Synergise AI

Large language models “versus” small: embrace diversity A short example showing how natural language understanding (NLU) benefits from a mix of models. 👉A critical task in NLU is understanding what “she” and “that choice” refers to in the following sentences: “Janet doesn't have a car. She said if she were to buy one, it would be red. Bob criticized that choice.” (Multiple words or phrases referring to the same thing are called coreferences.) Without the ability to associate references together, a large part of the meaning is lost. Ask GPT 3.5 Turbo, "In those sentences, what color car doesn't he like?” and it will answer correctly, "Bob doesn't like red cars." Not only did it correctly connect the references in the input sentences, it also referenced “he” back to “Bob” in the question. Great! 👉 Job done? Turbo can identify coreferences? A huge step toward extracting maximum information from documents such as the following? (Taken from “Speech and Language Processing. Daniel Jurafsky & James H. Martin. Copyright © 2021.”) <document>Victoria Chen, CFO of Megabucks Banking, saw her pay jump to $2.3 million, as the 38-year-old became the company’s president. It is widely known that she came to Megabucks from rival Lotsabucks. </document> (Can you spot the four references to Victoria Chen and the three references to Megabucks?) 👉 Let’s try it! User: "Resolve all coreferences in the following document. <Insert Chen document here>" 3.5 Turbo answer: "<document>Victoria Chen, CFO of Megabucks Banking, saw her pay jump to $2.3 million, as the 38-year-old became the company’s president. It is widely known that Victoria Chen came to Megabucks from rival Lotsabucks. </document>" Turbo resolved only one of the five references, correctly replacing “she” in the second sentence. 👉 Back to the drawing board Let’s try a much smaller but fine-tuned model on this particular task. The Fastcoref Python library’s model is about 600M parameters; about 33x times smaller then 3.5 Turbo. (A guess on the size of the GPT 3.5 Turbo model is 20 billion parameters - only OpenAI knows for sure.) Fastcoref answer: “Victoria Chen, CFO of Megabucks Banking, saw Victoria Chen, CFO of Megabucks Banking's pay jump to $2.3 million, as Victoria Chen, CFO of Megabucks Banking became Megabucks Banking's president. It is widely known that Victoria Chen, CFO of Megabucks Banking came to Megabucks Banking from rival Lotsabucks.” Fastcoref replaced all references to both Chen and Megabucks with the longest, most specific name. (Interesting that it included Megabucks Banking into the Chen longest name. Good or bad?) 👉 Wrapping up Modern cars incorporate 50 to 100 microprocessors of all sizes and function. In the future, I expect line-of-business apps to use dozens, if not hundreds, of models from large and general purpose to small and dedicated. The first microprocessors were used in cars in the late 1970s. What will the ecosystem of models look like in 40 years?

Create an AI roadmap to coordinate adoption – people and ethics   A fifth post on orchestrating organizational adoption of AI. The first four: process for creating the roadmap, components, business strategy, technology and people. 👉Reactions to AI Be prepared to work with different reactions to AI: -       Innovators, early adopters, active collaborators -       Pragmatists and fast followers -       Skeptics, critics and traditionalists -       Compliance-focused -       Group-focused (not organization) -       Naysayers and “wait and see”   👉Consider an AI center of excellence (CoE) Build a dedicated business unit responsible for overseeing and coordinating all AI projects within the organization. It: -       Identifies AI use cases and creates implementation strategy. -       Contains a diverse team; AI/IT experts, business executives, and domain specialists. -       Bridges the gap between executive decision-making and execution of AI initiatives. -       Creates a single vision for AI adoption throughout the enterprise. -       Standardizes best practices and maintains communication channels.   👉Governance and Ethics -       Responsible AI: serve the greater good while minimizing potential risks and unintended consequences. -       Develop safeguards against biases and discrimination in AI systems. -       Mitigate biases through data selection, transparent algorithms, and testing. -       Address ethical concerns such as privacy and data security. 👉 (Partial) Responsible AI checklist -       Design for fairness and consistency. -       AI learns from training data, verify that the data isn’t biased. -       Define what "fair" means for the specific context. -       Detect and mitigate bias as part of established process. -       Commit to transparency, including algorithms and notifications to end users when they’re interacting with an AI. -       Make people responsible and accountability for outcomes. -       Establish strategies for communicating AI-related problems to the public and regulatory authorities. -       Trustworthy AI means availability is scheduled and consistent, output generation is reliable, even when confronted with changing conditions, unexpected events, or data drift. -       Assess and address risks. -       Verify algorithms consistently produce accurate outcomes for each new dataset. -       Adhere to data regulations and employ data for predefined and mutually agreed-upon purposes. -       Protect the privacy of data collected by AI assistants, including (not) sharing with employees, partners and law enforcement agencies. -       Maintain awareness of the customer data collected, its intended use, and compliance with customer expectations. -       Provide customers control over their data, including the ability to opt in or out of data sharing. Give them a feedback mechanism. -       Protect AI assets from cybersecurity risks that might lead to physical and/or digital harm. —To Be Continued—

Create an AI roadmap to coordinate adoption – technology and people   A fourth post on orchestrating organizational adoption of AI. The first three: the process for creating the roadmap, its components and business strategy. 👉Algorithm and model primer -       Algorithms are mathematical procedures, i.e. they are a series of steps describing how to process numbers. -       Machine learning algorithms, in particular, analyze data and learn from it. -       A mathematical model is a set of equations that create relationships between some variables. -       A language model is a mathematical model plus algorithms designed to process some representation of human communication. -       A large language model (LLM) is a “big” language model. Big, as of this writing, is hundreds or more billions of variables (parameters). The parameters are learned through training.   👉Architecture and Platform When defining and integrating architecture and platforms for AI projects, ask these questions: -       What is the existing infrastructure, and how will AI integrate into it to for compatibility and scalability? -       What are the computational requirements? Are the hardware and software resources available? -       What data storage and management systems are in place to support AI data requirements, including volume, velocity, and variety? -       How will data security and privacy concerns be addressed? -       What are the specific AI frameworks and tools that align with your organization’s objectives and technology stack? -       What cloud or on-premises solutions are most suitable, considering factors like cost, scalability, and ease of deployment? -       How will data preprocessing, cleaning, and feature engineering be handled? -       Is real-time processing required, and how will the architecture support it? -       What is the strategy for model deployment and monitoring, including version control and performance tracking? -       How will ongoing maintenance, updates, and optimization of the architecture keep it aligned with changing project requirements and industry advancements? 👉Communicate, engage and on-board people -       Identify stakeholders. Understand their concerns, expectations, and roles in the AI adoption. -       Articulate the value that AI brings. -       Assemble teams with diverse skills. -       Organize training sessions. -       Start with a small-scale AI project -       Build a proof-of-concept. -       Create a communication plan. -       Encourage participation and feedback from all levels. -       Establish a mechanism for collecting feedback. -       Share case studies and success stories. -       Keep the organization updated on AI trends and advancements. -       Maintain a focus on AI adoption and its alignment with the company's strategic objectives. —To Be Continued— with governance and ethics

Create an AI roadmap to coordinate adoption – data strategy and infrastructure   A third post on orchestrating organizational adoption of AI. The first two posts covered the process for creating the roadmap and business strategy. 👉References -       “The AI Advantage: How to Put the Artificial Intelligence Revolution to Work (Management on the Cutting Edge)” by Thomas H. Davenport -       "Artificial Intelligence and Machine Learning for Business: A No-Nonsense Guide to Data Driven Technologies” by Steven Finlay -       “Enterprise Artificial Intelligence Transformation”, by Rashed Haq 👉Data Strategy Data quality often determines the efficacy of an AI system. Consequently, it is important to establish a complete data strategy to avoid “garbage in, garbage out”. Include: 1.     Management of data collection, storage, integration, and cleansing. 2.     Assurance of the provision of high-quality data. 3.     Verification and structuring of the data to achieve accuracy and usability goals. -       Automation is key. Manual data processing may be sufficient for small-scale model usage but enterprise-level use requires automated data pipelines. This principle holds even in cases where the volume of data is low; it is still advisable for quality reasons. -       Large volumes of data are not required. Relatively small yet high-quality datasets can be better than extensive yet low-quality datasets. -       A data strategy encompasses the establishment of systematic procedures for the acquisition of information, especially pertaining to results and outcomes, which subsequently generates data.   👉Infrastructure -       Infrastructure, in this context, includes aspects of making data accessible and providing the processing capacity to handle it. The infrastructure empowers the AI team to expand and configure models, and feed them with the inputs they need. -       Data may be stored within silos, inaccessible to other teams within the organization. This isolation arises from structural, organizational, and bureaucratic constraints. Data silos may be a significant barrier, as AI systems often need to access across the organization. -       AI systems often consume considerable computational resources. According to OpenAI, the computing power required to train popular AI models has been doubling roughly every 3.5 months since 2012. This exponential growth in computational demands should be factored into the planning. -       An important decision regarding AI infrastructure is the choice between cloud and on-premise solutions. Cloud computing offers scalable, flexible, and cost-effective infrastructure solutions. However, data transmission and storage concerns may necessitate on-premise. -       Selecting the right cloud provider involves evaluating performance, scalability, cost-efficiency, and compatibility with existing IT systems. —To Be Continued— with algorithms, architecture and platform

Our "Agile for AI" methodology helps you apply the principles of Agile to AI projects while incorporating the realities of AI workflows to help your project succeed. Get our free guide here: https://hubs.li/Q02hXKmq0

Interested in partnering with an AI implementation and deployment specialists? Look no further than Synergise AI to help your business scale in a time-efficient and cost-effective manner to accelerate your business growth and revenue. https://hubs.li/Q02hXMps0

YES to ALL of this! Great article by Jeff Hollan of Snowflake for VentureBeat - How data engineers should prepare for an AI world. Read here: https://hubs.li/Q02hXMLJ0

Thinking of using AI-generated photos or videos? They may soon require watermarks in the state of California, per Bloomberg Law. Learn more about lawmakers' efforts here to see how it could impact your business: https://hubs.li/Q02hXMSZ0

Learn the right questions to ask, which teams to involve, how to think about AI ethics and regulations, and more in our free AI Roadmap guide created by AI expert John Vogel, Ph.D.. Download the guide here: https://hubs.li/Q02hXKgV0

Sometimes it's a better use of your time and resources to do it yourself. Other times? Not so much. Our article breaks down the things to consider when deciding which direction to go in. https://hubs.li/Q02hXNYk0