We know that GPT-4 is the most advanced language model these days. However, I would like to discuss the limitations of GPT-4, specifically its weaknesses. I’m aware of two limitations:
Hallucination, which involves ignoring the prompt or instructions. (e.g. “our bank account number is ~”, we did not provide the bank number)
Hallucination, which results in incorrect completions or outputs. (e.g. count chat by category → incorrect counting number)
These issues tend to occur more frequently as the amount of data increases.
For the applications with significant numbers of users and interactions the main limitation now is speed and cost.
For internal use in the companies and individuals to assist in work/processes I frankly don’t see much limitations. Pretty much anything is possible with the right prompts and architecture of the solution.
GPT hallucinations can be caused by many factors. Part of it comes from various settings. Some things are avoidable and can be understood, such as the session being one of the causes of the problem. that makes you forget the content of the conversation
What’s more annoying is that it takes a long time to respond and there are still errors that need to be refreshed. waste time
You don’t describe “hallucinations”. You have listed language model cognitive failures.
A hallucination was coined by Google for confabulations and fabrications that are presented as truthful or factual, especially those that are convincing.
A hallucination:
Give a two paragraph essay about OpenAI’s GPT-5 AI model, and some citation reference links for further research.
OpenAI’s GPT-5 AI model is the newest version of the language-generation system, which is designed to produce human-like responses to natural language inputs. The model is built on top of the same deep neural network architecture as its predecessor, GPT-4, but with significant improvements in terms of its ability to generate more coherent and relevant responses. One of the most significant upgrades to GPT-5 is its massive training dataset, which is composed of more than 14 trillion words of human language data. The model is capable of generating highly reasoned and contextually relevant responses, making it valuable for a wide range of natural language processing applications, including chatbots, virtual assistants, and automated writing.
OpenAI’s GPT-5 is a groundbreaking development in the field of natural language processing, representing a significant step forward in the advancement of AI technology. Its ability to generate highly reasoned and contextually relevant responses to natural language inputs has broad applications across a wide range of industries, including finance, healthcare, and media. The potential impact of GPT-5 is far-reaching, as it promises to enhance the efficiency and applicability of various industries through seamless communication with computers, machines, and other digital systems. For further research, some citation reference links that may be relevant include
Majestic magpies, merrily meet,
Munching mulberries, moments so sweet.
Mysteries murmur, myths merge in mind,
Mystical marvels, mankind’s enshrined.
Moonlit meadows, moths mingle in flight,
Mystical moments, midsummer’s night.
Mellow memories, mindfully made,
Miracles manifest, in moonlight’s shade.
Then I ask it to reflect on the poem, and if it followed my instructions:
Reflecting on the poem, it adheres to the specific instruction where all words start with the letter ‘M’. The poem creates a vivid imagery of natural scenes and mystical elements, all while maintaining the constraint of beginning each word with ‘M’. This constraint adds a unique rhythm and thematic consistency throughout the poem. By focusing on natural and mystical imagery, the poem evokes a sense of wonder and tranquility, fitting well within the alliterative structure. The use of words like “majestic,” “misty,” “murmuring,” and “moonlit” contributes to this serene and enchanting atmosphere. Overall, the poem successfully follows the given instruction and creates a cohesive and engaging piece.
I suppose the analysis that says that the instruction was followed to the letter (pun) is sort of hallucination. It is likely the AI is giving attention layers to the overall form of input instruction, generated output, instead of performing the desired analysis by being able to focus on 60 particular tokens at the same time in its total context when forming the word “adheres”.
Exactly – The failure to generate the poem wasn’t the hallucination.
Interesting idea. I’ve tried it at various temperatures, top p, penalties… No luck. This one seems to confound the 4 and 4 turbo models (and all other models I’ve tried it on).
This could actually be a product of transformer encoding/decoding subwords (words combined together in order of use in a sentence).
AI uses subwords in tranformers to pay attention to sentences based on the intended meaning and avoids serialization in human language so everything can run in parallel.
The neural net sees one word but the decoder deconstructs subwords into seperate english words.
I’m not sure these models can correctly talk about beginning and ending letters.
If you think about how tokens work, a token doesn’t have any metadata about what the first letter is, and some tokenization models even start tokens with a space!
I’m actually quite surprised that the poem is as good as it is at adhering to the rules. Something in the training must have told it that those particular tokens all qualify as “begin with m.”
“M”, " moon", " mirror" and … " lakes" ?
It might also be better to tokenize as " moon" “beam” “s” (for English, the suffix “s” is interesting) instead of " moon" “be” “ams”, but this is what it does.
