| LLM AI - in Natural Language Processing | Symbolic AI in Natural Language Processing | |
|---|---|---|
| Definition | LLMs are probabilistic fluency models, and should be viewed as universal approximate knowledge sources | Based on high-level symbolic (human-readable) representations of problems |
| Learning Model | Scan, tag and identify patterns - the algorithm learns “rules” as it establishes correlations between inputs and outputs | Pre-established formula - rules are created through human intervention and then hard-coded into a static program |
| Cynefin Complexity Domain Application Scope | Chaotic and Complex | Complicated and Ordered |
| Human input | Tagging, labelling with key words | Identification of logical framework and “rules” |
| Stability | Tends towards hallucinations - requires “noise injection” | Stable |
| Accuracy | Increases with volumes of reference/learning data - larger learning sets. However, non-linear return; flawed (?) hypothesis that “size is the solution and training with even more data is the key” | Fact based origins in designing rules. Accuracy however decreases with increasing reference or comparison data - increasing complexity and noise (i.e. zooming out too far). |
| Bias | Integrates cultural artefacts, processing human generated meaning statements, that include the bias from the training data. | Inevitably reductionist assumptions in attempting to define the dynamics of any system through governing rules |
| “Meaning” model | No discerning value judgement capacity | Cognitive logic - right and wrong |
| Outcome traceability / understanding how the result came about | Not possible | Directly scrutable |
| Nature of outputs | Recycling utterances of existing content | Capable of writing new content for the knowledge base |
| Historical context | Took off in the 1990s due to higher available computing power and larger available learning data sets. | Originally lead the frontier of AI in the 1960s, often referred to as GOAI (Good Old Artificial Intelligence). Why did Symbolic AI fall away in the 80s and 90s? Because they had the wrong model of language, they didn’t have a sufficient language syntax theory. In comparison, current LLMs specifically and consciously don’t have any theory of language structure, syntax etc. |
| Facts: LLMs can only propose assertions as likely (“the odds are that…”), and in different instances might change the assertions. |
Causality: They capture correlations from text, which may or may not reflect the structure of causal reality.
Reasoning: They can capture likely alternatives but cannot identify conclusions as definitive.
Ephemera: They depend on pretrained models requiring enormous computational resources to train, resulting in a time lag in model coverage. Responses of the current version of ChatGPT are based on 2021 data.
Memories: They have no capability to learn long-term memories from interactions. •
Explanations: They cannot provide provenance information to account for their conclusions. | |