Should be afraid of GPT? Are you afraid of Excel?

What is inside a GPT model? I asked GPT-4 a few questions about its memory and algorithm. Here is the discussion.

Q1: Can we describe the ChatGPT algorithm as deep learning, backpropagation, attention mechanism based algorithm?

GPT-4:  Yes, the GPT (Generative Pretrained Transformer) family of models, including ChatGPT, can indeed be described as deep learning, backpropagation, and attention mechanism-based algorithms. Here’s how:

1. Deep Learning: GPT models are a type of deep learning model. They are transformer-based models that use multiple layers of transformers. “Deep” refers to the use of multiple layers in the model — the more layers, the deeper the model. These layers allow the model to learn and extract high-level features from the input data.

2. Backpropagation: This is the primary algorithm for performing gradient descent on neural networks. It iteratively adjusts the weights of neurons (in the transformer layers) to minimize the difference between the actual output and the desired output. This is how the model learns from the data.

3. Attention Mechanism: The transformer architecture used by GPT models includes what’s known as a self-attention mechanism (or scaled dot-product attention). This allows the models to weigh the importance of different words when generating a response. For instance, when generating a response, words that are more relevant to the context will be given more “attention” than those that are less relevant.

So, the statement is correct — ChatGPT and similar models are deep learning, backpropagation, and attention mechanism-based algorithms.

Q2: Is something missing from this description?

GPT-4: The description you provided is accurate and covers the essential elements of the GPT (Generative Pretrained Transformer) family of models, including ChatGPT. However, there are several other concepts and components that could be discussed for a more comprehensive understanding of these models. Here are a few:

1. Transformer Architecture: The transformer architecture is the fundamental structure upon which GPT models are built. This architecture is designed to handle sequential data for tasks like translation, summarization, and text generation.

2. Pretraining and Fine-tuning: GPT models are pretrained on a large corpus of text data, and then fine-tuned on a specific task. During pretraining, the models learn to predict the next word in a sentence, which helps them understand grammar, facts about the world, reasoning abilities, and some biases in the data. Fine-tuning then tailors this pretrained model to perform well on a specific task.

3. Positional Encoding: Since the transformer architecture doesn’t inherently account for the position of words in a sequence, GPT models use positional encodings to give the model information about the relative position of words in a sentence.

4. Layer Normalization: GPT models use layer normalization inside the transformer blocks to stabilize the learning process and reduce the training time.

5. Tokenization: Before processing, the input text is broken down into smaller pieces called tokens. GPT models use a specific type of tokenization called Byte-Pair Encoding (BPE) to effectively handle a wide array of languages and special characters.

Remember, even though these models have proven to be incredibly powerful tools for natural language processing tasks, they still have limitations. They generate outputs based on patterns learned during training and do not have real-world knowledge or understanding beyond their training data (which is cut off at a certain point). Furthermore, they don’t have beliefs, opinions, or emotions, and any such sentiments expressed are a reflection of the data they were trained on.

Q3:What size is the memory GPT-4 uses?

GPT-4: As of my knowledge cutoff in September 2021, OpenAI hasn’t released GPT-4 or any specific details about it, including its memory usage. The memory used by any version of the GPT model (such as GPT-3 or a potential GPT-4) would depend on a variety of factors, including the number of layers in the model, the number of neurons (or units) per layer, and the size of the model’s embeddings.

Generally, larger models require more memory. For example, GPT-3, with its 175 billion parameters, would need significant memory just to load the model weights. This does not even account for additional memory needed for operations during inference or training, such as maintaining activations for backpropagation.

To get specific numbers, you’d need to refer to documentation or announcements from OpenAI regarding GPT-4 or any other model that has been released after my last update.

Q4: Does the memory include data used during pretraining or data was used only for pretraining and then not used during a conversation?

GPT-4: The memory used by models like GPT-3 or GPT-4 does not include the data used during pretraining. The model does not store the training data in memory. Instead, it learns from the training data and adjusts its internal parameters (weights and biases) to predict the next word in a sentence.

During the pretraining phase, the model processes the training data in small chunks called batches. Each batch is loaded into memory, used to update the model’s parameters, and then discarded. Once the training phase is complete, the model retains no explicit memory of the training data.

What the model does retain is a kind of “implicit memory” in the form of learned patterns and structures in the data. This is stored in the model’s parameters. For example, the model learns syntactic and semantic rules of language, facts about the world, and to some extent, the biases present in the training data.

During a conversation or while generating text, the model does not have access to its training data. It generates responses based solely on the input it receives and the patterns it learned during training. Its understanding is limited to the patterns it learned up until its last training update and does not include any new information or events occurring after its training cut-off.

Q5: So it does not need any other memory than the words of vocabulary and its internal parameters. Is this correct?

GPT-4: That’s mostly correct. The memory usage of a model like GPT-3 or GPT-4 primarily consists of:

1. Model Parameters: These are the weights and biases that the model learns during training. These parameters are the “knowledge” that the model has acquired. For GPT-3, there are 175 billion parameters, and these would occupy a substantial amount of memory.
2. Vocabulary Tokens: These are the words, subwords, or characters that the model knows, derived from its tokenization process. Each token has an associated vector representation (an embedding), which is also part of the model’s parameters.

However, during the actual execution (whether training or inference), the model also requires additional memory for computations. This includes:

1. Intermediate Computations: For each input, the model needs to store intermediate values, such as layer activations, for the current pass through the network (whether forward or backward).
2. Input Data: The input data, whether a batch of data during training or a user’s query during a conversation, also needs to be stored in memory.

So, while the bulk of the memory is used for the model’s parameters and vocabulary, additional memory is required for ongoing computations and input data handling.

Q6: Given the above, can we say the GPT is analogous to Excel? Excel uses symbols (numbers) and rules of mathematics to compute number-based symbols. GPT uses symbols (tokens) and rules defined during pretraining to compute token-based symbols. Is this analogy correct?

GPT-4: In a very broad and abstract sense, your analogy holds some value. Both GPT and Excel take input symbols (tokens for GPT, numbers for Excel), apply a set of rules (the pretrained model parameters for GPT, mathematical operations for Excel), and generate an output.

However, the specifics of how they operate and what they are capable of are very different:

1. Nature of Symbols and Rules: Excel works with numbers and uses explicit mathematical operations as its rules. GPT, on the other hand, works with tokens that represent pieces of text and uses the patterns it learned during training as its rules. The “rules” that GPT learns are not explicit, they are encoded in the weights and biases of the model and learned from patterns in the training data.
2. Computation: Excel performs direct, deterministic computations based on user-defined formulas, while GPT generates probabilistic predictions based on learned patterns. GPT’s computations involve passing input through multiple layers of a neural network, and the output is a statistical prediction, not a deterministic calculation.
3. Capabilities: Excel is a spreadsheet software, used for things like data storage, management, and complex numerical computations. GPT is a language model, used for generating human-like text based on the input it’s given. Their areas of application are very different.

So, while the broad concept of “input symbols -> apply rules -> output symbols” is common between the two, the specifics of how this concept is realized are quite different in each case. It’s important to keep these differences in mind when making such an analogy.

* * *

Final comment. Think about it: Here  ends the discussion with GPT. But If the differences between GPT and Excel are in the nature of symbols and rules, the deterministic / probabilistic computation, and the type of software, should you be afraid of GPT? Are you afraid of Excel?

Fears about GPT technology often overlook the fact that Large Language Models (LLMs) lack the ability to generate scientific knowledge, as they possess no mechanisms for discerning reality. Science employs observation and experimentation to depict and elucidate reality, foster our understanding of the world, develop intelligence and problem-solving. Science requires questions and hypotheses that fall within the realm of language and thought, and necessitates observation, experimentation, testing, and verification. This is a continuous process of refining, expanding, and revising the understanding of reality. Currently, such capabilities are far beyond the reach of LLMs, which are merely language models devoid of any capacity to grasp reality and produce science.

Large Language Models (LLMs) like ChatGPT generate symbols based on patterns they’ve learned from the symbols they were trained on, but they do not have the ability to understand or interact with the real world. They cannot observe, experiment, or verify the associations of the symbols they have learned. While they can progress towards generating imaginative and fabricated content (fabulation in creating symbols), they cannot advance towards producing scientific knowledge, therefore to exceed human intelligence based on science.

No doubt, GPT models and technology are disruptive as they allow further automation of tertiary jobs, but like any other digital tools do not pose any existential threat to humanity.