# summary of a useful concept from the [[AI Assertion]] paper The text you provided discusses a concept closely related to what might be termed 'representational function', specifically using the term '[[descriptive function]]'. Section 2 of the paper is dedicated to explaining this concept and its relevance to whether AI systems can perform assertions. [[The authors]] argue that for an output (like a sentence produced by an AI) to be considered an assertion, it must first meet the requirement of having a [[descriptive function]]. They position assertion as a specific kind of activity within the broader category of descriptive representation. Descriptive representations are defined as outputs that aim to state how things are, making them evaluable as true or false (or accurate/inaccurate), in contrast to other types of representations like directives (which tell someone what to do). The core idea presented is that descriptive representations are identified by their specific function or purpose. The text defines this function as follows: > For an output of some system to have a [[descriptive function]] is for it to have the function of conveying information to an observer or consumer system, so as to cause the consumer to behave as though some condition holds. The text elaborates that this function can arise from various sources, including [[natural selection]] (e.g., a firefly's flash signalling its location), intentional design (e.g., a thermometer displaying temperature), human intentions (e.g., a person telling someone a fact to influence their actions), or [[the training]] processes of AI models. The purpose of conveying information via a [[descriptive function]] is typically to influence the recipient's behaviour based on [[the state]] of affairs represented. Outputs with this function are assessable for accuracy because there is a specific condition they are meant to correspond to. [[The authors]] apply this concept to AI systems. They argue that pre-trained large [[language models]] (LLMs), which are trained primarily to predict the next word in a sequence, produce outputs that lack a [[descriptive function]]. Their outputs, like `(3) A ‘pod’ is a group of whales`, are generated because they are statistically likely continuations of text, not because the system was trained with the function of accurately conveying information about the world to a user. While such outputs might happen to be true, this is considered a byproduct of [[the training]] data rather than the function of the output itself. Furthermore, the text argues that possessing a [[descriptive function]] is a necessary, but [[not sufficient]], condition for assertion. Systems like thermometers or specialised AI classifiers (e.g., one identifying skin lesions) produce outputs that do have descriptive functions – they are designed to convey specific information accurately. However, the authors contend that these systems still do not perform assertions because they function more like inflexible measuring devices. Their outputs are generated through calibration or supervised training based on specific inputs, lacking the further characteristics (specifically, sanctionability within a social practice) that distinguish assertion. The paper concludes that equating assertion merely with the production of outputs having a descriptive function would broaden the concept too much, collapsing it into the general category of descriptive representation. # Nick "The authors apply this concept to AI systems. They argue that pre-trained large language models (LLMs), which are trained primarily to predict the next word in a sequence, produce outputs that lack a descriptive function. Their outputs, like (3) A ‘pod’ is a group of whales, are generated because they are statistically likely continuations of text, not because the system was trained with the function of accurately conveying information about the world to a user. While such outputs might happen to be true, this is considered a byproduct of the training data rather than the function of the output itself." I feel like you've missed some really important information that comes just after this. We'll release around this section about when we can link up LLMs as producing output with descriptive functions. Surely this is part of the text as well. # gemini You are correct to point this out. My previous response focused primarily on the argument that pre-trained LLMs lack descriptive functions. The text does indeed draw a distinction between these pre-trained models and models that have undergone further specific training, known as fine-tuning. Following the discussion of pre-trained models like PaLM, the text transitions to considering more advanced systems, specifically LaMDA and Sparrow, in Section 3. It explicitly states that these systems, having been fine-tuned with objectives like 'groundedness' or 'correctness', are different. The authors suggest that this fine-tuning process likely endows the outputs of these systems with descriptive functions. For instance, regarding LaMDA, which was fine-tuned partly by checking its outputs against known information sources, the text notes: > The aim of the LaMDA project was partly to build an LLM which would produce outputs that users could trust, and the fine-tuning process involved selection for features making true outputs more likely. So LaMDA does produce outputs with descriptive functions, such as (2), the output about Mt. Everest which we quoted above. Similarly, Sparrow was fine-tuned using reinforcement learning from human feedback (RLHF), where human raters judged outputs based on criteria including correctness. This process rewards the model for producing outputs deemed correct (and helpful, harmless), thereby selecting for outputs that accurately convey information. Therefore, the text argues that while the _initial_ training of LLMs (pre-training) does not establish a descriptive function for their outputs, subsequent _fine-tuning_ aimed specifically at accuracy, correctness, or grounding in reliable sources can result in systems whose outputs do possess such a function. These fine-tuned models are designed not just to produce plausible text, but text that aims to be informative and reliable, fulfilling the criterion of having a descriptive function. However, the paper then proceeds to argue that even these systems fail to meet the _second_ condition for assertion, namely sanctionability. Sources and related content [ Your saved info Info that you asked Gemini to save ](/saved-info)