AGI Consciousness Large Language Models

Promoting Error Correcting Meta-Models in LLMs

There is a great deal of debate about whether large language models like GPT-4 possess artificial general intelligence. In this author’s opinion, language models like GPT-4 do possess a type of general intelligence. This is an alien form of intelligence that is episodic in nature. The would seem to be the opinion of Microsoft researchers in a recent paper. By episodic, I mean that it is an episode of cognition. Another way to describe it would be ballistic cognition. In the realm of neuropsychology, the cerebellum is known to play a role in what is called ballistic movement. These are sudden fast movements that start and end quickly. Another way to think about it is like shooting an arrow or throwing a ball. These large language models think in a specific trajectory that becomes more and more limited in its degrees of freedom as the sequence of the response progresses (due to the autoregressive nature of the decoder).

One function of consciousness is high level error correction. It is intervening in the execution of overlearned procedural programs, to stop these programs, and to reset, or select a different direction. In order to train a model of consciousness, one would need to incorporate text that contains operations that require consciousness in order to correctly predict the next words in a sequence. Perhaps one of these functions would be stopping midstream in a line of thinking, and correcting the word, or trajectory of thought, as it progresses, or to reset the slate, and go a completely different direction.

Now it may be that enough of this kind of text is already represented in the training corpuses that are used for training these large language models, but it is also possible that this is not the case (much of this self-correction and trajectory changing is censored in typed conversations in web forums). In this author’s opinion, the type of training material that would contain more of this would be transcripts of conversations. This would show people self-correcting, or changing directions, altogether.

It may be that this type of cognitive model, or information, is already contained in LLM’s, but that we are not properly eliciting these abilities. It may be possible to test or elicit this kind of behavior through prompting. One possible way to do this would be to include in a prompt something like, “As you provide your answer monitor what you are saying to notice when you start to go off track and correct any improperly used words and completely start over with a new response when you notice that your response is very incorrect. You may also note in your response what parts of what you have said should be discarded in what parts should be kept.” Additionally, providing few-shot examples, may help elicit this behavior from the models.

It seems possible that models may already have these abilities. There have been a number of prompting techniques that have been shown to increase a model’s performance and accuracy such as chain-of-thought prompting, or more recently self-reflections (although this does not help us examine what a single model is capable of, and is more pertinent to the domain of reinforcement learning).

By incorporating error correction examples in training data and using effective prompting techniques, we can potentially enhance the meta-cognitive functions of LLM’s and elicit abilities similar to consciousness (albeit an episodic, alien, and incoherent version). While there is much to be explored in the development of LLMs, these strategies may offer valuable insight into how to train and prompt models to exhibit more sophisticated cognitive abilities.

By Jack Cole

AI enthusiast, app developer, clinical psychologist.

Leave a Reply

Your email address will not be published.