Unlocking the Power of Generative AI An Executive's Guide.pdf
- 1. Unlocking the Power of Generative AI: An Executive's Guide May 2023 TM
- 2. Outline The Rise of Generative AI (Gen AI) Large Language Models Powering Generative AI Building with Generative AI Leveraging the Power of Generative AI for your business Using MLOps to create business value of Generative AI The Katonic advantage
- 3. Uber Time to reach 100 million monthly active users No. of months ZOOMING AHEAD ChatGPT TikTok Instagram Pinterest Spotify Telegram 2 9 30 (2 yrs 6 mnths) 41 (3yrs 5 mnths) 55 (4 yrs 7 mnths) 61 (5 yrs 1 mnth) 70 (5 yrs 10 mnths) ChatGPT’s explosive global popularity became an inflexion point for AI’s public adoption. ChatGPT – The Big Bang Moment for Gen AI katonic.ai
- 4. katonic.ai Large Language Models are not limited to just text ChatGPT by Open AI Bard by Google BLOOM by BigScience Galactica AI by META DALL-E2 by OpenAI Imagen by Google Stable Diffusion by Stability AI MidJourney v4 by MidJourney Vall-Eby Microsoft for speech PointEby OpenAI for 3D objects Imagen Video by Google Chinchilla by Deep Mind for text katonic.ai
- 5. Windows 3/'95 Internet iOS/Android LLMs/AGI Data processing activities Communication based activities On-the-go activities Knowledge activities Range of human activities that software has "eaten" Major platform launches that have enabled new types of applications, over time 1990s+ Mid-late 1990s+ 2007-8 2022+ 1990s+ katonic.ai LLMs Are The Engine | AI Applications are the Product
- 6. Finance & Legal: Draft and review documents, patents and contracts; find, summarise and highlight important points in regulatory documents; find and answer specific queries from large documents; scan through historical data to recommend a course of action. Marketing & Sales: Automate SEO-optimised content generation, enhance ad bids, hyper-personalise communication and deployment, create product user guides by persona, analyse & segment customer feedback, hyper-capable chatbots for upsell and cross-sell . Customer service: Natural-sounding, personalised chatbots and virtual assistants can handle customer inquiries, recommend swift resolution, and guide customers to the information they need. HR & Recruitment: Smart-shortlist of candidates, risk assessment of candidates, self- service of HR functions via chatbots and automation Information technology: Advanced code writing code and documentation, code review and error detection, and accelerated software development, auto-complete data tables, generate synthetic data. katonic.ai Potential opportunities and use cases
- 7. Data Security , privacy and cost is a BIG hurdle katonic.ai
- 8. 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Banking Insurance Capital Markets Software & Platforms Energy Communications & Media Retail Health Aerospace & Defence Automotive High Tech Travel Utilities Consumer Goods & Services Chemicals 40% of working hours across industries can be impacted by Large Language Models (LLMs) Why is this the case? Language tasks account for 62% of total worked time in the US. Of the overall share of language tasks, 65% have high potential to be automated or augmented by LLMs. Based on their employment levels in the US in 2021 Work time distribution by industry and potential AI impact: Higher potential for automation Higher potential for augmentation Lower potential for augmentation or automation Non - language tasks Nearly 6 in 10 organisations plan to use ChatGPT for learning purposes and over half are planning pilot cases in 2023. Over 4 in 10 want to make a large investment. Adapt or be left behind Companies must reinvent work to find a path to generative AI value. Business leaders must lead the change, starting now, in job redesign, task redesign and reskilling people. katonic.ai
- 9. Generative AI Tech Stack Applications API Layers MLOps Foundation Models Build Your Own Closed Source Open Source GPT-3.5 DALL.E 2 LaMDA Codex CLIP DALL.E 2 BLOOM Stable Diffusion Infrastructure Add guardrails Embed Knowledge Add Skills How Can Businesses Use Generative AI, Today? Generative Models katonic.ai
- 10. Option Explanation Cost Subscribe and Use Embracing off-the-shelf tools leveraging LLMs that are already available. Minimal Consume with Guardrails Build Guard Rails by adding pre and post-processing restrictions to off the shelf LLM's $ Augment Use database lookups to tailor LLMs to an organisation’s needs. $$ Fine Tune Using fine-tuning to tailor LLMs to an organisation’s needs $$$ Build your own Build and Train your model from scratch with your data $$$$ Off the Shelf katonic.ai What are the Options? Customise Train your own
- 11. Using paid subscriptions or corporate user plans of Generative AI tools like ChatGPT, Jasper, Notion etc. for trial and training of employees without exposing confidential company data. Use-cases limited to the generation of low-quality and low-risk content. LIMITATIONS RECOMMENDATION Limited to publicly available info Acceptable only for trial and training of employees. BENEFITS Fastest turnaround time Cost limited to subscription fees Cybersecurity Concerns Fabricated Information. Copyright Issues Data Privacy Deepfakes Strongly recommend avoiding sharing of any confidential information. Off the Shelf - Benefits and Limitations Requires the least LLM training technical skills. Cost limited to subscription fees Can leverage the best-performing LLMs in the market Good for prototyping apps and exploring what is possible with LLMs. katonic.ai
- 12. katonic.ai Guard Rails 1. Customer query. 2. Check for Guard Rails defined by the organisation 3. The request is processed or filtered and all responses are stored for audit and training 4. Response, by way of LLM, sent back to user. LLM App, hosted by the organisation Customer service appinterface Consume with Guardrails
- 13. katonic.ai Knowledge base LLM Augment 4. Response, by way of LLM, sent back to user. App, hosted by the organisation Customer service appinterface 3. Articles from knowledge base and customer query are processed by the LLM to construct a response. 1. Customer query. 2. Lookup of relevant articles, using keywords from customer query. Open Source
- 14. katonic.ai Fine Tune Dataset Enterprise Data Transfer learning Custom Model User Prompt Interface Output Foundation models are trained on massive publicly available data sets. Transfer learning enables companies to build on top & fine-tune these models for their use case with less intense requirements. Training Foundation model
- 15. Organisations can boost the capabilities of their applications by integrating them with LLMs by consuming Generative AI and LLM applications through APIs and tailor them, to a small degree, for your own use cases through prompt engineering techniques such as prompt tuning and prefix learning. LIMITATIONS RECOMMENDATION Not appropriate where the model needs to have a wide-ranging understanding of the content in the knowledge base, as only a limited a mount of data can be passed to the LLM. An affordable and powerful way to quickly leverage the power of generative Ai for your business BENEFITS Model trained on organisations data which is publicly not available . More affordable than organisations further training (“fine-tuning”) an LLM Data security as data resides in your own environment. An intermediate step for most businesses. The LLM will only use the data passed to it, along with the user’s original query, to construct a response. Customise - Benefits and Limitations katonic.ai
- 16. katonic.ai Build your Own Continuously Improve Enterprise Source Systems Add guardrails Embed Knowledge Add Skills Monitoring Your Enterprise Model LLM Training Stack
- 17. katonic.ai Build your Own - Examples Smaller 3B- 7B but specialised models can have a strong business value. Trained on a combination of web data that's already out there and internal Bloomberg data. Example: BloombergGPT Bloomberg trained a 50B LLM on combination of web data + internal Bloomberg data Outperforms existing open source models on finacial tasks Example: BioMedLM Growing evidence that training domain - specific LLMs ( medical, legal, etc ) are more accurate than using a generic LLM Example: BioMedLM, a 3B parameter LLM only on PubMed publications
- 18. Organisations training their own LLM gives them a deep moat: superior LLM performance either across horizontal use cases or tailored to your vertical, allowing you to build a sustainable advantage, especially if you create a positive data/feedback loop with LLM deployments. katonic.ai LIMITATIONS RECOMMENDATION Very expensive endeavor with high risks. Need cross-domain knowledge spanning from NLP/ML, subject matter expertise, software and hardware expertise. Best if you need to change model architecture or training dataset from existing pre-trained LLMs. BENEFITS Specialised models are smaller and can be deployed on significantly cheaper hardware Specialised models are significantly more accurate for the same resource budget Gain full control of training datasets used for the pre-training, Typically, you have or will have lots of proprietary data associated with your LLM to create a continuous model improvement loop for sustainable competitive advantage Less efficient than Customise option as it leverages existing LLMs, learning from an entire internet’s worth of data and can provide a solid starting point Build your Own - Benefits and Limitations
- 19. What is the role of MLOps? katonic.ai
- 20. katonic.ai AI Does not work out of the Box Generic Bespoke Domain Complexity Prototype Production Quality requirements AI works "out of the box" Complex, high accuracy applications Developement required
- 21. Complex Use Cases Requires Significant Development GPT-4 BERT Clip GPT-4 Domain Use case Fortune 500 pharma Information extraction Image classification Global ecommerce Chat intent classification Top US bank Document classification Legal data case study katonic.ai 60% 60% 43% 59% Foundation model performance
- 22. Complex Use Cases Requires Significant Development GPT-4 BERT Clip GPT-4 Domain Use case Foundation model performance Fine Tuned model performance Fortune 500 pharma Information extraction Image classification Global ecommerce Chat intent classification Top US bank Document classification Legal data case study katonic.ai 60% 86% 60% 85% 43% 71% 59% 83%*
- 23. Select a foundation model katonic.ai Define the task Classification Entry Extraction Translation Other ... and more. Check model license PowerML Fine tune full model or modify only last layers; Other layers are frozen Or use other strategy Use the validation data set to evaluate the performance of the model Repeat previous steps until you achieve satisfactory results. Use model on held- out test set to confirm its performance on unseen data Model architecture Hyperparameters Optimizer Set Up Run through multiple epochs Adjust hyperparameters Train model on your specific tasks/data Monitor loss/accuracy on validation set. Gather, pre- process, label Split new data: Training set Validation set Prepare Data Choose a fine-tuning strategy Configure the model Fine-tune the model Evaluate Iterate and improve Test the model Deploy or use the model Key Steps To Derive Value Out Of Generative AI
- 24. Data Preparation: Create training data & continuously update data. Model Training: Feed the data into a model for training Model Deployment: Deploy trained models into production (live). Model Monitoring: Monitor models for performance, accuracy, data sways or data drifts. Automation: Automate model for retraining, version control, rollback or update basis performance. [ [ [ [ [ Data Preparation Model Training Model Deployment Model Monitoring Automation katonic.ai An MLOps platform allows you to manage this complete process end to end with high accuracy, reliability and efficiency Role Of MLOps
- 25. How can Katonic Help? katonic.ai
- 26. Creative scientific process of data scientists + = Professional software engineering process Releasing ML Models into production safely, quickly, and in a sustainable way. katonic.ai Unified Platform powering your Generative AI strategy
- 27. [ [ [ [ [ Data Preparation Model Training Model Deployment Model Monitoring Automation Deploy with few Clicks Katonic LLM Playground Test best-in-class foundation models for your business and your specific data to build sustainability. Katonic has curated over 70 + popular LLMs for you to effortlessly experiment and prototype flows with drag-and-drop components. Katonic partners or integrates with all of the leading AI models, from open-source to closed- source.Deploy from our curated list of LLMs or any open-source LLM in a few clicks securely on your infrastructure Develop, test, and productionize - all in one platform. The only full- stack platform for powering your Generative AI strategy—including Data preparation, Distributed training , fine-tuning, security, model safety, model evaluation, and enterprise apps. One Unified Platform katonic.ai How Katonic can help accelarate your LLM Journey
- 28. Achieve Higher Accuracy Easily parallelize and distribute workloads across multiple nodes and GPUs. Bundled with State of the art tools and techniques Fastest Performance at Scale Easily access the capabilities of your custom LLM through Intuitive UI. Ease of Use Public cloud, private data centres, bare metal, Kubernetes cluster — Katonic runs anywhere Run Anywhere Fully Supported by Katonic Experts every step of the way. Enterprise Support katonic.ai Katonic Advantage
- 29. Automate your cycle of intelligence Thank You katonic.ai