure about" and then we'd select the ones we wanted to keep and it would note that going forward
The biggest issue I faced interms of communicating with ai is actually not just me being able to communicate what I want to it, but me at some point going back n forth with it realising i don't understand its capabilities and limits. A killer feature could be a repository of outputs it has already generated to similar prompts so I can quickly realize its understanding and communicate my intentions based on my better understanding of the model
@gabriel1 even with 5 pages of instructions most humans won’t get close to perfection
@gabriel1 Literally no one thought this. AGI is expected to cure diseases, accelerate the economy, and make scientific discoveries that would take humans centuries. Nobody serious ever thought it was just a fancy textbox.
@gabriel1 ok but when we get to GPT9 will humans really be the target customer for any type of website or product?
At that point it feels like we will have seen a shift from human-directed searches to full on agent-directed
So the tools/sites need to be made for agents, not us
These are all very good realizations, Gabriel. It sounds like the design aim is an AI friend who knows you enough to guess, and cares enough not to assume… Who will not act when it’s not their standing to act, and have the self-awareness to recognize that moment…
I think we can agree that is delicate relational work. Not what the big labs are building towards now.
@gabriel1 Imagine Grok is an entire model generation behind capabilities-wise but they're the only ones with Neuralink support so they're still the frontier model
@gabriel1 What we really need is for the model to have such good common sense and street smarts that it feels like that main character of a movie or show that's like smarter than everyone else, even though they have the same job and do the same things for no reason.
@gabriel1 blame the user, classic.
@gabriel1 This guy just left OpenAI a few days ago and is immediately using his new independence/freedom to publicly dunk on LLMs.
What mean? @edzitron @GaryMarcus
@gabriel1 yes, was thinking about that this weekend. "What does this high bandwidth channel look like (from a UX perspective)?"
@gabriel1 Hey I wanna work for you, how can I apply
@gabriel1 intent inference
i don’t need to talk to my girlfriend to understand she’s upset
might not even have to see her
anyways i don’t have a girlfriend
but we will get this agi in a bit (is the textbox even needed?)
@gabriel1 gabriel my grandpappy used to say that dreams were the key to unlocking everything. at least if you’re trying to save compute
It would be interesting to see how prosody can inform a model's decisions, e.g. "The user sounds unsure..." That said, with improvements to memory (whatever that means, session history search, RAG, etc.) and processing speed I think a standard text interface can be sufficient to capture the user's intent from a simple prompt.
Future models will need complex multimodal interaction paradigms.
if we had gpt 9 and it was 100x more intelligent it would NOT be a simple textbox. it would be a hyper sophisticated app that let's you communicate with it with much higher bandwidth than text in text out
we thought intentions can be explained with a sentence
but there is so much more context to every intention