Conference Introduction & AI Engineering State 00:00
The conference opens with an overview, combining updates on the event and the broader state of AI engineering.
Attendance surged with 3,000 last-minute registrations, illustrating high interest and causing logistical stress.
The conference doubled the number of tracks from the previous year to cover more AI topics while acknowledging "decision fatigue."
Organizers aim to be more responsive and technically focused than competing events, using attendee surveys to shape content.
Attendees are encouraged to complete an ongoing survey to inform future conference planning.
The event is positioned at the frontier of AI engineering, likening the gathering to the 1927 Solvay Conference in physics.
Evolution of AI Engineering & Standard Models 03:01
The speaker reflects on the evolving focus of previous conference editions, noting a trajectory from defining AI engineer roles to emphasizing agent engineering.
AI engineering has gained status and profitability, transforming perceptions—referencing the success of simple GPT wrappers.
Simplicity is highlighted as a recurring lesson in effective AI solutions, citing Anthropic and Greg Brockman's remarks and simple scaffolds outperforming complex systems.
AI engineering is described as still being in an early, "emperor has no clothes" period, suggesting there remains substantial opportunity for innovation.
The industry is compared to the formative decades of physics (1940s–1970s) when foundational models were established, prompting a question: What will the "standard model" in AI engineering be?
The talk reviews existing standard models from other engineering fields (e.g., ETL, MVC, CRUD, MapReduce) and discusses their partial relevance to AI development.
The future of Retrieval Augmented Generation (RAG) is debated, with some claiming it is becoming outdated due to newer methods like long context and fine-tuning.
Several candidate standard models for AI engineering are introduced:
LLM OS: Originating in 2023, updated for multimodality and external tool connectivity via MCP.
LLM SDLC: Software Development Life Cycle frameworks are evolving, with early stages (like LLM and monitoring) now largely commoditized and available for free.
Business value and complexity emerge later, particularly in evaluation, security, and orchestration—areas now featured as conference tracks.
Building effective agents has a growing playbook, but definitions are still in flux across organizations (Anthropic, OpenAI).
The speaker argues for a top-down, descriptive model of "agent" terminology, highlighting concepts like intent, control flow, memory, planning, and tool use.
The distinction between "workflow" and "agent" in AI is seen as less important than the degree of value delivered through human input versus valuable AI output.
The evolution of AI output ranges from autocomplete systems (low input/output ratio) to "ambient agents" that provide value with minimal human input.
The speaker's own application ("AI news"), although not technically an agent, delivers value via a structured process: scrape, plan, recursively summarize, format, evaluate.
This repeated approach for different data sources is abstracted into a generalizable model for AI-intensive applications: Sync, Plan, Analyze, Deliver (SPAD).
AI engineering techniques include processing results into knowledge graphs, structured output, and code artifacts.
Examples given include ChatGPT with Canvas and Claude with Artifacts, representing the delivery of code as an AI output.