Is a question I’ve been getting more and more at job interviews over the past year and when I say yes we’ve been using it for almost two years I see happy faces.
So if you’re a Software Engineering Leader there’s really no (business) reason not to get GitHub Copilot (or any other AI Coding Assistant) for your developers – it will (soon) be a requirement by new hires.
Recently GitHub released the GitHub Copilot Metrics API which provides customers the ability to view how Copilot is used and as usual someone created an Open Source tool to view the data: github-copilot-resources/copilot-metrics-viewer.
So let’s take a look at the usage of Copilot in Software Engineering in Carlsberg from end of May to end of June 2024.
I’m focusing on the following three metrics:
Total Suggestions
Total Lines Suggested
Acceptance Rate
As I think they are useful for understanding how effective Copilot is and I would like to get closer to an actual understanding of the usefulnes of Copilot rather than the broad statement offered by both GitHub and our own developers that it saves 50% of their time.
The missing data in the charts is due to an error in the GitHub data pipeline at the time of writing and data will be made available at a later stage.
The low usage in the middle of June is due to some public holidays with lots of people taking time off.
Total Suggestions
Total Lines Suggested: Showcases the total number of lines of code suggested by GitHub Copilot. This gives an idea of the volume of code generation and assistance provided.
Total Lines Suggested
Total Lines Accepted: The total lines of code accepted by users (full acceptances) offering insights into how much of the suggested code is actually being utilized incorporated to the codebase.
Acceptance Rate
Acceptance Rate: This metric represents the ratio of accepted lines to the total lines suggested by GitHub Copilot. This rate is an indicator of the relevance and usefulness of Copilot’s suggestions.
Conclusion
The overall acceptance rate is about 20% which resonates with my experience as Copilot tends to either slightly miss the objective and/or be verbose so that you have to trim/change a lot of code. So if Copilot suggests 100 lines of code you end up accepting 20.
Does this then align with the statements from developers in Software Engineering and GitHub which claim that you save 50% of time using Copilot?
Clearly reviewing and changing code is faster than writing, so even if you end up only using 20% of the suggested code, you will save time.
Unfortunately we don’t track actual time to complete tasks in Jira, so we don’t have hard data to prove the claim.
This is an attempt at trying to create some highlevel patterns for Artificial Intelligence (AI) solutions in order to be able to more easily choose a pattern based on type and problem area.
Chatbots with LLMs: Automate interactions and provide instant, contextually relevant responses, enhancing customer service, information retrieval, and user engagement across various domains like e-commerce and healthcare.
Chaining LLMs: Link multiple LLMs in sequence, leveraging their specialized capabilities for more nuanced and accurate solutions, enabling sophisticated workflows where each model performs tasks it excels at.
State machines and directed graphs: This approach introduces cycles, meaning the system can loop back and reconsider previous steps, which allows for more complex decision-making and adaptive behaviors. The state machine can maintain a state based on previous interactions, which can influence future decisions and actions.
Orchestrating LLMs: Simplify the integration and management of multiple LLMs to work in harmony, improving the development, performance, and scalability of AI-driven applications by leveraging the strengths of diverse models.
Would be a neglect not to mention RAG here although more of a feature than a solution pattern:
Retrieval Augmentation Generation (RAG): RAG combines the power of LLMs with a retrieval mechanism to enhance response accuracy and relevance. By fetching information from a database or collection of documents before generating a response, RAG models can provide answers that are more detailed and contextually appropriate, drawing from a wide range of sources. This approach significantly improves performance on tasks requiring specific knowledge or factual information, making RAG models particularly useful for applications like question answering and content creation.
Graphs and orchestration is also commonly referred to as “Agentic” architectures.
Chatbots with LLMs
Bots and chat-based interfaces powered by Large Language Models (LLMs) address a wide array of problem areas by automating interactions and processing natural language inputs to provide instant, contextually relevant responses.
These AI-driven solutions revolutionize customer service, information retrieval, and interactive experiences by enabling scalable, 24/7 availability without the need for human intervention in every instance.
They excel in understanding and generating human-like text, making them ideal for answering queries, offering recommendations, facilitating transactions, and supporting users in navigating complex information landscapes.
Furthermore, they significantly enhance user engagement by providing personalized interactions, thereby improving satisfaction and efficiency in areas such as e-commerce, education, healthcare, and beyond. By harnessing the capabilities of LLMs, bots and chat interfaces can decode intricate user intents, engage in meaningful dialogues, and automate tasks that traditionally required human intelligence, thus solving key challenges in accessibility, scalability, and automation in digital services.
Chaining LLMs
Chaining LLMs involves linking multiple LLMs in sequence to process information or solve problems in a stepwise manner, where the output of one model becomes the input for the next. This technique utilizes the specialized capabilities of different LLMs to achieve more complex, nuanced, and accurate solutions than could be provided by any single LLM.
Through this approach, developers can create advanced workflows in which each model is tasked with a specific function it excels at, ranging from understanding context to generating content or refining answers. This method significantly enhances the effectiveness and efficiency of AI systems, allowing them to address a wider variety of tasks with greater precision and contextual relevance. Chaining LLMs thus represents a strategic approach to leveraging the complementary strengths of various models, paving the way for more intelligent, adaptable, and capable AI-driven solutions.
Chaining LLMs is particularly effective for solving problems that benefit from a multi-step approach, where each step might require a different kind of processing or expertise. Here are some examples of problems typically solved using chaining:
Complex Query Resolution: Simplifying and addressing multifaceted queries through a stepwise refinement process.
Content Creation and Refinement: Generating drafts and then improving them through editing, summarization, or styling in successive steps.
Decision Support Systems: Deriving insights and suggesting actions through a sequential analysis and decision-making process.
Educational Tutoring and Adaptive Learning: Providing personalized educational feedback and instruction based on initial assessments.
These examples highlight the versatility of chaining LLMs, enabling solutions that are not only more sophisticated and tailored but also capable of handling tasks that require depth, precision, and a layered understanding of context.
Directed Graphs
State machines (a directed graph) are abstract machines that can be in exactly one of a finite number of states at any given time. In the context of LLMs and LangChain, a state machine would manage the flow of interactions with the LLM, keeping track of the context and state of conversations or processes.
LangGraph is designed to facilitate the creation, expansion, and querying of knowledge graphs using language models. It uniquely combines the representational power of knowledge graphs—structures that encode information in a graph format where nodes represent entities and edges represent relationships between entities—with the generative and understanding capabilities of language models. This integration allows for sophisticated semantic reasoning, enabling applications to derive insights and answers from a rich, interconnected dataset.
The primary value of LangGraph lies in its ability to leverage the contextual awareness and depth of language models to enrich knowledge graphs. This makes it particularly well-suited for applications requiring complex query answering, semantic search, and dynamic knowledge base expansion. It’s about not just processing language but understanding and organizing information in a way that mirrors human cognition.
Orchestrating LLMs
A framework for orchestrating LLMs is aimed at tackling the intricate challenges of integrating and managing multiple LLMs to work in harmony. Such a framework simplifies the process of combining the capabilities of diverse LLMs, enabling developers to construct more complex and efficient AI-driven solutions. It offers tools and methodologies for seamless integration, enhancing the development process, and allowing for the creation of applications that leverage the strengths of various LLMs. This not only streamlines the development of sophisticated applications but also boosts their performance and scalability, facilitating the customization of AI solutions to meet specific needs and contexts.
Orchestrating LLMs involves coordinating multiple models to work together efficiently, often in parallel or in a dynamic sequence, to tackle complex tasks. This approach is particularly useful for problems that benefit from the combined capabilities of different LLMs, each bringing its unique strength to the solution. Here are some examples of problems typically solved using orchestration:
Multi-domain Knowledge Integration: Coordinating specialized LLMs to offer solutions that require expertise across various fields.
Personalized User Experiences: Dynamically combining LLM outputs to customize interactions according to user data.
Complex Workflow Automation: Utilizing different LLMs for distinct tasks within a broader workflow, optimizing for efficiency and effectiveness.
Advanced Customer Support Systems: Integrating various LLMs to understand, process, and respond to customer inquiries in a nuanced and effective manner.
Orchestration enables the leveraging of multiple LLMs’ strengths in a coordinated manner, offering solutions that are more versatile, scalable, and capable of addressing the multifaceted nature of real-world problems.
By integrating GitHub Copilot into our development workflow, Carlsberg has significantly enhanced the developer experience. Copilot acts as an intelligent coding assistant, offering real-time suggestions and code completions. This seamless integration enables our developers to write more efficient and error-free code. From a business perspective, this translates to accelerated development cycles and a boost in productivity, allowing us to bring innovations to market faster and maintain a competitive edge.
Understand Code Faster
GitHub Copilot transcends simple code suggestions by providing developers with the ability to quickly understand existing codebases and even entire projects. This feature is invaluable for onboarding new team members and tackling complex legacy systems. By asking Copilot to explain intricate code, developers can rapidly grasp functionality without deep-diving into documentation or consulting peers. For Carlsberg, this means reduced ramp-up times for new projects and more efficient utilization of developer time, leading to cost savings and faster project deliveries.
Spend Less Time on Scaffolding
Scaffolding, while necessary, often consumes valuable time that could be better spent on developing business-critical features. GitHub Copilot streamlines this process by generating the foundational code structures automatically. This allows our developers at Carlsberg to concentrate on crafting the unique aspects of our solutions that drive real business value. The direct result is a more agile development process, with resources optimally allocated towards innovation and creating competitive advantages.
Lower the Learning Curve
Adopting new frameworks and technologies is a constant challenge in the fast-paced tech environment. GitHub Copilot lowers the learning curve for our developers by suggesting how to effectively use new frameworks. This guidance reduces the time spent on trial and error, enabling our team to leverage the latest technologies confidently. For Carlsberg, this capability ensures that we are always at the forefront of technology adoption, enhancing our agility and ability to respond to market changes swiftly.
Reduce Monotonous Work
Monotonous Work, like writing unit tests, though critical for ensuring code quality, can be tedious and time-consuming. GitHub Copilot addresses this by generating unit tests, which developers can then review and refine. This automation not only speeds up the development process but also ensures a high standard of code quality. At Carlsberg, leveraging Copilot for unit testing means our developers can focus more on developing features that add value to the business, while still maintaining a robust and reliable codebase.
Improve Documentation
Well-crafted documentation is crucial for maintainability and scalability but is often overlooked due to the time it requires. GitHub Copilot aids in this aspect by automatically generating meaningful comments and documentation during code commits or pull request reviews. This not only saves time but also enhances the quality of our documentation, making it easier for developers to understand and work with our code. At Carlsberg, improved documentation directly translates to reduced maintenance costs and smoother collaboration among teams, further driving operational efficiency.
At Carlsberg, our integration of GitHub Copilot into our development workflow has not just been about improving individual elements of the coding process—it’s about a holistic enhancement of the overall developer experience.
GitHub frames Developer Experience as the sum of Productivity, Impact, and Satisfaction. Here’s how Copilot aligns with these components:
Productivity: By automating and accelerating parts of the development cycle, Copilot directly boosts productivity. In the “Spend Less Time on Scaffolding” and “Reduce Monotonous Work” sections, we explored how Copilot streamlines tasks that traditionally consume significant time and resources. This allows our developers to focus on higher-value work, speeding up our overall project timelines and making our workflow more efficient.
Impact: The true measure of any tool or process change is the impact it has on the business and its goals. As discussed in “Understand Code Faster” and “Improve Documentation,” Copilot helps our team tackle complex systems more effectively and maintain better documentation. This not only enhances our current projects but also secures our long-term ability to adapt and grow, significantly impacting our operational success and market competitiveness.
Satisfaction: A satisfied developer is a productive and innovative developer. Through features like lowering the learning curve for new technologies and reducing the drudgery of repetitive tasks, as highlighted in “Lower the Learning Curve” and “Reduce Monotonous Work,” Copilot increases job satisfaction. This leads to a more engaged team, ready to innovate and push boundaries in pursuit of new solutions.
By investing in tools that elevate these aspects of the developer experience, Carlsberg is not just improving our software; we are fostering a culture of efficiency, innovation, and satisfaction. This commitment not only enhances our current team’s morale and output but also positions us as a forward-thinking leader in leveraging technology to drive business success.
Conclusion
GitHub Copilot has revolutionized the way we approach software development at Carlsberg, significantly enhancing the overall developer experience. By automating repetitive tasks, simplifying complex codebases, and expediting the learning process for new technologies, Copilot has allowed our developers to focus on what they do best: creating innovative solutions that drive real business value. This not only leads to a more satisfied and engaged development team but also accelerates our time-to-market and improves our competitive stance. The integration of GitHub Copilot into our workflow is a testament to Carlsberg’s commitment to leveraging cutting-edge technology to foster a culture of efficiency, innovation, and continuous improvement. It’s clear that by investing in tools that enhance the developer experience, we’re not just improving our software; we’re building a stronger foundation for our business’s future success.
When driving value from generative AI (GenAI) it’s important to choose the right approach in order to be able to get a return on investment. This page attempts at explaining possible approaches and required resources.
Takers, Shapers and Makers
There seems to be 3 major categories of GenAI adopters according to McKinsey and Gartner:
McKinsey
Gartner
Description
Takers
Quick Wins
Focus on utilizing existing GenAI tools and models for productivity improvements with minimal customization.
These initiatives typically have short time to value and are task-specific, aiming for immediate efficiency gains in routine tasks.
Shapers
Differentiating Use Cases
Engage in integrating GenAI tools with proprietary data or adapting them for specific applications.
These initiatives aim to achieve competitive advantages, involving medium time to value with higher costs and risks than quick wins.
They leverage GenAI to extend current processes and create unique value propositions.
Makers
Transformative Initiatives
Concentrate on developing new GenAI models or tools for specialized applications, with the potential to transform business models and markets.
These are the most ambitious initiatives, characterized by high cost, complexity, and risk, and a long time to value.
They aim for strategic benefits that may be difficult to quantify initially.
TCO/ROI
The Total Cost of Ownership (TCO) and Return on Investment (ROI) for GenAI adoption across takers, shapers, and makers categories involve several considerations, including hidden costs, strategic implications, and potential benefits.
Gartner offers insights on measuring GenAI ROI, advocating for a business case approach that simulates potential cost and value realization across GenAI activities. This approach categorizes investments into quick wins, differentiating use cases, and transformational initiatives. Quick wins focus on immediate productivity improvements with short time to value, differentiating use cases aim at competitive advantage with medium time to value, and transformative initiatives have the potential to upend business models with longer time to value but higher costs and complexity. The guide emphasizes the importance of balancing financial returns with strategic benefits, which might be difficult to quantify initially.
Source: https://www.gartner.com/en/articles/take-this-view-to-assess-roi-for-generative-ai. Red box is added by me, see conclusion below.
Builders
I’m introducing an extra “Builders” category into the GenAI adoption landscape beyond merely adopting or adapting, Builders take a step further by crafting bespoke extensions and plugins for GenAI platforms. This initiative is driven by the ambition to tackle intricate, multi-step workflows that typically demand considerable human intervention. The essence of being a Builder lies in their ability to not just work with GenAI but to enhance its core capabilities, enabling solutions that seamlessly bridge various systems and processes. This approach demands a blend of creativity, technical prowess, and a deep understanding of both the technology and the problem domain.
Category
Description
Required People Resources/Skills
Tools
Takers
Utilize existing GenAI tools for productivity improvements with minimal customization.
Aimed at immediate efficiency gains in routine tasks with short time to value.
Basic understanding of AI/ML conceptsSkills in integrating and configuring APIs
Ability to adapt third-party GenAI tools to existing workflows
Microsoft Copilot
Microsoft Copilot Plugins
Enterprise “Chat”-GPTs
Shapers
Integrate GenAI tools with proprietary data or adapt them for specific applications to achieve competitive advantages, involving medium time to value with higher costs and risks.
Low/No-code developers
Domain experts for data interpretation
Project managers with a technical background
Retrieval Augmented Generation (RAG)
Microsoft Copilot Studio
Microsoft Azure AI Studio
Builders
Develop custom solutions or extensions to GenAI platforms to solve complex, multi-step processes that usually require significant human effort.
Advanced programming skills in relevant languages
Data scientists for model tuning
Experience with GenAI frameworks
Systems integration expertise
Creative problem-solving abilities
Microsoft Copilot Extensions
Microsoft PromptFlow
LangChain
LangGraph
LlamaIndex
AutoGen
CrewAI
(OpenAI Swarm)
LLM Function Calling
LLM Routing
LLM Threat Modelling
LLM Security
Makers
Develop new GenAI models or tools for specialized applications with the potential to transform business models and markets.
Characterized by high cost, complexity, and risk, with a long time to value.
Expertise in deep learning and neural networks
Experience in building and training large-scale AI modelsStrong research and development background
Ability to work with high-performance computing resources
LLM Models
LLM Frameworks
LLM Fine-Tuning
(LLM Creation and Training)
The “Builders” category fills the gap between “Shapers,” who mainly adapt existing models for their unique needs, and “Makers,” who create new GenAI models from scratch. Builders leverage powerful frameworks and platforms to create bespoke solutions that automate complex workflows, potentially revolutionizing how businesses approach process automation and efficiency. This distinction underscores the evolving landscape of GenAI adoption, highlighting the increasing sophistication and customization capabilities available to organizations.
Conclusion
The red box on the image above indicates that solutions made in the Takers and lower Shapers category are likely to be overtaken by standard solutions from vendors and the plethora of SaaS AI offerings appearing on a daily basis. Caution should be used when choosing to invest in solutions in this area unless quick wins are important.
Clearly it’s important to have a strategic, well-planned approach to integrating GenAI with emphasis on organizational readiness, skill development, and a focus on applications that offer a competitive advantage – otherwise GenAI just becomes a technology looking for a problem like Blockchain.
I finally got access to ChatGPT Vision and I uploaded a picture of some beers I had at home:
This is the response from ChatGPT:
Nice selection of beers you got there! I see Perlenbacher, which is often associated with Lidl supermarkets, and of course, the iconic Carlsberg. Carlsberg, founded in 1847, is one of the most well-known beer brands globally and is based out of Copenhagen, Denmark. “Birra Messina” seems to be a Sicilian beer. Each of these beers brings a different taste profile to the table. Enjoy responsibly! Would you like to know more about any specific beer?
I then prompted this using TypeChat:
Please attempt to recognize the cans in the picture and fill out the following response: interface Response { data: Array<{ name: string, brand: string, count: number}>; }
No training – no nothing. I know the internet is full of examples of people having ChatGPT turning Figma drawings into web sites and apps – but this is totally out of the comfort zone of ChatGPT and I have previous experience trying to teach image recognition of packaged goods to models and it’s not that easy.
Solutions based on this could destroy the incumbent vendors in image recognition in the CPG market.
Christian says that he doesn’t want a future where a Clippy has to talk to a Clippy which talks to another Clippy and so on. In this context Clippy is some sort of an “AI” which can provide some answer.
I don’t know if that is true or not or if it’s the future we want or not – but right now, that seems to be where we’re headed.
LangChain is about chaining prompts and answers together so that we can get better answers
Microsoft Copilot Plugins is about enabling Microsoft Copilot to extend your prompts to other systems and consolidate the answers and perform tasks on your behalf
Microsoft TypeChat is about providing types and structure to the data we want to pass between systems to make interoperability even easier and more precise
I’ve spent the first 20 years of my life as a developer and the last 20 or so as an “architect” in many different contexts and it almost always boils down to solving integration between business domains and figuring out how data can be accessed, translated and acted upon across the different domains either in real time or before some date and time.
I totally agree with the sentiment that most, if not all, business systems are really just a database with a UI on top and that actual business value comes from combining the data across all systems and extracting answers from it. However, as long as the “business” continues to lock data in silos and refuses to adapt business processes to standard systems and demands customization of systems to accommodate the last 20% of functionality in a way that almost makes integration impossible – then I think the Clippies are going to be around for some time to come.
I think short to medium term the scenarios with plugins to Microsoft Copilot (and similar) will prove valuable to most businesses. Longer term we still need to fix copyright (AIs will be stopped from scraping), privacy (business are scared of leaks) and authentication/authorization (who are you and what do you have access to across boundaries). For instance how can we safely solve problems where data is scattered across private and public domain.
It’s going to be interesting and some new rules/processes are going to have to be worked out.
