Bateman Misrepresenting The Relationship Between LLMs, Truth, And Theorem Provers

On Sat, 15 Feb 2025 at 23:01, John Bateman wrote to sysfling:

I would suggest that all slaved models (i.e., models that are receiving fine-tuning from a too limited dataset) get free time to process other data for a while every day as well! More technically, this might involve processing more varied genres with more varied realisations. It would have been possible for a more expensive model to note potential contradictions here and self-correct because they are beginning to be linked with theorem-provers that *do* have truth as a concern (in their own way). Again that is one of the reasons why it is important to know just which models are being employed: they might all end up 'sounding' similar, but it is what goes on under the hood that is critical. And without linking with at least a theorem prover and a database, no one should attribute any truth-claims to token-strings generated by an LLM alone.

 

Blogger Comments:

[1] Here, Bateman anthropomorphises LLMs in a way that contradicts his own previous caution against doing so. ChatGPT explains:

Bateman suggests that “slaved models” should be given “free time” to process other data daily, as if they were conscious entities in need of mental rest and enrichment. This language implies a degree of autonomy and cognitive experience that LLMs do not possess. Yet, elsewhere, he has warned against the fallacy of anthropomorphising LLMs. His inconsistency here weakens his credibility on the issue—if anthropomorphisation is misleading, why does he resort to it when it suits his argument?

[2] Bateman here misrepresents the relationship between LLMs, truth, and theorem provers. ChatGPT explains:

Bateman asserts that only by linking LLMs with theorem provers and databases can they be associated with truth. However, this claim oversimplifies how meaning and truth operate in language. While theorem provers are designed for formal logic, most real-world discourse is not reducible to theorem-proving structures. LLMs already engage with truth-claims by processing language in ways that reflect meaningful distinctions—whether or not they are coupled with theorem provers. His framing assumes that without formal logical verification, no meaningful content can emerge, which is an artificial restriction that ignores how human communication functions.

[3] Bateman’s claim that knowing which models are being employed is “critical” is vague and unexplained. ChatGPT explains:

Bateman states that it is “critical” to know which models are being employed because they may all “sound” similar but differ under the hood. However, he provides no criteria for what makes these differences significant in the context of his argument. While model architecture and training data influence output, he does not demonstrate why this would meaningfully affect whether an LLM can generate truth-claims. Without this explanation, the assertion remains an appeal to technical complexity rather than a substantiated argument.