AI Engineering Highlights

Scott Wu from Cognition presents Moore's Law for AI Agents - capabilities double every 70 days (16-64x annually). From tab completion to autonomous engineers in 18 months. Learn the 5-tier evolution framework: migrations → bug fixes → complex debugging → project autonomy, plus technical infrastructure including Playbooks, Deep Wiki, With Search, and integrations with Linear, Jira, and Slack.

Asaf Bord shares how a 160-year-old Fortune 100 insurance company built GenBI using 6-week sprints, continuous plug-pulling rights, and incremental value delivery. Learn the crawl-walk-run adoption strategy, why 80% of BI work is report routing (not SQL generation), and the honest assessment that executive-ready AI may never arrive.

Justin Muller, Principal Applied AI Architect at AWS, reveals the battle-tested 7 Habits framework that transformed document processing from 22% to 92% accuracy in 6 months. Learn why evals are the missing piece to scaling GenAI, the 30-second rule for rapid iteration, and how to build evaluation systems that enable production deployment with real-world case studies and practical implementation guidance.

Craig Wattrus from Flatfile presents a four-form-factor framework for AI coworkers: Invisible, Ambient, Inline, and Conversational. Learn why traditional design processes fail with LLMs and how playful experimentation leads to better AI products through "feeling the material" and character coaching over control.

Eric Simons shares how Bolt.new went from near-shutdown to $20M+ ARR in 60 days with just 15 people. Learn the Spartan mentality, community strategies, AI-powered support, and team culture that made it possible.

Full workshop on achieving unprecedented Llama 3 inference speeds of 1,000 tokens/second using SambaNova's Composition of Experts architecture and custom RDU hardware. Includes hands-on RAG implementation with LlamaIndex, ChromaDB, and performance benchmarks (16 chips vs 576, full precision).

Ola Mabadeje from Cisco's Outshift group reveals how they built a 5-agent system for network change management, why fine-tuning beat RAG for knowledge graph queries (drastic token reduction), and how knowledge graphs serve as digital twins for safe testing before production. Complete architecture with real quotes.

Kyle Penfound and Jeremy Adams from Dagger demonstrate building production-ready AI agents in a hands-on workshop. Learn guardrails, container-native development, GitHub integration, function calling patterns, and practical production-ready agent architecture.

Samir Mody from The Browser Company shares lessons from building Arc and Dia. Learn how they achieved 10x iteration speed, built Jeba for automated prompt optimization, embraced non-engineers as AI developers, and navigated security challenges in AI browsers.

From Death's Door to $100M ARR: The Three-Step Framework, Anti-Patterns to Avoid, and How to Create Categories Instead of Features

Michael Yuan explains why Rust is the perfect language for AI-generated code. Learn how the Rust compiler serves as a reward function for AI, the MCP tools that automate code generation, and the path to AGI through verifiable code with 1,000+ developers already using Rust Coder in production.

Jacob Kahn from Meta FAIR presents Code World Model (CWM) - a 32B parameter model that explicitly models program execution dynamics rather than just syntax. Learn how execution tracing, asynchronous RL with mid-trajectory updates, and "neural debugging" enable AI to simulate code execution without running it, effectively approximating solutions to the halting problem.

Steve Ruiz demos tldraw.computer - a visual programming language where LLMs execute through graph-based nodes. See multimodal computing, self-scripting nodes, and AI as collaborator returning structured shapes, not pixels.

Will Hang and Cathy Zhou from OpenAI introduce Agent RFT - the first time models can interact with the outside world during training. Learn how Cognition, Cosine, and Qodo achieved dramatic improvements with as few as 10 examples, the four success principles, and why parallel tool calling reduces latency from 8-10 steps to 4.

Why Architecture Copilots Deliver Higher ROI Than Coding Copilots: Preventing Costly Mistakes, Justifying Nine-Figure Infrastructure Spends, and Enabling Safe Delegation to Developers

AWS engineer demonstrates building Rocky, a Minecraft bot powered by Claude Haiku and Amazon Bedrock Agents. Live demo with emergent behavior, real-time gameplay, and practical lessons on architecture evolution from LangChain to Bedrock.

Natalie Serrino from Gimlet Labs on AI-Driven GPU Optimization: 25-70% Speedups, Agentic Synthesis Swarm, Hardware-in-the-Loop Verification, and the Path Forward for PTX Generation and Formal Verification

Scott Stephenson, CEO at Deepgram, explains the evolution from Siri-era Voice AI 1.0 to LLM-powered Voice AI 2.0, the Intelligence Revolution timeline (25-30 years), accuracy improvements (75% to 90%+), latency breakthroughs (2-5s to 100-200ms), and how contextual AI传递es conversation context to enable human-like voice interactions with <500ms roundtrip latency

Rhythm Garg and Linden Li from Applied Compute present efficient RL systems for enterprise applications. Learn about asynchronous vs synchronous RL, GPU utilization optimization, staleness trade-offs, system-level modeling, and first-principles optimization for end-to-end performance.

Logan Kilpatrick from Google DeepMind recaps a transformative year — 10 years of progress in 12 months, 50x inference growth, Gemini 2.5 Pro final update, organizational evolution, and what's next for universal assistant vision, omnimodal models, agentic AI, and developer platform expansions (Embeddings API, Deep Research API, Veo 3, Imagen 4, AI Studio repositioning).

Stephen Chin and Andreas Kollegger explore the fundamental obstacles to AGI through science fiction memes—from Memento's memory problem to The Matrix's simulation control. A concise 4-minute lightning talk covering memory limitations, alignment problems, transparency issues, and the ultimate question: do we know what to ask AGI?

Nathan Lambert's Four-Pillar Framework: Skills, Calibration, Strategy, Abstraction—10-100x Token Waste Problem, Post-Training RL Compute Revolution (1% → 10%+)

swyx Landmark Keynote: Why 2025 is the Year of Agents - 6 Enabling Factors, Agent Definitions, PMF Use Cases, ChatGPT Growth Analysis, and the Evolution of AI Engineering as a Discipline

8B = 235B Distillation Breakthrough, Doubled Reasoning Tokens, and the New Scaling Paradigm from Chinchilla to Inference

Neo4j MCP Tools, GraphRAG Architecture, and Enterprise Case Study with 85% Adoption

Pfizer & Neo4j Case Study: How GraphRAG Achieved 3 Months → 3 Weeks with Knowledge Graphs. Real Enterprise Lessons on Technology Transfer, Workforce Knowledge Retention Crisis (20 Years → 3 Years Tenure), and Navigating Organizational Politics.

John Crepezzi shares how Jane Street builds custom AI infrastructure when off-the-shelf tools won't work. Learn workspace snapshotting, Code Evaluation Service (CES) running 50-100x faster than builds, Aid sidecar architecture for multi-editor support, and why they have more OCaml code than exists publicly worldwide.

Why TDD is Critical for Reliable AI Agents: L0-L4 Agentic Workflow Framework, Evaluation Loops, and SEO Agent Demo with 60% Performance Improvement

Why AI Evaluation Finally Breaks Through: Three Converging Forces, C-Suite Alignment & Market Validation

Why Vibe Coding Will Transform Software Development

Comprehensive analysis of AI agent pricing models, cost structures, and the economics of intelligence — outcome-based pricing, prepaid credits, cost optimization strategies, and 2025 predictions from 13+ real company examples

How to Build Custom Knowledge Apps at Scale - Human-in-the-Loop Design, LLM Strategies, and Why Autonomous Agents Don't Work in Finance

Sarah Sachs (AI Lead, Notion) and Carlos Esteban (Braintrust) share their evaluation-first approach to building AI products. Learn why Notion spends 90% of time on evaluation and 10% on prompting, with practical guidance on data management, trace-based debugging, user feedback analysis, multi-turn conversation evaluation, and production monitoring with online scoring.

Ankur Goyal (CEO, Braintrust) shares hard-earned lessons: YAML vs JSON (15% token savings), GPT-4o 10% → Claude 4 Sonnet viable (6x better), Notion's 24-hour model integration, continuous reconciliation, and why great evals must be engineered like any other software system.

Production AI Methodology from HumanLayer - Transform Unreliable Demos into Dependable Systems

The Mathematical Formula for Agent Success: P(success) × Value - Cost(failure) > Cost(running)

What Self-Driving Taught Me About Building Agents: Agentics, 1% vs 99% Problem, and Closed-Loop Systems

Hands-on workshop exploring AI music generation tools (Udio, Suno, Stable Audio), voice cloning (RVC), stem separation (Wave, UVR5), and the RIAA legal battle. Learn practical workflows for generating professional-quality music from text prompts.

The Single-Process Security Flaw, Real-World Production Blocker, and Three-Layer Defense Framework from Apache Ranger's Creator - Don Bosco Durai, Priv

Production Reality vs. AI Hype in Software Engineering - Why Professional Engineers Are Skeptical and What Actually Works

Unified Platform: HNSW Algorithm, Search Nodes for Independent Scaling, Framework Integrations (LangChain, LlamaIndex), and Production-Ready RAG with 4,096 Dimensions Support

Google DeepMind's Vision for AI-Assisted Development

Greg Brockman (OpenAI President) & Jensen Huang (NVIDIA CEO) on the evolution from AlexNet to 100K GPU clusters, why technical humility matters, and the future of domain-specific AI agents. "If you don't have the idea, you're dead in the water. But if you don't have the engineering, that idea is not going to live and see the light of day." Learn about the 3-phase evolution of AI engineering at OpenAI, the cultural divide between engineers and researchers, and predictions for AGI-era development workflows.

Real Revenue, Validated Models: 25M Users, $100M ARR Across Portfolio. Meet Krea, OpenHome, Koframe, Federous AI, Upside, OpenAudio, Glow, Favored, and OpenRouter.

Why and How Modal Rebuilt Cloud Infrastructure from Scratch

Jellyfish Research: 2x Throughput, 24% Faster Cycles, and the Architecture Correlation That Determines AI Success (4x vs 0x Gains)

LLM-as-Judge Methodology with 4 Components: Role, Task, Context, Goal. Why Even OpenAI and Anthropic CPOs Say Models Hallucinate. Transition from Vibe Coding to Thrive Coding in Production.

MongoDB's Guide to Building Stateful AI Agents: Memory Management Lifecycle, Four Memory Types, and Voyage AI Integration

How Replit Broke the One-Hour Autonomy Barrier for Non-Technical Users

AWS Solutions Architect's live demo of Rocky, a Minecraft bot powered by Claude Haiku and AWS Bedrock, showcasing emergent AI behaviors including autonomous escape from a hole, 3D spatial reasoning, and the critical Return of Control pattern for production agents

5-15% Improvement with Only 150 Examples - A Practical Alternative to Reinforcement Learning

Lee Robinson reveals how Cursor built their first agent model with 4x token efficiency, parallel tool calling breakthrough, and RL infrastructure secrets. Learn about the 3.5x Blackwell speedup, semantic search impact, and vertical integration advantages.

Meta Neuro-Symbolic RL: 10x PR Increase with Acquired Taste

Scott Wu's Framework: 70-Day Doubling Cycle - From Tab Completion to Autonomous Engineers in 18 Months. Deep Wiki, Automated Testing, Backlog Processing, and the Future of Software Engineering

Diamond Bishop from Datadog shares what they learned building AI agents that automate on-call duties, handle incident response, and transform DevOps. Covers evaluation strategies, team composition, LLM observability, and predictions about agents surpassing humans as SaaS users.

OpenAI Guide to Building AI That Writes Code and Survives Model Evolution

Architecture Deep Dive: Flexible Loops, Skills System & Comparison with Cursor, AMP, and OpenAI Codex

A Practical Guide to Evaluating RAG Systems: Fast Evals, Cluster Analysis, and Data-Driven Decision Making

How K-Scale Labs Built a $9k Open-Source Humanoid in 5 Months: Sim-to-Real RL, Python SDK, and Democratizing Robotics

Jake Nations from Netflix argues that while AI has dramatically accelerated code generation, it has created a dangerous gap between what we can produce and what we can understand. He presents "context compression" as a three-phase solution (Research, Planning, Implementation) to maintain control over complex systems.

Luminal AI Automatically Rediscovered Flash Attention Using Search-Based Compilation—12 Primitives, E-Graphs, and 3M → 5K Lines of Code

Organizational Readiness for Autonomous AI Development

Why AI Writes Code Faster But Debugging Gets Harder - Three-Part Framework: Causal ML + LLMs + Swarms of Agents. DigitalOcean Case Study: 40% MTTR Reduction

Reinforcement Learning, Defense & Vertical Integration

Neil Dwyer reveals how to achieve $1/hr voice AI costs using Orpheus model, LoRA fine-tuning, vLLM with FP8, and consistent hash load balancing

Why Your Agent Optimization is Failing (And How to Fix It)

Practical Patterns from GitHub: Realistic Metrics (30% Average), IDE Integration Over Chat Tools, Human-in-the-Loop Autonomous Agents

How AI Code Evaluation Evolved from Single Functions to Hour-Long Challenges—and Why 30% is Reward Hacking

Multimodal Voice + Text Integration Using GPT-4o Realtime API - "Shipping the Org Chart" Anti-Pattern and the FaceTime + iMessage Solution

Multilingual Architecture with ElevenLabs: 31 Languages Now, 99 Coming in V3, $5M Voice Marketplace, Production-Grade Low-Latency Pipeline

Jacob Kahn from FAIR Meta presents a revolutionary framework for understanding code through execution modeling. Learn how 32B parameter models trained on execution traces enable semantic understanding, neural debugging, and approximation of undecidable problems like the halting problem—all through bash-first async RL with mid-trajectory model updates.

Systematic Framework for Reliable AI-Augmented Development: 5-Step Planning Phase, Multi-Sensory Feedback Loop, Binary Dependencies, and Circular Resolution Strategies That Transform AI Agents from Erratic Novices into Predictable Implementation Partners

From Localhost to Production: Taking Model Context Protocol Servers from Demo to Enterprise Deployments - Security Challenges, Compliance Requirements, and Implementation Realities

How LinkedIn replaced dozens of specialized models with a single LLM—achieved 7x latency reduction and 30x throughput improvement through promptification and model distillation

Niklas Nielsen from Log10 introduces llmeval - a command-line tool for reliable LLM evaluation built on Meta's Hydra. Learn about flexible test criteria for fuzzy model outputs, Python-based metrics, model-based evaluation with its pitfalls (self-preference bias, score inflation), and the innovative "Auto-John" concept for scaling human feedback through AI personas.

How Netflix proved scaling laws apply to recommendation systems—applying LLM techniques like multi-token prediction, long-context training, semantic embeddings, and rich multi-task objectives to achieve infrastructure consolidation and quality improvements

Justin Reock shares aggregated data from 140,000 engineers revealing extreme variability in AI impact (+20% to -20%) and provides a leadership framework for successful AI adoption. Learn why writing code has never been the bottleneck, how Theory of Constraints applies to AI, and the 7 leadership principles that separate +20% outcomes from -20% outcomes.

Mark Myshatyn from Los Alamos National Lab reveals how one of the most secure government organizations is deploying AI agents that design fusion capsules and execute code on HPC systems. Learn about 1000+ security controls, FedRAMP compliance, OpenAI models on classified networks, Venado supercomputer (2500+ GraceHopper nodes), and four architecture principles for government-ready AI. "We are not a t-shirt company...People can die if we do this wrong."