# From Software Developer to AI Engineer Antje Barth, AWS | AI Engineer World's Fair 2024 | 19:48 Source: https://www.youtube.com/watch?v=90Ir4TMMbS0 Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/from-software-developer-to-ai-engineer Published: 2024-07-24 Tags: agent-skills, agents, code-generation, rag, tool-use ## TL;DR - AI engineers need a working grasp of foundation models, customization, function calling, and responsible AI practices. - Developer tools such as Amazon Q Developer can handle code explanations, code generation, feature work, and code transformations. - Building AI applications requires choosing models carefully, using services such as Amazon Bedrock, and learning how to create agent workflows. ## Summary Antje Barth presents five practical steps for moving from software development into AI engineering. She starts with the technical basics, including how foundation models work, why their outputs can be unexpected, and how to customize them. She then shows how AI developer tools can reduce routine work through Amazon Q Developer, with examples covering code explanation, function generation, and larger planned changes. The talk moves into prototyping with AI, where developers choose models, consider prompting, retrieval-augmented generation, or fine-tuning, and add privacy controls and guardrails. Barth explains how Amazon Bedrock gives applications access to multiple foundation models through a unified Converse API. A Minecraft agent demo illustrates how tools let an agent reason through actions and solve a task. She closes by encouraging developers to stay current through community events and learning resources. ## Key ideas ### AI engineers still need to understand how models behave [00:38](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=38s) Barth says developers no longer need to be full machine learning researchers or data scientists to put AI projects into production. They do need a basic understanding of foundation models, including why a model can produce an unexpected result inside application code. She also names model customization, fine-tuning for specific use cases and datasets, and functions that give an application access to other systems. For people starting out, she recommends the course Generative AI with Large Language Models, created with Andrew Ng and the DeepLearning.AI team. The course is available through DeepLearning.AI and Coursera. ### Natural-language tools can reduce routine software work [02:14](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=134s) Barth describes a software development workflow in which much of the time goes to boilerplate code, documentation, and maintaining old codebases. She says this can reach up to 70 percent of the work, leaving a smaller share for creative tasks. Natural language can now be used in both directions: developers can ask AI to turn English into code, such as rewriting a README, or turn code into English by documenting functions. AWS created Amazon Q Developer around this shift. Barth presents it as a development assistant that uses agents for work beyond ordinary code completion. ### Amazon Q Developer can plan and apply larger code changes [05:01](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=301s) The Amazon Q Developer demo begins with a developer asking how to create a serverless application. Q returns instructions that include installing the AWS SAM CLI and creating a project. It can explain generated code or unfamiliar legacy code when the developer sends a selection from the IDE. For larger tasks, the developer uses the /dev command to request a DynamoDB search-by-category function with detailed output requirements. Q proposes files such as template.yaml and search-by-category.mjs, then produces a change set showing the difference from the current project. The developer can review it and insert the code. ### AI prototypes require choices about models, data, and controls [08:28](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=508s) Barth describes prototyping as a sequence of decisions that starts with defining the use case. Developers then compare models and decide whether prompting, retrieval-augmented generation, or fine-tuning fits the application and its data. Responsible AI policies belong throughout the development workflow. Barth specifically calls for private and secure data, along with guardrails in the application. She also mentions the operational questions that arise once agents are part of an application, including how to keep them up to date and how practices such as GenAIOps, MLOps, and LLMOps apply. ### Amazon Bedrock lets developers compare models through one managed service [09:32](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=572s) Barth argues that applications will usually need more than one model because different use cases have different requirements. Amazon Bedrock gives developers access to foundation models from providers including AI21 Labs, Anthropic, Cohere, Meta, Mistral, Stability AI, and Amazon's Titan family. She says Bedrock also includes tools for customization, retrieval-augmented generation workflows, and agents, with control over application data. She calls out the recent addition of Claude 3.5 Sonnet on Bedrock. The service is presented as a place to experiment with and integrate several model choices. ### The Converse API standardizes calls across supported models [11:01](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=661s) Barth explains that Bedrock's unified Converse API uses a common invocation method across models. Developers can use the same message structure and inference parameters while Bedrock translates them into the format expected by the selected model. The API handles system, user, and assistant prompts and returns a consistent output format. Her Python example creates an SDK client, builds a list of messages, selects a model ID, and passes inference parameters in one call. Function calling is also supported through a tool list, where each tool gives the model access to a function. ### Agents can solve tasks by combining tools into a workflow [14:15](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=855s) The agent demo uses Amazon Bedrock Agents to control a Minecraft character called Rocky. The agent uses Claude 3 Haiku, instructions that describe how the bot should play, and action groups linked to code. Available actions include jumping, digging, finding a player's location, and moving to a location. When asked to come to the player, the agent identifies the requesting player, finds that player's location, and maps a path from its current position. Barth uses this example to show an agent combining several tools to solve a task even though there is no single tool named 'come to me'. ### Staying current requires active participation in the developer community [17:59](https://www.youtube.com/watch?v=90Ir4TMMbS0&t=1079s) Barth's final step is to stay up to date because the field is changing quickly. She announces that the AWS Loft in San Francisco is being turned into an AI engineering hub for workshops, events, and meetups. She invites attendees to suggest topics and to volunteer as speakers or meetup hosts. She also points people to additional Amazon Q sessions, the AWS booth, the Bedrock community tutorials, and the GitHub project behind the Minecraft agent. Her advice is practical: keep learning through resources and direct contact with other builders. ## Notable quotes - "You have to understand at a basic level how Foundation models work why they're sometimes producing output that you don't expect in your application code." (01:04) - "Amazon Q developer is much more than just code completion." (07:40) - "There's likely no one model to rule them all." (09:43) - "All of those things happened all in that blink of an eye and there's agentic workflows making all of that happen." (17:17) ## Tools & references mentioned - Amazon Q Developer - Amazon Bedrock - AWS SAM CLI - Generative AI with Large Language Models - DeepLearning.AI - Andrew Ng - Coursera - Claude 3 Haiku - Claude 3.5 Sonnet - AI21 Labs - Anthropic - Cohere - Meta - Mistral - Stability AI - Amazon Titan - Converse API - Minecraft - GitHub ## Who should watch - You are a software developer who wants a practical starting point for building applications with generative AI. - Your team is comparing foundation models or considering Amazon Bedrock for model access, customization, retrieval, or agents. - You want to see what an AI coding assistant can do with legacy code, generated functions, and larger planned changes. ## Related talks - [Developer Experience in the Age of AI Coding Agents](https://aietalks.com/talks/developer-experience-in-the-age-of-ai-coding-agents) (Max Kanat-Alexander, Capital One, 18:20) - [The 4 Patterns of AI Native Development](https://aietalks.com/talks/the-4-patterns-of-ai-native-development) (Patrick Debois, 14:11) - [How to Become an AI Engineer from a Fullstack Background](https://aietalks.com/talks/how-to-become-an-ai-engineer-from-a-fullstack-background) (Reid Mayo, Rema AI, 10:19) - [The 1,000x AI Engineer](https://aietalks.com/talks/the-1-000x-ai-engineer) (Swyx, AI Engineer Summit, latent.space, smol.ai, 09:27) - [AI Leadership](https://aietalks.com/talks/ai-leadership) (Alex Lieberman, Morning Brew and 10X & Kath Korevec, Google Labs & Katelyn Lesse, Anthropic & Michele Catasta, Replit & Lisa Orr, Zapier & Steve Yegge, Sourcegraph and AMP & Gene Kim, IT Revolution & Bill Chen & Brian Fioca, OpenAI & Martin Harrysson & Natasha Maniar, McKinsey & Yegor Denisov-Blanch, Stanford & Itamar Friedman, Qodo & Olive Song, MiniMax & Asaf Bord, Northwestern Mutual & Lei Zhang, Bloomberg & Samir Mody, The Browser Company & Max Kanat-Alexander, Capital One & Arman Hezarkhani, 10X & Justin Reock, DX & Dan Shipper, Every & Mel Lutzky, Graphite, 8:16:05)