Comments by "DefaultFlame" (@DefaultFlame) on "TheAIGRID"
channel.
-
59
-
17
-
12
-
7
-
7
-
6
-
5
-
5
-
4
-
4
-
4
-
3
-
3
-
3
-
2
-
2
-
2
-
Wow, a model that only is trained on one form of data only understands/generates things based on that data. /s
I thought this was freaking obvious? We essentiall want to make an artificial human, but only training them on one form of data will obviously not produce a human.
Train it only on language and it will only understand the dimension of language. Train it only only on visual data and it will only understand visual data, not the physics behind that data. We can see that with more, higher quality, and more diverse data if performs better at generalizing in the limited dimension it has been trained on. So, with that knowledge we can take the next steps.
Human brains do not become full human beings by only giving it one single type of stimulus. Humans also start out with hardware that's already specialized by evolution over three and a half billion years for acting in the world and reacting to stimuli, with carefully calibrated reward and punishment mechanisms.
We need to train multimodal models in environments with as much different kinds of data as we can give it. Basically, we need to create the matrix to train AIs in it. It needs to receive visual, sound, language, and kinetic/touch data. What we are currently doing is the equivalent of building/growing the various cortexes of the human brain separately and then expecting each cortex to somehow be a full human. The human brain is a holistic, interconnected, interdependent structure, with feedback from our bodies. Humans also do not grow up in a solipsistic world, we grow up around other humans.
When we are making AGI we are trying to essentially recreate and take shortcuts to do what evolution did by random chance and through selection pressure over 3.5 billion years into just a few months or years. We have the advantage of guiding the process by intelligent design instead of evolution and natural selection, but we are still trying to create digital intelligent life.
I'm not saying that it cannot be done, I'm saying that the approaches we are currently pursuing are inherently limiting and will not result in AGI. (Not without the incredible luck of actually stumbling onto the golden BB of just the right combination of shortcuts.)
2
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
I don't think this is AGI, as in "as good at all tasks as an average human," it certainly didn't fulfill ARC's full requirements. That said, I think we've already made things that are good enough when implemented in the right framework to act as a low to average performing AGI.
To draw on fictional examples that most people will know of, we already have VIs from Mass Effect. Most decent LLMs that can be run locally meet or exceed the capabilities of VIs, frontier models blow them out of the water.
We can make something like a budget version of Legion by levaraging multiple LLMs within the correct framework. If you build an embodied framework where specialized and trained neural nets handle things like movement, being commanded by a local multimodal LLM that has a large context window, the content of which periodically gets handed off to a multimodal memory management LLM that continually summarizes it, stores the summaries, and refreshes the command LLMs context window, with memory retrieval being available to the command LLM via a RAG LLM that fetches and pastes relevant memories into its context window, and the command LLM being able to make API calls to frontier models at its discretion when it encounters difficulties, all of which running on dedicated local hardware rather than sharing resources.
This roughly mimics how humans work. Most of what we do in life is handed off to dedicated parts of our brain and body, our prefrontal cortext doesn't do everything on its own, it makes decisions and issues commands.
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1
-
1