Artificial Narrow Intelligence (ANI)

Artificial General Intelligence (AGI)

Artificial Super Intelligence (ASI)

1950s–1960s

1970s–1980s

Late 1980s–1990s

2000s

2010s

2020s

Future Prospects

Supervised

Is trained on a labeled dataset, where the input data is paired with corresponding output labels.

The goal is for the model to learn the mapping from inputs to outputs, allowing it to make predictions on new, unseen data.

Example: Predicting house prices based on features such as square footage, number of bedrooms, and location, using a dataset where each house's price is provided.

Unsupervised

Deals with unlabeled data, and the algorithm aims to discover patterns, relationships, or structures within the data without explicit guidance.

It explores the inherent structure of the dataset.

Example: Clustering customer data based on purchasing behavior, where the algorithm groups customers with similar buying patterns together without any predefined labels.

Reinforcement

Involves an agent interacting with an environment and learning to make decisions by receiving feedback in the form of rewards or penalties.

The agent learns to take actions that maximize cumulative rewards over time.

Example: Teaching a computer program to play a game. The program takes actions in the game environment (e.g., moving a character) and receives rewards or penalties based on its performance, learning to improve its strategy over time.

Training Steps

Data Collection: Gather a large, diverse dataset relevant to the task.

Data Preprocessing: Clean, normalize, possibly augment the data.

Model Design: Choose an appropriate neural network architecture for the task.

Weight Initialization: Initialize weights randomly.

Training Loop:

1. Pass input data through the network to obtain predictions.
2. Compute the loss by comparing predictions with true labels using a loss function.
3. Calculate gradients of the loss with respect to the weights.
4. Adjust the weights in the direction that minimally reduces the loss, typically using an optimizer like SGD, Adam, etc.

Evaluation: Assess model performance using a validation set and adjust hyperparameters as needed.

Challenges / Risks

Overfitting/Underfitting: The model learns noise and details from the training data that do not generalize, or the model is too simple to capture the underlying pattern in the data.

Improper Initialization: Poor weight initialization can lead to slow convergence or training stagnation.

Vanishing/Exploding Gradients: Gradients become too small or too large, hindering effective weight updates.

Class Imbalance: Disproportionate representation of classes can bias the model towards the majority class.

Poor Hyperparameter Choices: Incorrectly chosen hyperparameters can lead to suboptimal training results.

Inadequate Evaluation: Not using proper evaluation metrics or validation techniques can misrepresent model performance.

Computational Constraints: Large models require significant computational resources, which can limit experimentation.

Differences

Depth of Layers: Deep learning networks have many hidden layers (deep architecture), whereas regular neural networks typically have fewer hidden layers.

Feature Learning: Deep learning networks automatically learn hierarchical feature representations from raw data, while regular neural networks may require manual feature engineering.

Complexity and Abstraction: Deep learning models can capture more complex patterns and higher levels of abstraction compared to regular neural networks.

Computational Resources: Deep learning models generally require more computational power and data for training due to their complexity and depth.

Performance: On tasks involving large-scale data and complex patterns, deep learning networks often outperform regular neural networks, especially in fields like computer vision and natural language processing.

History

1998 — LeNet-5: Digit recognition, approximately 60,000 parameters.

2012 — AlexNet: ImageNet competition, image classification. 60 million parameters.

2014 — GoogleNet (Inception v1): ImageNet competition, 22 layers deep, 4 million parameters, introducing the inception module.

2015 — ResNet (Microsoft): ImageNet competition, 152 layers and introduced residual learning to ease training of networks.

2018 — BERT (Google): language representation model, 110 million parameters for base model, 340 million for large model. It was used for NLP tasks.

2019 — GPT-2 (OpenAI): large-scale unsupervised language model, 1.5 billion parameters and was used for various NLP tasks.

2020 — GPT-3 (OpenAI): autoregressive language model, had 175 billion parameters, strong performance on many NLP tasks.

2022 — PaLM (Google): language model with 540 billion parameters, trained on a high-quality text corpus, state-of-the-art few-shot learning results.

2023 — GPT-4 (OpenAI): multimodal LLM, had 1.76 trillion parameters and could take image and text as inputs, producing text outputs.

Images

ImageNet: Used for visual object recognition software research, over 14 million images and thousands of object categories.

COCO (Common Objects in Context): Features object detection, segmentation, and captioning, over 200,000 labelled images, 80 categories.

PASCAL VOC (Visual Object Classes): Images for classification, detection, and segmentation tasks with 20 different object categories.

CIFAR-10 and CIFAR-100: Two datasets with 60,000 32x32 color images in 10 and 100 classes, respectively, used for object recognition.

Cityscapes: Focuses on semantic understanding of urban street scenes, with high-quality pixel-level annotations of 5,000 images in 50 different cities.

LSUN (Large-scale Scene Understanding): Contains around one million labeled images for each of 10 scene categories and 20 object categories.

Places: A scene-centric database with more than 10 million images depicting 400+ unique scene categories.

Kinetics: A large-scale, high-quality YouTube videos.

Open Images Dataset: About ~9 million images, with image-level labels, object bounding boxes, object segmentation masks, and visual relationships.

MS COCO (Microsoft Common Objects in Context): A large-scale dataset for object detection, segmentation, and captioning.

Text

BooksCorpus and English Wikipedia: Used for BERT and its variants, providing a diverse range of topics and language styles.

Common Crawl: web crawl, used by GPT, RoBERTa, offering a broad snapshot of the internet's text.

WebText2: Utilized by OpenAI for training GPT models, compiled from web pages and designed to represent a wide array of internet text.

CC-NEWS: News articles, used in conjunction with other corpora like BooksCorpus and English Wikipedia for training models like RoBERTa.

STORIES: A subset of Common Crawl focusing on narrative content, used for training language models to understand and generate story-like text.

Reddit: Derived from Reddit posts, used for training GPT-2 and GPT-3 to understand conversational language and diverse topics.

Giga5: Part of the training dataset for XLNet, consisting of English Gigaword (5th Edition), a comprehensive archive of newswire text.

ClueWeb: Used by XLNet, web pages intended for research on information retrieval, language modeling, and web search.

Common Crawl (filtered): A refined version of the Common Crawl dataset, filtered for quality and relevance, used for training GPT-3.

GitHub code: A corpus of public GitHub repositories used by models like CODEX for understanding and generating programming code.

Reasons

Improving Model Performance: Clean data helps in training more accurate and reliable models.

Reducing Noise: Prevent model from learning incorrect patterns due to irrelevant or noisy data, may lead to better generalization to new, unseen data.

Handling Data Sparsity: Ensure that models can handle rare or unseen words or n-grams effectively.

Mitigating Bias: Reducing biases present in the training data, which is crucial for building fair and unbiased models.

Enhancing Model Robustness: Eliminating duplicates and inconsistencies contribute to the robustness of the model, making it less prone to overfitting and more stable in its predictions.

Techniques

Document/Sentence Level

• Data Filtering: Removing low-quality text entries that contain excessive spelling and grammatical errors to improve the overall quality of the dataset.
• Deduplication: Eliminating duplicate sentences or documents in the dataset to prevent the model from overfitting on repeated instances of the same or similar text.

Word Level

• Text Preprocessing: Standardizing text by converting to lowercase, removing special characters, and stemming words to their root form to reduce the complexity of the language the model needs to learn.

Handling Vocabulary

• Converts text into tokens (words, subwords, characters).
• Manages vocabulary size for model learning.
• Ensures efficient text processing and understanding.

Dealing with Out-of-Vocabulary (OOV) Words

• Uses strategies like BytePairEncoding (BPE), WordPieceEncoding, and SentencePieceEncoding.
• Breaks down rare/unseen words into sub-word units.
• Reduces OOV issues and improves generalization.

Efficient Training and Inference

• Breaks text into smaller units.
• Reduces computational complexity.
• Enables training on larger datasets and faster inference.

Capturing Semantic and Syntactic Information

• Subwords capture linguistic units (prefixes, suffixes).
• Enhances understanding of language structure and meaning.

Supporting Multilingual Models

• Language-agnostic tokenization (e.g., SentencePiece).
• Tokenizes multiple languages without specific rules.
• Essential for multilingual LLMs understanding/generating text across languages.

Consider the sentence

"I walked to the store after I left home."

Without position encoding, the model might not distinguish between "I walked to the store" and "I left home" as sequential actions, potentially confusing the order of events.

With proper position encoding, the model can understand that "left home" came before "walked to the store," preserving the narrative's temporal sequence.

Position encoding is required of Large Language Models (LLMs) and other transformer-based architectures because these models do not inherently capture the sequential order of input tokens.

Unlike Recurrent Neural Networks (RNNs) or Long Short-Term Memory networks (LSTMs), which process sequences in order and thus have an intrinsic understanding of position, transformers process input tokens in parallel.

Position encoding is necessary to provide the model with information about the order of tokens, which is crucial for understanding language where the meaning often depends on the sequence of words.

Position encoding allows the model to understand word order and the relationships between words in a sentence, which is essential for tasks like translation, question answering, and text generation.

Model pre-training in the context of Large Language Models (LLMs) involves training the model on a massive amount of (usually) unlabeled text data, typically in a self-supervised manner.

This foundational step allows LLMs to acquire fundamental language understanding capabilities, which can be useful across a wide range of language-related tasks.

1. Data Preparation: The model is exposed to a vast corpus of text data, which could include web pages, books, articles, and other forms of written language.
2. Self-Supervised Learning: The model learns by predicting parts of the input data. E.g. words in the input text are masked or hidden. This process enables the model to learn word meanings, grammar, and the ability to form coherent text.
3. Adjusting Model Weights: The model learns by gradually adjusting its weights to minimize the difference between its predictions and the actual data.
4. Iteration: Passing the entire dataset through the model multiple times (epochs).
5. Outcome: A model that has a general understanding of the language it was trained on.

Masked Language Modeling (MLM): Randomly masking out tokens in the input and training the model to predict the masked tokens based on their context.

Autoregressive Language Modeling: Training the model to predict the next token in a sequence given the previous tokens, thus learning to generate text.

Next Sentence Prediction (NSP): Given a pair of sentences, the model predicts if the second sentence is the subsequent sentence in the original document, used in early models like BERT.

Contrastive Learning: Learning representations by contrasting positive pairs (related data points) against negative pairs (unrelated data points).

Replaced Token Detection (RTD): Training a model to distinguish between original tokens and those replaced by a generator in the input sequence, as used in ELECTRA.

BooksCorpus: A collection of text from books, providing a diverse range of narratives, styles, and vocabulary.

English Wikipedia: The entirety of English Wikipedia articles, offering a wide array of knowledge across countless subjects.

Common Crawl: A massive web crawl that includes text from billions of web pages, providing a rich and varied source of language use and information.

WebText2: A dataset compiled from web pages, designed to represent a wide array of internet text.

CC-NEWS: A corpus consisting of news articles, capturing a variety of writing styles and topics from journalistic sources.

STORIES: A subset of Common Crawl focusing on narrative content, useful for understanding storytelling and narrative structures.

Reddit: A dataset derived from Reddit posts, useful for conversational language and a wide range of topics discussed by users.

Giga5: Part of the training dataset for XLNet, consisting of English Gigaword (5th Edition), a comprehensive archive of newswire text.

ClueWeb: A dataset used by XLNet, consisting of web pages intended for research on information retrieval, language modeling, and web search.

Task-Specific Data: After pre-training on a large corpus, the LLM is fine-tuned using a dataset that is specific to the desired task, which could be anything from sentiment analysis to question-answering. This dataset is usually much smaller than the pre-training corpus and contains labeled data that provides clear examples of the task.

Continued Training: During fine-tuning, the model's weights are further adjusted to minimize the loss on the new task-specific dataset. This process involves using an optimization algorithm, such as stochastic gradient descent, to update the weights based on the error the model makes in its predictions.

Preserving General Knowledge: Fine-tuning aims to preserve the broad language understanding the model has gained during pre-training while also specializing its knowledge to perform well on the specific task. This is often

a delicate balance, as too much fine-tuning can cause the model to "forget" its general capabilities (a phenomenon known as catastrophic forgetting).

Hyperparameter Adjustment: Fine-tuning often requires adjusting hyperparameters, such as the learning rate, to ensure that the model's pre-trained knowledge is not overwritten too quickly. The learning rate might be set lower than during pre-training to make smaller, more precise updates to the model's weights.

Evaluation and Iteration: The fine-tuned model is evaluated on a validation set to monitor its performance. Based on these results, the model may be fine-tuned iteratively with adjustments to hyperparameters or the training procedure to optimize its performance on the task.

The step of alignment when building Large Language Models (LLMs) is a crucial process aimed at ensuring that the model's outputs are aligned with human values, goals, and ethical standards.

Addressing Unintended Behaviors: LLMs, by default, predict the next token based on probabilities learned during training. Outputs can be untruthful, biased, or harmful. Alignment seeks to steer the model away from such behaviors towards more desirable responses.

Incorporating Human Feedback: Reinforcement Learning from Human Feedback (RLHF), where human raters evaluate the model's outputs and provide feedback. This feedback is used to adjust the model's parameters and improve its alignment with human preferences.

Fine-Tuning with Aligned Objectives: Using datasets that include instructions, positive/negative examples, or specific guidelines that reflect the desired aligned behavior. This fine-tuning process helps the model learn to generate responses that are more aligned with the provided instructions and examples.

Argmax Selection

Simplest method where the token with the highest predicted probability is selected as the next token.

Commonly used in deterministic models where the goal is to generate the most likely sequence based on the model's predictions. Example: In a language model, after computing the probabilities of all possible next tokens, the token with the highest probability is chosen.

Beam Search

Beam search maintains multiple hypotheses (beams) at each step and expands them until a complete sequence is formed. It selects the beam with the highest overall probability at the end.

Commonly used in machine translation where maintaining multiple potential sequences can improve accuracy by considering context over longer spans.

Sampling

Instead of always picking the token with the highest probability, sampling involves randomly selecting a token based on its probability distribution. This introduces variability and can help generate more diverse text.

Variations:

• Temperature Sampling: Adjusts the randomness of predictions by scaling the logits before applying softmax. A higher temperature increases randomness, while a lower temperature makes predictions more deterministic.
• Top-k Sampling: Only considers the top-k most probable tokens and samples from them, effectively narrowing down choices to a manageable subset.
• Top-p (Nucleus) Sampling: Selects from the smallest set of tokens whose cumulative probability exceeds a threshold p, allowing for dynamic adjustment based on context.

This process is essential because LLMs, such as GPT-3 or BERT, are trained to predict the next word or sequence of words based on the input they receive. The quality and structure of the prompt can significantly influence the relevance, coherence, and accuracy of the model's response.

LLMs are dependent on prompts as they serve as the interface for human-model interaction, providing context and instruction that shape the model's behavior.

Without well-designed prompts, LLMs may generate outputs that are off-topic, factually incorrect, or fail to grasp the user's intent.

Effective prompt engineering can mitigate these issues by incorporating clear instructions, examples, or structured queries that align the model's responses with desired outcomes, making it a critical skill for leveraging the full potential of LLMs in various applications.

Computer Vision

• "Generate a high-resolution image of a cat."
• "Create a new artistic rendering of a landscape."
• "Translate this sketch into a realistic image."
• "Enhance the resolution of this low-quality photograph."
• "Remove the background from this image while preserving the foreground object."
• "Generate a cartoon version of this portrait."
• "Modify the lighting conditions in this image."
• "Apply a specific artistic style to this photograph."
• "Colorize this black and white image."
• "Generate a realistic image of a car from a textual description."

Multi-modal

• "Generate a descriptive caption for this image."
• "Create a video sequence based on this textual storyline."
• "Translate this English text into a corresponding image."
• "Generate an audio description for this visual scene."
• "Produce a video with synchronized audio based on this script."
• "Generate a textual summary of this video clip."
• "Create a slideshow presentation from this text document."
• "Translate this image into a sequence of musical notes."
• "Generate a storyboard based on this audio narration."
• "Create a comic strip from this dialogue script and accompanying images."

Human-like Interaction

• Understands and generates human-like text
• Interacts with users and comprehends instructions
• Executes tasks with complex reasoning and multi-step processes

Foundation on LLMs

• Uses Large Language Models (LLMs) as a foundation
• Learns from examples and improves over time

Augmented Capabilities

• Accesses external databases and uses APIs
• Incorporates updated information for dynamic environments
• Adapts to new challenges continuously

Versatile and Intelligent Systems

• Acts as personal assistants or customer service agents
• Components of larger autonomous systems

Data Privacy and Security Concerns

One of the prime challenges is security and data privacy. Handling proprietary and sensitive data can pose serious risks.

For instance, the ChatGPT outage (March 2024) incident mentioned in the iplocation.net article underscores the vulnerabilities and privacy violations that can occur, emphasizing the need for robust security measures like encryption, access controls, and regular security audits to safeguard sensitive data.

Integration with Existing Workflows

The challenge of smoothly integrating generative AI into existing business workflows is significant.

For example, financial institutions considering the use of language model to determine fraud will likely find the integration challenging due to the difference in operational methodologies between legacy systems and generative AI technologies.

Legacy systems have a very specific way of operating, and introducing generative AI, necessitates finding new ways to either create integrations or adopt new capabilities that enable reaching the same outcomes more effectively and efficiently.

Expertise Requirements

High level of expertise is required to implement, customize, and maintain effectively, a barrier for businesses without this expertise in-house.

For instance, the limited talent pool challenge mentioned in the Neoteric article points out that the demand for engineers experienced in generative AI development is high, and the talent pool is limited.

This is further complicated by the rapid evolution of generative AI technologies, making it difficult to have extensive experience in using technology that was released only a short time ago.

Enterprise-level

At the enterprise level, Gen AI solutions are integrated into the operations of a specific business function or across multiple functions, becoming a part of the organization's core processes and systems.

This approach is designed to enhance the overall efficiency, productivity, and innovation of teams and departments.

Personal/Individual

On the personal or individual level, Gen AI solutions are adopted by individual employees to enhance their own productivity and performance in their specific roles.

This approach allows for personalized use of Gen AI tools to meet unique job requirements and personal work styles.

Characteristics

• Broad Scope: Targets improvements in processes, services, or products that impact the entire organization or significant parts of it.
• Collaborative Use: The solution is accessible, used by multiple team members, facilitating collaboration and shared benefits.
• System Integration: Requires integration with existing IT infrastructure, databases, and applications to ensure seamless operation and data flow.

Benefits

• Efficiency Gains: Automates repetitive tasks and optimizes workflows, leading to significant time savings and operational efficiency at scale.
• Enhanced Decision-Making: Provides valuable insights and data analysis, supporting better-informed decision-making across departments.
• Innovation Acceleration: Enables rapid prototyping, content generation, and idea exploration, fostering a culture of innovation.

Challenges

• Complex Integration: May involve significant challenges in integrating with existing systems and ensuring compatibility and scalability.
• Change Management: Requires effective change management strategies to address resistance, train employees, and ensure adoption.
• Data Privacy and Security: Must address concerns related to data privacy and security, especially when handling sensitive information.

Characteristics

• Focused Application: Tailored to address specific tasks or challenges faced by an individual, such as content creation, data analysis, or coding.
• Personal Productivity: Directly impacts the efficiency and output of individual employees, enabling them to accomplish more in less time.
• Ease of Adoption: Often involves standalone applications or plugins that are easier to adopt and require minimal integration efforts.

Benefits

• Customization: Allows for customization and personalization of the Gen AI tool to fit the individual's needs and preferences.
• Skill Enhancement: Empowers employees to enhance their capabilities and tackle more complex tasks with AI assistance.
• Immediate Impact: Provides immediate productivity gains and benefits to the individual, contributing to job satisfaction and performance.

Challenges

• Varied User Experience: The effectiveness can vary significantly between individuals based on their ability to leverage the tool.
• Lack of Standardization: May lead to inconsistencies in work quality or processes if different tools are used by employees within the same team.
• Knowledge Gaps: Requires individuals to have or develop an understanding of how to effectively use Gen AI tools for their specific needs.

Agents at scale

Multi-agent simulations, social agents, digital twins.

Neuro-symbolic

Combine logical reasoning, problem solving and rule-based systems with the “flexibility” and scalability of neural models.

Structured knowledge

Graph databases, Graph Neural Networks, reasoning.

Multimodal AI Models

The emergence of multimodal AI models that can understand and generate content across different forms of media, such as text, images, audio, and video, is a significant trend. These models are expected to become more intuitive and dynamic, allowing for more natural interactions with AI systems.

Small Language Models (SLMs)

While large language models have been the focus, there is a growing trend towards developing powerful SLMs that require less computational power and can be fine-tuned for specific tasks or industries, meeting legal and regulatory requirements.

Autonomous Agents

The development of autonomous agents, which are software programs designed to accomplish specific objectives without human intervention, is on the rise. These agents are expected to improve customer experiences by providing highly contextualized interactions in various sectors.

Open Models vs. Proprietary Models

There is an ongoing debate and development regarding open models, which are expected to become comparable to proprietary models in terms of capabilities. This trend is likely to democratize access to generative AI technologies.

Education

Generative AI is being used to create personalized lesson plans and learning materials that cater to individual student needs and learning styles. AI-powered systems analyze student data to generate customized curriculum.

AI tools assist educators in creating comprehensive and customized course materials, including lesson plans, lecture notes, and educational content, saving significant time and effort.

Generative AI is used for automating grading, generating quizzes, and providing feedback, as well as creating virtual simulations for immersive learning experiences.

AI-based tutoring systems offer clarification on various topics and recommend study plan adjustments, providing interactive learning content and instant feedback.

Research

Tools like ResearchRabbit and Genei are being used to notify researchers about new publications, visualize networks of papers, and summarize existing literature. These tools help researchers identify relevant studies and key insights, particularly useful for comprehensive literature reviews.

Generative AI is being employed to draft and revise academic papers, submissions, and presentations. Grammarly and Microsoft's Editor and Designer assist in writing, while other tools help aggregate and organize literature.

Researchers are encouraged to disclose the use of AI in their work, and professional organizations and publishers are releasing guidelines for the acceptable use of AI during publication and review.