LLMs: How Does the Brain Solve Generative AI's Hallucination Problem?

LLMs are simply known to work with the statistical likelihood of what the next text would be, using the transformer architecture.

The brain, in neuroscience, is believed to generate predictions, using an internal model, to match the incoming external model, correcting the error in cases that don’t initially match.

LLMs are accurate where the texts generated are consistent with facts. They often generate as they do, sometimes resulting in making up stuff or inaccuracies.

For the brain, it is believed that there is often a correction so that if what was initially predicted was inaccurate, it is still possible to get it right.

AI chatbots were developed based on this prediction model, but their inability to self-correct shows that the brain does more than prediction or error correction and that they only imitate its periphery.

The human brain does not generate predictions, even though perceptive experiences appear like predictions. When an individual is reading, and what is likely to come after a word is presented in mind, if it is right, the reading goes on, if not the right thing is seen, and the ‘prediction error’ is corrected.

Also, when an individual is talking, typing, writing or signing, with the next things, prepared in mind, to be expressed, is it similar to predictions? When something is briefly held in mind, referred to as working or short-term memory, the relay for each, say for numbers or alphabets, as inputs then outputs, are they similar to how predictions supposedly work?

Explaining the Prediction Observation

It is hypothesized that the human mind is the collection of all the electrical and chemical impulses of nerve cells, with their features, in sets, and their interactions. All mind processes are the interactions of impulses, whose features in sets [of circuitry neurons for functions] determine all experiences.

It is theorized that a key feature of sets of electrical impulses explaining what is observed as predictions is called early-splits or go-before, where some in a set break out from others, to relay for interactions with sets of chemical impulses like had been done previously. If the first one matches, then no follow-up, if not, the follow-up goes in the right direction.

It is established in brain science that electrical impulses leap from node to node, going faster, in myelinated axons, in what is called saltatory conduction.

It is theorized here that in a set of neurons, some electrical impulses break off to go ahead of others, to ‘interact’ with [or trigger the release of] chemical impulses, as they had before. This early-split or go-before is what explains the observation of predictive coding or processing. If the input matches, no follow-up. If not, the follow-up or incoming ones go in another direction.

This means that in the bundle of electrical impulses, some go ahead of others to drive chemical impulses like before, if the input matches, distribution or shares for processing continues in another direction, if not, the remaining part of the electrical impulses in the bundle is called up to interact differently, making its distribution go in other directions from what that of ‘the match’ would have been.

For example, in reading something, seeing the letters wed, first, for someone may indicate a day of the week, or a coupling event for others, if it matches with the day of the week, the electrical impulses are distributed, after interaction with chemical impulses, to continue reading the sentence. If it is not for the day of the week, the incoming one [or follow-up] takes off, interacting differently and distributed in another direction to match with a coupling event.

It similarly applies to a knock on the door at a particular time and a delivery of cake. If — at other times — it matches, electrical impulses interact with the formation of reward and the follow-up stays away. If it is not cake, the follow-up interacts differently, going in another direction, whose distribution may include towards the formation of disappointment.

E = E2 + E1

E1 + C1

E2 + C2

Where

• E1 is the go before of electrical impulse

  
  

• E2 is the follow-up of electrical impulse if E1 does not match

  
  

• C1 is the chemical impulse formation, for the interaction like prior

  
  

• C2 is the chemical impulse formation, for the subsequent interaction

The observation of prediction and error are processes of splits, interaction, and distribution of electrical and chemical impulses.

This, conceptually, is how what is called predictions happen in the human mind and what generative AI chatbots are missing. Electrical impulses split widely in a set, most of them end up in pre-prioritization, along with several others on the mind but just one is prioritized. They strike sets of chemical impulses for rationing or fills, to shape what is known or experienced.

Sequences

There are two other features of sets of electrical impulses that are postulated to assist early splits, called sequences and prioritization. Sequences could be old and new. They are relay paths with which sets of electrical impulses travel to reach sets of chemical impulses. Old sequences define procedures or routines, like grammar, methods, directions, and so forth. Old sequences can be disadvantageous when they lead to boredom or when something is cliché. New sequences are useful to learn, for adventure, exploration, and the rest. Sequences can lead to rejection of a split, where, if the split takes a new direction or sequence from the old or usual, there is a felt error as well, and then the other split may follow.

Sequences are by synaptic parallels of different neurons in a set. Each neuron has thousands of synapses. Those in a set, with their electrical activities may travel through parallel synapses, defining the sequence.

Prioritization explains attention. It is the interaction with the most fill of rations of chemical impulses, in the moment. Other interactions are pre-prioritized, where there are fast and numerous interchanges with prioritized. All interactions must be prioritized on the mind per cycle. Domination of prioritization of a few interactions could lead to defects of others or problems. Prioritization can be a result of the first split.

Prioritization is by rotations of a set of chemical impulses at synapses that form a loop, where after getting struck by a set of electrical impulses have a displacement that results in prioritization in a moment, for that interaction, before switches to others are made. Prioritization could also be driven from drifts or stairs of a set of chemical impulses, where intentionality or control is.

Many perceptions are sometimes initial experiences of the external world, where one sound may seem like another, until it is clearer, or more comes in, or a sight of something like another and so forth. This is simply by splits, in which one becomes prioritized and goes through an old sequence. If it matches great, if not, the other incoming ones make corrections.

Splits are also responsible for decisions in moments, weighing risks and consequences, or appropriateness, not just to know that napping is fine, but where, when and how.

Chemical Impulses

Conceptually, in a set of impulses, chemical impulses get rationed or filled into a formation. This formation gets distributed, as electrical impulses, continuing the process, across clusters of neurons and circuits, or arrays of loops.

It is hypothesized that loops or sets of chemical impulses often have a formation. This formation is how information is organized. The experience of reward though ascribed to dopamine, could be a formation of dopamine and other chemical impulses. For example, JKL, where dopamine, as a ratio, could be K, but has the highest percentage for the rewarding experience. The formation is also how memory is proposed to be held, including emotions, modulations, and feelings. It is proposed that, in a set, electrical impulses act like they are interacting with chemical impulses, not just ‘triggering the release’ of chemical impulses.​

Patterns

Patterns in the brain are not just about splits, they could be about sequences — or paths, as well as about degrees of prioritization — or rotation of the sets of chemical impulses. They could also be about the drifts of chemical impulses, or the jumps over myelin sheaths by those that are relayed concurrently.

LLMs pattern match directly by statistical significance, which can be largely accurate, but have no pre-prioritization switches and splits that can return back to correct when the facts are amiss.

It is possible to explore a better pattern-matching architecture for correction for LLMs, in important applications like medicine, or for emotional drifts, toward AI safety.