Genarative AI
Generative AI
Generative AI, or GenAI, is capable of creating new content like text, images, music, or code by learning patterns from existing data. Unlike traditional AI which relies on explicitly programmed rules, GenAI uses machine learning models to understand and reproduce the patterns within data. Through this process, GenAI can produce novel outputs that are not explicitly programmed, making them valuable for applications requiring creativity and variability.
Generative AI systems are based on artificial neural networks (ANNs), which are computing systems inspired by the human brain. An ANN consists of interconnected nodes (neurons) organized into layers: an input layer, one or more hidden layers, and an output layer. Think of a neuron as an object with a value between zero and one. This value is the neuron’s activation, and the higher the number, the more strongly it influences an output. For many GenAI models, the input value is a prompt from a user. That data is converted into numerical values that will be processed by the neuron layers.
The input layer of a feedforward neural network is made up of neurons represented by x1, x2, …xn. Each one receives a feature or attribute from the input data. In a task like text generation, each of these neurons may correspond to a specific word or token from a prompt.
The hidden layers in this network perform the core computations. Each neuron in a hidden layer receives inputs from the neurons in the previous layer and processes these inputs using a weighted sum. A weighted sum is the sum of all inputs to a neuron. Each input is multiplied by its corresponding weight, plus a bias, (a value calculated during training that adjusts the output of a neuron independently of the input values). The next step is an activation function, a function that determines the output of a neuron based on its input and then passes the output to the next layer. These hidden layers are where the network learns to detect and abstract patterns from the data. Multiple hidden layers allow the network to build increasingly complex representations of the input data. For instance, in image creation, the early layers might capture basic shapes or textures, while deeper layers capture more abstract features like style or composition.
The output layer in a feedforward neural network is responsible for generating final outputs denoted by y1, y2, …ym. In GenAI models, this layer generates the content the user has requested. Considering the text generation example again, the output layer neurons would contain words or tokens that have been produced by feeding the input prompt through the network.
The connections between neurons have weights that are adjusted during the training process. Training a feedforward neural network involves several key steps to minimize errors in the network's predictions and improve its performance.
Initially, the input data is passed through the network layers in a forward pass, where each layer processes the input and passes it to the next layer. Once the network produces an output, the difference between this predicted output and the actual target output is calculated, resulting in an error. This error is measured using a cost function, which quantifies the discrepancy between the predicted and actual outputs.
The network's weights are fine-tuned using backpropagation. This process involves propagating the error, computed as the difference between predicted output and actual output, backward through the network. The information flow in reverse allows the training algorithm to determine how much change is needed for each weight so that it can reduce this error or gradient. Gradient descent is an optimization technique that allows each weight to be adjusted by calculating its degree of change. The magnitude of each adjustment is determined by the learning rate, a parameter that controls the size of the steps taken toward the minimum error. After calculating these gradients, the weights are readjusted.
Generative Adversarial Frameworks
A Generative Adversarial Network (GAN) is an ML framework for training GenAI models. GANs consist of two neural networks: a generator and a discriminator. The generator creates fake data that aims to resemble real data, while the discriminator evaluates the authenticity of the data, distinguishing between real and fake samples.
The training process of GANs is adversarial, meaning that the two networks compete against each other. The generator's goal is to produce data that is indistinguishable from real data. The discriminator's goal is to accurately identify whether the input data is real or generated. During training, the generator receives feedback from the discriminator, which helps it improve the quality of the generated data. This iterative process continues until the generator produces highly realistic data that can fool the discriminator.
GenAI Use Cases
Generative AI is becoming more popular due to its ability to assist humans in completing various tasks. Even in the early stages of GenAI, it assists developers and engineers with design, writing code, debugging, and other tasks in which some form of output is expected.
Software developers can interact with GenAI products like OpenAI’s ChatGPT and GitHub Copilot using natural language in a web browser or in an extension to an Integrated Development Environment (IDE) like Microsoft VS Code. Within an IDE, GitHub Copilot scans code and text in a file and suggests auto-completion of code that logically fits as the next line or lines. A developer can autogenerate entire functions by simply typing out some well-worded inline documentation and pressing a hotkey that maps to accepting the suggested code. In addition, Copilot has a text chat in which you can discuss concepts and approaches with the language model. These development models can also generate documentation based on existing code, improving readability.
Due to the complexity of an environment, in modern software development, error logs can contain messages that do not seem directly related to what you are trying to capture. Using GenAI, developers can copy an error output or log in to these models to receive either the solution to an issue or a list of steps that they can take to debug further.
GenAI is also being used to automate and enhance the design process. AI models can generate design layouts, graphics, and user interfaces based on specific requirements or aesthetic preferences. For example, Adobe's AI-powered tools can assist designers by automatically generating color palettes, suggesting design elements, and creating variations of a design concept. For example, the Adobe toolkit generative fill feature allows you to select a region within an image and instruct the AI to make an addition such as “put a penguin here” or something to that effect. The incredible results are hard for a human to match without extensive training and practice. Capabilities such as this speed up the design process and encourage creativity by providing designers with new ideas and inspiration.
Generative AI Inference Challenges
Trained machine learning models draw conclusions from new data in a process called inference. The training and fine-tuning performed on a model will correlate with the quality of its inferences, leading to results that vary in alignment with user expectations. For example, a model trained on millions of images of cows can generate an image of a cow exceedingly well, but it will never be able to render an image of a giraffe.
To illustrate the inference process, refer to the figure shown. When a user provides input, such as "Create a work of art for me," the GenAI model processes this input through its layers, drawing on the patterns that it learned during training. The resulting output is a new piece of art that combines elements from the training data with a degree of variability.
Since a small set of training data was used, the model's ability to generate diverse and novel outputs is limited. This constraint means that when prompting the model with a vague input like "Make art," you will invariably see elements of each datum in the training set reflected in the output. The output artwork blends characteristics from the training images, imitating imagination with randomness, but the output is not truly imaginative.
The quality of a generative model’s inferences determines its overall performance across different systems and applications. For example, real-time image generation demands that the model creates high-quality images quickly for use in virtual reality or gaming apps. The same is true of text completion in user applications like chatbots or writing assistants; fast and accurate inference is needed to give relevant responses coherently and quickly.
GenAI Challenges and Limitations
Although generative AI has made significant progress in creating believably human content, it has some challenges to its continued growth.
Training Data Quality and Cleanliness
When training AI models, data cleanliness direclty affect the quality of the output. "Clean data" is valid, accurate, complete, and consistent. Biased, incomplete, or noisy data can lead to incaccurate, unfair, or inappropriate outputs. For instance, if the language used in the data used to train a language model is biased, then such a model will produce biased and harmful content.
High-quality training data can only result from thorough data preprocessing, which involves cleaning, normalizing, and mitigating bias. Training data also requires the use of diverse and representative datasets so that the output is fair and accurate across different contexts. Addressing data quality issues is essential for building trustworthy generative AI systems that can be reliably used in various applications.
Ethical and Social Implications
As is the case with any new powerful technology, malicious actors will certainly harness the power of GenAI to do bad things. The potential to misuse AI to create malicious software or damanging false videos known as deepfakes is troubling. Deepfake technology utilizes generative models to alter images and video resulting in seemingly real but fraudulent representation of people. such deepfakes can be used maliciously to disseminate false information about someone, slander them, or create counterfeit materials that pose significant danger to privacy, safety and confidence.
To address the social and moral implications that GenAI may cause, strong measures must be put in place to guarantee responsible creation, deployment, and use of GenAI systems. These include bias detection methods and guidelines outlining acceptable usage practices. These guidelines should ensure openness about how AI works throughout its lifecycle from conception through production until retirement at the end-of-life stage
Neural Network Complexity
The process of creating or improving a GenAI model is complex. With an Artificial Neural Network (ANN), software developers that try to debug an error in code cannot inspect the line on which an error occurs and track that error through a codebase as they usually can do. GenAI models powered by deep neural networks consist of many interconnected layers of neurons. Each Layer transforms the input data in complex ways, with the final output being the result of numerous computations across these layers
The complex architecture and high number of parameters make it impossible to trace how any specific input produces a givin output, effectively operating in a black box.
Resource Implications
The computational complexity of training and inferring from generative models, especially those using deep learning, demands significant resources. Models require powerful GPUs and large datasets for high-quality results, posing accessibility and scalability challenges for smaller organizations. Training Generative Adversarial Networks (GANs) on high-resolution images, for example, demands specialized hardware and high energy consumption, leading to significant costs and environment impacts.
Efforts to address these challenges include developing efficient algorithms, optimizing model architectures, and leveraging advancements in hardware technology and cloud-based AI services. For instance, specialized GPUs designed by companies like NVIDIA support massive computations required by GenAI, enabling higher throughput and lower latency.
Hallucinations
Mitigating hallucinations is another challenge in GenAI. The term "hallucination" refers to when AI that is capable of natural language processing produces outputs that are not based on reality or fact. For example, when prompted to create a logo for "Excellent Corp.," the model may produce something that looks mostly normal but, upon closer inspection, contains strange symbols derived from something present in the training data.
Generally, output that contains a hallucination is useless. However, in text generation, hallucinations may be difficult to distinguish from fact and could lead to the spread of false information.
GenAI Bias and Fairness
Bias in AI systems can take forms such as racial or gender-based discrimination that can reinforce cultural stereotypes and lead to unequal treatment. Generative models can be biased and unfair because they inherit the biases present in their training data. A language model trained on biased data may produce content that reflects those same biases. Understanding and identifying bias in generative models is essential to ensure that AI models perform by ethical standards. In the following figure, the red squares represent biased data in a training dataset, whereas the blue squares represent clean data. The output from a model trained on each representation of a training dataset will produce outputs that can be biased, somewhat biased, or not biased at all. You can decrease bias by reducing biased data or adding unbiased data to the training dataset.
Ensuring fairness and equity in generative models requires various techniques to reduce bias. One such technique is data augmentation, which involves adding different and representative samples to the training dataset. Additionally, fairness constraints and bias detection algorithms can be incorporated during the model’s training process to identify and correct any bias patterns that may exist within it. Continuous monitoring and evaluation of the model’s outputs should be done to address any new biases that may arise before deploying the model for use.
The ethical implications of biased generative AI outputs must also be considered in deployment. Ethical considerations include the potential societal impact of generative AI, such as spreading misinformation through deepfakes or creating harmful content. Developers and stakeholders must address fairness and accountability in AI systems to build trust and acceptance among users. Adherence to ethical guidelines and standards involves establishing clear ethical frameworks for AI development and deployment, promoting transparency in model design and training, and fostering collaboration between AI researchers, ethicists, and policymakers.
GenAI Resource Optimization
Optimizing the use of hardware resources, such as GPUs and TPUs, is essential for efficient inference in GenAI models. Techniques like model pruning, quantization, and parallel processing can enhance performance by reducing the computational load. Model pruning involves removing unnecessary parameters from the neural network, making the model smaller and more efficient without significantly compromising accuracy. Quantization reduces the precision of the model's weights, producing lower-precision computations that require less power and memory. Finally, parallel processing is the simultaneous processing of multiple tasks, or splitting one task into many threads. With parallel processing, each compute resource works on its share of the overall task, speeding up computation time dramatically.
Managing the energy consumption of generative models during inference is crucial for minimizing environmental impact and ensuring sustainable AI deployment. Generative models can consume substantial amounts of energy, contributing to the carbon footprint of AI applications. Balancing performance and energy efficiency involves optimizing algorithms and hardware to reduce energy usage. Data centers running generative AI workloads can benefit from energy-efficient practices such as using renewable energy sources and implementing cooling solutions that reduce overall energy consumption.
Optimizing cloud or on-premises infrastructure can reduce the costs of deploying and maintaining GenAI models. Techniques like serverless computing and distributed inference can help achieve cost-effectiveness by dynamically scaling resources based on demand. Serverless computing allows developers to run code without provisioning or managing servers, while distributed inference uses multiple machines to handle large-scale inference tasks that improve efficiency and reduce costs.
Generative AI vs Traditional AI
Traditional AI and GenAI have very different technical approaches. Each offers unique advantages, but also face distinct challenges and limitations.
Traditional AI
Traditional AI relies primarily on rule-based systems. The traditional AI model uses symbols, also called tokens, to represent some element of knowledge while also using rules to analyze these symbols and come up with conclusions. This approach, with well-defined rules that are explicit and transparent for decision-making, is relatively easier to understand and debug.
One example of a traditional AI application is a rule-based financial system designed to detect fraudulent transactions. The system is based on a set of rules representing known patterns that indicate some likelihood of fraudulent activity. One transaction, such as a purchase from a specific store in New York City on January 5th, 2024 for $2500, arrives in a JSON format and then is sent to the inference engine. The inference engine uses its various business logic rules to analyze the transaction details to produce a verdict.
The inference engine must draw on a knowledge base to generate its output. The knowledge base contains a large set of historical records or transactions, known fraud patterns, specific and predefined in-house rules for identifying strange or suspicious activity, and even rules relating to regulatory requirements.
Some examples of these rules may include:
IF amount > $2000 THEN flag as potentially fraudulent
IF location is not within 50 miles of user's typical locations, THEN flag as potentially fraudulent
IF merchant is not in the list of frequent merchants, THEN flag as potentially fraudulent
Each rule in the knowledge base is evaluated against the recent purchase. Then, the results of the analysis, including a detailed enumeration of each rule and exactly which parts of the rule failed, are sent back to the inference engine. The output generated by the inference engine includes a final ruling on whether the transaction is likely to be fraudulent, as well as a summary of why the transaction was flagged, and a confidence score.
Traditional AI is well suited to highly structured environments. However, these systems are often severely limited in flexibility and scalability. They will almost always require manual updates to the list of rules and logic that govern their decision-making if the system needs to be changed to handle new or unexpected scenarios. As the number of rules and scenarios grows, the number of interactions between those rules can grow exponentially, and the system becomes more challenging to manage.
Gen AI
Conversely, the goal of GenAI is to come up with some type of novel content or decision-making by making use of immense datasets. GenAI uses neural networks and machine learning to accomplish this goal. This type of AI outperforms humans in extremely complicated environments, where the number of variables is more than could be typically handled by distinct rules.
Rather than being used in a fraudulent transaction detection system, GenAI would be more likely to be used financial advisor chatbot. If you try to code a traditional AI application to give financial advice, the sheer number of variables involved would make it nearly impossible to cover them all comprehensively. GenAI and neural networks still have the same level of complexity under the hood. However, with GenAI, the focus for the developer shifts from listing specific rules to tuning the model based on inputs to achieve a low error rate. In the end, the developer may not be able to explain how the model came up with its output, but the system may still be able to do so by discovering commonalities between historical and current data.
The process of using a GenAI model begins with an interaction, usually via a prompt provided by the user. That prompt is turned into tokens through a process called tokenization. A token can be a word, a fragment of a word, or even a single character, depending on the tokenization scheme used by the model.
The initial input tokens are processed by the model to generate a probability distribution for the next possible token. This generation process is iterative, where each new token generated is influenced by both the original prompt and all previously generated tokens. After each token is generated, the model incorporates it into the sequence and uses the updated sequence as the input for the next iteration. This iterative process continues until:
The model generates a special token (often referred to as an "end-of-sequence" token) that signals the completion of the output
Another stopping criterion is met, such as reaching a predefined token limit
GenAI offers significant advantages, such as remarkable creativity and innovation, adaptability to new tasks with minimal retraining, and effective personalization. However, the inherent complexity of these models, while enabling sophisticated outputs, can also lead to unpredictability. GenAI models can inherit biases present in the training data, requiring continuous updating and careful data selection to ensure relevance and fairness. Additionally, the resource-intensive nature of training these models presents significant challenges.
GenAI vs. Traditional AI Data Requirements
Due to their technical approaches and application domains, traditional AI and generative AI have vastly different data requirements.
Traditional AI
Traditional AI systems depend heavily on explicit, structured data and predefined rules. They are designed to operate within well-defined boundaries and often require clear and concise data inputs. When the AI system needs to make a decision, it consults this structured data to apply the relevant rules and return conclusions.
MYCIN, a medical expert system, processes patient data using predefined diagnostic rules and recommends a set of applicable treatments. The historical dataset includes many medically relevant details, including patient details like age, gender, symptoms, medical history, test results, and confirmed diagnoses. When a new patient presents with some problematic symptoms, the system compares the new patient's data against this historical dataset. The AI system uses this structured data stored in its knowledge base to apply the predefined rules to suggest a diagnosis for the new patient based on similar patterns.
Traditional AI
P001
45
Male
Fever, Cough, Fatigue
Hypertension
Elevated WBC, Positive for Streptococcus
Bacterial Infection
P002
30
Female
Headache, Nausea
Migraine
Normal MRI, Normal CT
Migraine
P003
65
Male
Chest Pain, Shortness of Breath
Coronary Artery Disease
Elevated Troponin, Abnormal ECG
Myocardial Infarction
Because of this dependency on well-defined and structured data, traditional AI systems require less data overall and are easier to manage. They can outperform even while using smaller datasets and simpler data formats. However, these small datasets also limit a traditional AI model’s flexibility and its ability to expand easily. Traditional AI systems struggle to process ambiguous data, and they cannot efficiently handle unstructured data such as free-form text or complex images. This limitation puts a hard cap on the number of relevant traditional AI applications and limits them to well-defined domains in which the data is predictable and consistent.
Generative AI
In contrast, GenAI needs vast amounts of data for training to learn and to generate new, realistic outputs. These huge diverse datasets allow GenAI models to understand and mimic complex patterns and structures. For example, GenAI models like Generative Adversarial Networks (GANs) can use pictorial datasets to learn the general features of cats and their environments. Then, the model can be used to generate images that include cats with realistic backgrounds.
Generative AI
Domestic Cats
70,000
Images of domestic cats in various environments, poses, and breeds.
Backgrounds
10,000
Images of various backgrounds to help the model generate realistic environments.
However, since the dataset shown categorizes these images simply as "Domestic Cat," any images of these cats that the model generates will likely be lacking in some details. The cat images may have the right shape and size, but might not resemble any specific breed. Or they might look like a breed that does not exist at all due to mixing traits from different breeds. The model with the dataset shown cannot distinguish between breeds or identify them by name. To train a GenAI model to generate cat images of a single specific breed, or one of several breeds, a developer needs to expose it to thousands of images of various breeds along with metadata that somewhat describes the data that the model is seeing.
Generative AI
Persian Cat
10,000
Images of Persian cats in various environments and poses.
Siamese Cat
10,000
Images of Siamese cats in various environments and poses.
Maine Coon
10,000
Images of Maine Coon cats in various environments and poses.
This dataset, with labels and other metadata included, helps the model to tell the difference between different breeds by learning their unique characteristics. By linking these visual features with various metadata about their breeds, the model's ability to recognize these breeds, or to generate them accurately, improves drastically. The images that are generated may more clearly show off the distinctive spotted coats of a Bengal cat or the hairless body of a Sphynx cat. The model will now be able to produce much more detailed and realistic images of a selected cat breed.
The ability to handle unstructured data is a significant advantage of GenAI. These models create a wide range of outputs in formats as disparate as text and images to even music or code. However, this advantage comes with high data requirements. GenAI models need extensive datasets to function effectively, and vague or low-quality training data will lead to low-quality outputs. This requirement can be a significant barrier because acquiring and processing such large datasets demands substantial resources in terms of time, storage, and computational power.
Future Trends in AI
The future of AI is marked by significant advancements and evolving applications that promise to reshape various industries and aspects of daily life.
Several emerging trends are shaping the future of AI, driving innovation and creating new opportunities:
Edge AI: Deploying AI models on edge devices, such as smartphones and IoT devices, allows for real-time data processing and decision-making without relying on cloud infrastructure. This trend is crucial for applications requiring low latency and high privacy.
Explainable AI: As AI systems become more complex, there is a growing need for explainable AI to ensure transparency and trust. Research in this area focuses on developing models that provide clear and understandable explanations for their decisions.
AI Ethics and Governance: Ensuring ethical AI development and deployment is becoming increasingly important. Responsible development involves addressing biases in training data, ensuring fairness, and establishing guidelines for responsible AI use.
Transfer Learning: Transfer learning enables models to use pre-trained knowledge on large datasets and apply it to specific tasks with less data and training time. This approach makes deep learning models more efficient and scalable.
ML and DL Integration: The integration of machine learning and deep learning techniques can lead to more robust AI solutions by combining the strengths of both approaches. For instance, using machine learning for feature selection and deep learning for feature extraction and prediction can enhance model performance and versatility. This hybrid approach allows for more efficient data processing and improved accuracy in complex tasks.
AI Language Models
Language models are sophisticated AI systems designed to understand human-language inputs, generate their human-language outputs, and manipulate language. These models use deep learning techniques to process and produce text or voice, enabling a wide range of applications such as chatbots, text summarization, and translation services. By analyzing vast amounts of language data, language models can perform complex tasks that mimic human language comprehension and generation.
Core Concepts
The reason language models are so important in the field of AI is that humans speak different languages. It is important to note that the only way most people will be able to interact with advanced AI is by communicating with it using their own language. A user prompts a model to make it perform some action on their behalf. Therefore, language models can accept inputs and generate outputs in both text and spoken formats.
Language models can be designed to interact with many external systems and services through Application Programming Interfaces (APIs). A properly trained language model can take information from its interaction with a user and use it to query an external database, send a command to a smart home device, or retrieve information from an online service. For example, once a model has been provided with some permissions, passwords, or secure tokens, it could act on a user’s behalf to:
Access a reservation service's API
Search for available times
Complete the booking process without much user intervention
The combination of natural language processing and integration with other technologies enhances the functionality and utility of AI systems, bringing them closer to truly understanding and acting on human intentions in various domains. The AI systems become capable of acting on complex, multi-step goals, making them powerful tools for automating tasks and providing seamless, intelligent assistance. It is likely that in the future, language model integrations will become more complex and interact with other types of AI to accomplish incredible things faster than humans could ever hope to.
Types of Language Models
There are two distinct categories within the generative AI domain, large language models (LLMs) and small language models (SLMs), each with unique characteristics and applications.
LLMs are the more advanced of the two types of language models. They require much larger datasets to be trained properly. They can possess billions of parameters within the neural networks that they depend on to enable their deep learning approach. This approach allows the models to generate coherent and contextually relevant text across various tasks. For example, OpenAI’s ChatGPT, with its 175 billion parameters, is an LLM capable of performing diverse language tasks, including writing essays, answering questions, and generating conversational responses.
SLMs are a simpler type of language model that requires a smaller number of parameters and contains less computational complexity. While they can also be trained using deep learning techniques, SLMs are usually optimized for specific tasks, with a smaller boundary as to what is in scope and what is not.
An SLM’s smaller size and efficiency are achieved via techniques such as knowledge distillation, pruning, and quantization that allow SLMs to operate efficiently with limited resources. Knowledge distillation transfers knowledge from a pre-trained LLM to a smaller model, capturing its core capabilities without the full complexity. Pruning removes less useful parts of the model, and quantization reduces the precision of its weights, further reducing its size and resource requirements.
SLMs often do not require the extensive datasets and computational power that LLMs do. Therefore, SLMs are more accessible for practical applications in which resources are constrained. A custom chatbot model trained on a specific customer service dataset with a few million parameters exemplifies an SLM, providing targeted responses without the extensive resource requirements of an LLM.
LLMs vs. SLMs
Decision makers need to understand the difference between large language models (LLMs) and small language models (SLMs) so that they can be informed when deciding which type of AI application they want to build. Knowing the strengths and limitations of each type of model helps in selecting the right approach for specific needs. This foundational knowledge ensures that organizations can optimize resources and make informed decisions that align with their specific objectives.
The simpler architecture of SLMs allows them to be deployed in resource-constrained environments such as embedded systems or mobile applications. Additionally, SLMs are more appropriate for applications that require more security due to their smaller codebases and fewer potential surfaces for security breaches. They are also better for the environment, using less energy and memory. This efficiency also makes them ideal for edge computing and real-time applications. Because SLMs consume fewer computational resources, they offer significant cost savings and reduced environmental impact.
SLMs are easier to train, making them suitable for practical applications in which efficiency and speed are paramount. They are most applicable for tasks that do not require a deep understanding of the text being read and that lack extensive text generation. For instance, SLMs can quickly classify emails as spam or not, or extract keywords from short texts. Their deployment in scenarios where quick, on-the-fly processing is needed further underscores their utility in specific, targeted applications.
Conversely, LLMs are characterized by their training on vast datasets. This characteristic enables them to handle a wide range of language tasks. The large datasets allow them to generalize and abstract out meaning, generate complex and diverse text, understand the context in which an input was provided, and perform multi-step reasoning. They are likely to perform well across diverse tasks without extensive task-specific training. LLMs have been made to generate human-like responses in chat applications, translate text, and write essays. This versatility makes them suitable for a wide range of applications, from automated customer service to creative content generation. The ability to generate nuanced and context-aware responses further distinguishes LLMs, making them indispensable for sophisticated AI applications that demand higher levels of accuracy and complexity.
Last updated