Does this basic typology shed any light on expected progress in AI capabilities and AI alignment? Or must one reach for more sophisticated concepts in order to attain any insights?
For example, Yann LeCun claims that autoregressive models like the GPT series can’t reach AGI, because their errors necessarily accumulate. But it’s unclear to me how relevant this stricture is, once you embed a GPT model in a larger architecture which provides other types of feedback.
I’m also unclear on how these concepts relate to AIXI theory, which provided the classic idealized example of a superintelligence, and which was the PhD topic of DeepMind’s chief scientist Shane Legg. (Remember DeepMind?)
once you embed a GPT model in a larger architecture which provides other types of feedback
It seems rather urgent that we obtain a compact terminology for this.
A recent post here refers to “agentized LLMs”.
Elsewhere on the web, one may also find references to “GPT-based agents”, “LLM-based agents”, and “transformer-based agents”. (edit: and “scaffolded LLMs”)
The general idea is of an agent which consists of a language model supplying a verbal “stream of consciousness” that governs a more conventionally structured software architecture.
Transformers technically are an architecture which is completely orthogonal to training setup. However their main advantage in parallelization over the time dimension allows large training speedup and thus training over very large datasets. All of the largest datasets generally are not annotated and so permit only unsupervised training. So before transformers SL was the more dominant paradigm but foundation models are trained with UL on large internet size datasets.
Of course GPT models are pretrained with UL and then final training uses RLHF.
In the pre-GPT era of machine learning, one often heard that the basic typology of ML is supervised learning, unsupervised learning, and reinforcement learning.
The new typology for the era of “transformers” or “foundation models” seems to be autoregressive, autoencoding, and sequence-to-sequence.
Does this basic typology shed any light on expected progress in AI capabilities and AI alignment? Or must one reach for more sophisticated concepts in order to attain any insights?
For example, Yann LeCun claims that autoregressive models like the GPT series can’t reach AGI, because their errors necessarily accumulate. But it’s unclear to me how relevant this stricture is, once you embed a GPT model in a larger architecture which provides other types of feedback.
I’m also unclear on how these concepts relate to AIXI theory, which provided the classic idealized example of a superintelligence, and which was the PhD topic of DeepMind’s chief scientist Shane Legg. (Remember DeepMind?)
It seems rather urgent that we obtain a compact terminology for this.
A recent post here refers to “agentized LLMs”.
Elsewhere on the web, one may also find references to “GPT-based agents”, “LLM-based agents”, and “transformer-based agents”. (edit: and “scaffolded LLMs”)
The general idea is of an agent which consists of a language model supplying a verbal “stream of consciousness” that governs a more conventionally structured software architecture.
Transformers technically are an architecture which is completely orthogonal to training setup. However their main advantage in parallelization over the time dimension allows large training speedup and thus training over very large datasets. All of the largest datasets generally are not annotated and so permit only unsupervised training. So before transformers SL was the more dominant paradigm but foundation models are trained with UL on large internet size datasets.
Of course GPT models are pretrained with UL and then final training uses RLHF.
… and in between, instruction tuning uses SL. So they use all three paradigms.