LLM Conversational Level

The LLM Conversational Level (LCL) of a piece of text is the size of the smallest general-purpose LLM that can correctly decode the writer's intention in the text, given that mainstream larger LLMs also concur in the meaning. Texts where the writer's intention has changed or where the writer's intention cannot be decoded by any LLM have an undefined LCL.

Conversationally, the LCL of an individual is bounded above by the LCL of a text that that individual cannot decode.

Examples

Word Salad 1

Emmett Shear @eshear

The incredibly clever trick that all frontier AI labs use which makes LLMs so fabulously powerful: they barely regularize at all, resulting in models which are massively overfit. This would be a problem, except they’re overfit to the domain of all human cultural knowledge.

Aug 30, 2025

rizz or bust @rizzorbust

Replying to @eshear

what a word salad

emmett also reminding us he was ceo of OAI for 24hrs The incredibly clever trick that all frontier AI labs use which makes LLMs so fabulously powerful: they barely regularize at all, resulting in models which are massively overfit. This would be a problem, except they’re overfit to the domain of all human cultural knowledge.

Aug 30, 2025

Gemma 3n 4b flawlessly comprehends the message, so the LCL is bounded above by 4B.

I couldn't get a smaller Gemma model at the time since OpenRouter's Gemma 2b doesn't generate and just returns 400. It is interesting to observe what happens when you use smaller models, though. And Llama 3.2 conveniently has a few available. The comparison rapidly illustrates the difference.

Repeating the experiment has the smaller model more often describe the tweet as word salad, while the larger model concludes that more rarely, though it does not often explain the tweet correctly. The largest models like Opus 4.1 understand it every time and never make a mistake. So Twitter user rizzorbust likely exceeds LCL 1B but is lower than 4B and perhaps lower than 3B.

Word Salad 2

Roshan George @arjie

Replying to @jstephencarter

Citizens *are* prioritized. They have the best folk working for them rather than just whomever they could find. This is the beauty of America: anyone can be an entrepreneur and when they are, they can hire the guy they think can do the job best.

Jan 2, 2025

Stephen Carter @jstephencarter

Replying to @arjie

Silly word salad. I can’t hire an eight-year-old. I can’t hire certain types of criminals. It’s outrageous that I can hire a non-citizen when they are American citizens capable of doing the job, and there absolutely are.

Jan 2, 2025

As an example, the latter tweet refers to the former as "silly word salad", implying that it is a "confused or unintelligible mixture of seemingly random words and phrases" (as described by Wikipedia). If an LLM were to be able to decode the text it would imply that it is not, in fact, unintelligible, and consequently it would imply that the reader who considers it word salad has an LCL bounded by that of the text.

This is a pretty big commercial model so it's not a surprise it gets it right

This Llama 3.2 is quite small, having only 3B parameters (which is still large compared to the GPT-2 1.5B model)

Based on this, we can safely conclude that the text has an LCL of at most 3 billion. We must also conclude that the person unable to comprehend the original tweet has an LCL lower than 3 billion as well.