To me the most interesting thing about ICL is that one can teach an LLM something new or correct through chatting, and the LLM learns and can apply new understanding in the next prompt as chat history. It's not so much that the ICL is using something that had been learned during training, although that surely happens, but that the LLM is learning from new context because there is reasoning going on. This learning, in turn, can be applied to the creation of synthetic training material and new memories which completes the cycle.
If they are only able to compose tasks they’ve learned at training time… can they learn various combinators? I think arbitrary computations can be expressed through compositions of a finite collection of combinators (iirc?). I suppose the detail of “how can the tasks be composed” is also important.
I'm not even sure humans "generalize" I have never been a big fan of generalized intelligence hypothesis for humans or AI. That makes me think it's even closer that we can have human-level AI, not the opposite.
Suppose you have some novel input-output function, and you provide it as a black box to a collection of humans. They are able to decide on inputs to query it with, and observe the outputs. Suppose also that you give them strong motivation to be able to predict the outputs for inputs they haven’t tried yet. I think this is a task that humans can do reasonably well at?
While we prompt in a released LLM, some of Us are exploring the capabilities of the system for generalization (generalization power) ... Once some of those some us prompters find a boundary in the system in which it can not go beyond its after-training potential ... If the system can not be re-feed and re-train with boundary condition prompts and answer space for boundary condition prompts ... we get an LLM system with generalization frontiers with difficulties for 'reasoning' at those 'knowledge frontiers' ... that's the point where we notice the biases in the data sets used to train the model and/or the zeitgeist of the human noosphere in the most exhaustive databases... it seems that implementing multimodality, embodiment, and re-feed of boundary condition exchanges into a less static' paradigm is the way to prevent 'games of language regurgitation' at boundary conditions in LLMs ... Once the system gets the potential for identifying boundary conditions exchanges and re-feed itself with those exchanges, and the potential to train itself under such scenarios ... then, we can begin to think that the system seems to be going into a path for AGI ... while the system is limited by a data set, it can play with the data, find relationships and patterns that humans had not found, mimic humans exchanges, simulate reasoning capabilities, simulate and regurgitate generalizations, etc ... but it is at boundary conditions of knowledge when/where Deep Generalizations happens ... Not all humans show GI, some of them, Yes ... an electronic/digital system capable of emulating/replicating Human Levels of GI can perform GI at non-human processing speeds ... that doesn't imply ASI ... just faster AGI ... ASI implies the effective modification of energy-matter systems ... Intelectual Generalization doesn't imply performing abilities in physical action space ... A God as a thinker is less than an inert atom nuclei vibrating in a vacuum ...
Generalization is interesting, but there may not be anything in place for a model to across layers and tokens. It's like making someone rethink what they've already interpreted. The model may like (been trained) to sound confident than have no chance at looking like it. Like you suggested it might be consuming too much to properly put things in their place for the next prediction. It could be worth taking advantage of large context lengths with this, making efficient whatever physical or spiritual phenomenon is going on at this point.
@@happypeoplearereal I Agree, I suspect that humans in the chain are yet required and today the probability of finding the 'master architectures' for integrating multimodal systems seems to have more probabilities to merge from human ingenuity rather than from the complementary machine extension ... but who knows ... maybe, there is an economic( by architecture simplicity ) paradigm for integrating multimodality, like a sandbox of narrow agents trained to integrate their outputs in the sandbox driving to an emergent ability for integrating modalities in cohesive forms ... but that simple paradigm is already - technically - difficult to implement, and a 'bet' ... Humans in the chain seem to be the less expensive but seem to require humans with a high degree of experience and performing skills, specialized knowledge, high resilience to work in multidisciplinary scenarios, a high degree of discernment for spotting their trends and cognitive biases, and faster consensus in collective intelligence scenarios ... then, we realize that the bet in the sandbox is not as expensive as the humans in the chain ... AGI becoming dependent on 2 'miracles' for happening ... one by chance ( the sandbox), the other by assuming the existence of collective intelligence in humans ... by the pieces of evidence, chance + a bit of 'humans in the chain noise' seems to be the trend to become factual... Would that drive into AGI? I don't know ... ... but I suspect that that would bring hints that AGI is a possible physical phenom ... ... no reason to jump into a hype bandwagon, no reason to jump into pessimistic and skeptical bandwagons ... ... but an irrational reason, for being superstitious, and crossing your fingers expecting the multimodal sandbox with humans in the chain would bring the hints for AGI ... ... the state-of-the-art is just testing hypothesis ... and the technology is in its infancy ... ... the PR agents are bringing money and the masses are barely aware of what is happening ...
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
To me the most interesting thing about ICL is that one can teach an LLM something new or correct through chatting, and the LLM learns and can apply new understanding in the next prompt as chat history. It's not so much that the ICL is using something that had been learned during training, although that surely happens, but that the LLM is learning from new context because there is reasoning going on. This learning, in turn, can be applied to the creation of synthetic training material and new memories which completes the cycle.
The meat of the paper was how they interpreted the hypotheses and tested them. I'll have to get a copy myself.
that's interesting, I had been wondering about something tangential to Hypothesis 2 a month back.. cool stuff!
If they are only able to compose tasks they’ve learned at training time… can they learn various combinators? I think arbitrary computations can be expressed through compositions of a finite collection of combinators (iirc?).
I suppose the detail of “how can the tasks be composed” is also important.
Isn't it very hopeful from a safety standpoint, if ICL is only capable of compositional expression?
I'm not even sure humans "generalize"
I have never been a big fan of generalized intelligence hypothesis for humans or AI.
That makes me think it's even closer that we can have human-level AI, not the opposite.
Suppose you have some novel input-output function, and you provide it as a black box to a collection of humans. They are able to decide on inputs to query it with, and observe the outputs. Suppose also that you give them strong motivation to be able to predict the outputs for inputs they haven’t tried yet.
I think this is a task that humans can do reasonably well at?
While we prompt in a released LLM, some of Us are exploring the capabilities of the system for generalization (generalization power) ...
Once some of those some us prompters find a boundary in the system in which it can not go beyond its after-training potential ...
If the system can not be re-feed and re-train with boundary condition prompts and answer space for boundary condition prompts ... we get an LLM system with generalization frontiers with difficulties for 'reasoning' at those 'knowledge frontiers' ...
that's the point where we notice the biases in the data sets used to train the model and/or the zeitgeist of the human noosphere in the most exhaustive databases...
it seems that implementing multimodality, embodiment, and re-feed of boundary condition exchanges into a less static' paradigm is the way to prevent 'games of language regurgitation' at boundary conditions in LLMs ...
Once the system gets the potential for identifying boundary conditions exchanges and re-feed itself with those exchanges, and the potential to train itself under such scenarios ...
then, we can begin to think that the system seems to be going into a path for AGI ... while the system is limited by a data set, it can play with the data, find relationships and patterns that humans had not found, mimic humans exchanges, simulate reasoning capabilities, simulate and regurgitate generalizations, etc ... but it is at boundary conditions of knowledge when/where Deep Generalizations happens ...
Not all humans show GI, some of them, Yes ... an electronic/digital system capable of emulating/replicating Human Levels of GI can perform GI at non-human processing speeds ... that doesn't imply ASI ... just faster AGI ...
ASI implies the effective modification of energy-matter systems ...
Intelectual Generalization doesn't imply performing abilities in physical action space ...
A God as a thinker is less than an inert atom nuclei vibrating in a vacuum ...
Generalization is interesting, but there may not be anything in place for a model to across layers and tokens.
It's like making someone rethink what they've already interpreted. The model may like (been trained) to sound confident than have no chance at looking like it.
Like you suggested it might be consuming too much to properly put things in their place for the next prediction.
It could be worth taking advantage of large context lengths with this, making efficient whatever physical or spiritual phenomenon is going on at this point.
@@happypeoplearereal I Agree, I suspect that humans in the chain are yet required and today the probability of finding the 'master architectures' for integrating multimodal systems seems to have more probabilities to merge from human ingenuity rather than from the complementary machine extension ...
but who knows ... maybe, there is an economic( by architecture simplicity ) paradigm for integrating multimodality, like a sandbox of narrow agents trained to integrate their outputs in the sandbox driving to an emergent ability for integrating modalities in cohesive forms ... but that simple paradigm is already - technically - difficult to implement, and a 'bet' ...
Humans in the chain seem to be the less expensive but seem to require humans with a high degree of experience and performing skills, specialized knowledge, high resilience to work in multidisciplinary scenarios, a high degree of discernment for spotting their trends and cognitive biases, and faster consensus in collective intelligence scenarios ...
then, we realize that the bet in the sandbox is not as expensive as the humans in the chain ...
AGI becoming dependent on 2 'miracles' for happening ... one by chance ( the sandbox), the other by assuming the existence of collective intelligence in humans ... by the pieces of evidence, chance + a bit of 'humans in the chain noise' seems to be the trend to become factual...
Would that drive into AGI? I don't know ...
... but I suspect that that would bring hints that AGI is a possible physical phenom ...
... no reason to jump into a hype bandwagon, no reason to jump into pessimistic and skeptical bandwagons ...
... but an irrational reason, for being superstitious, and crossing your fingers expecting the multimodal sandbox with humans in the chain would bring the hints for AGI ...
... the state-of-the-art is just testing hypothesis ... and the technology is in its infancy ...
... the PR agents are bringing money and the masses are barely aware of what is happening ...