Yolo Backend Login Guide: How to Set Up a Secure Login System
Yolo Backend Login Guide: How to Set Up a Secure Login System
Published on February 28, 2025 | Dive into Day 2 of the AI Engineer Summit 2025, focusing on agent engineering with insights from industry leaders. Introduction to Agent Engineering at AI Engineer Summit 2025 Welcome to our recap of Day 2 at the AI Engineer Summit 2025, held in New York City and live-streamed to over 13,000 viewers worldwide! Hosted by NLW (CEO of Superintelligent) and Ksenia Sova (founder of Turing Post), this event shifted focus from leadership (Day 1) to the builders shaping the future of AI agents. With a lineup featuring experts from Google, Anthropic, OpenAI, and more, Day 2 explored the practical and theoretical aspects of agent engineering. Why agents? They’re the hot topic of 2025, bridging decades of AI research with today’s scalable, real-world applications. Key Themes of Agent Engineering The summit highlighted several core themes: From Theory to Practice: AI agents, rooted in 1950s symbolic AI and 1980s multi-agent systems, are now operational thanks to LLMs and automation frameworks. Real-World Use Cases: Hear how agents are transforming finance (e.g., Jane Street, BlackRock) and beyond. Open Challenges: Scaling, accuracy, and memory remain hurdles for AI engineers to tackle. “What’s actually happening now—things that are real, live, not just theoretical—is a key focus,” said NLW. Opening Remarks: Setting the Stage with Swyx The State of AI Engineering Swyx, co-founder of the summit and editor of Latent Space, kicked off the day by contextualizing agent engineering in 2025. He noted the discipline’s maturation, with AI engineering emerging as distinct from MLE and software engineering. “2025 is the year of agents—if we say it enough, it might come true,” he quipped, acknowledging skepticism but pointing to real progress. Why Agents Now? Swyx outlined why agents are gaining traction: Improved Capabilities: Better reasoning, tool use, and model diversity (e.g., OpenAI’s market share dropping from 95% to 50%). Cost Reduction: GPT-4-level intelligence costs have dropped 1,000x in 18 months. RL Fine-Tuning: Reinforcement learning is enhancing agent performance. He also highlighted use cases like coding and deep research, urging engineers to avoid overused demos like flight-booking agents. Sayash Kapoor: Challenges in Building Reliable Agents Sayash Kapoor, author of AI Snake Oil, challenged the audience to address agent reliability. He identified three key issues: 1. Evaluation Difficulties Examples like DoNotPay’s exaggerated claims and Lexus Nexus’s hallucination-prone legal tools show that evaluating agents is tough. Kapoor’s team at Princeton found top agents struggle to reproduce scientific papers (less than 40% success on CoreBench). 2. Misleading Static Benchmarks Unlike LLMs, agents interact with dynamic environments, requiring cost and multi-dimensional metrics. Princeton’s Holistic Agent Leaderboard (HAL) evaluates accuracy alongside cost, revealing trade-offs (e.g., Claude 3.5 vs. OpenAI’s o1). 3. Capability vs. Reliability Capability (what an agent can do) doesn’t guarantee reliability (consistent success). Kapoor likened this to early computing’s reliability struggles, urging a shift to reliability engineering. “Closing the gap from 90% to 99.999% reliability is the AI engineer’s job,” Kapoor emphasized. Google’s Gemini Deep Research: Building a Research Agent Mukund Sundararajan (Software Engineer) and Arush Sankholkar (Product Manager) from Google shared how they built Gemini Deep Research, a web-browsing research agent available in Gemini Advanced. Motivation and Challenges The goal: comprehensive answers for complex queries (e.g., “How do I get a shot put scholarship?”). Challenges included: Asynchronous UX: Adapting a synchronous chatbot for 5-minute research tasks. User Expectations: Communicating when deep research is worth the wait. Long Outputs: Making thousand-word reports digestible. Solutions Google’s approach: Research Plan: Users see and edit a plan before research begins. Transparency: Real-time display of browsed websites. Artifacts: Pin reports for follow-ups, with sources cited. Technical Hurdles Building a robust agent required: State Management: Handling failures in long-running tasks. Iterative Planning: Reasoning with partial web data. Context Management: Balancing growing context with recency bias. Future Directions: Expertise-driven insights, domain-specific outputs, and multimodal capabilities (e.g., coding, data analysis). Anthropic’s Barry Jang: Building Effective Agents Barry Jang from Anthropic shared practical lessons from their blog "Building Effective Agents": 1. Don’t Build Agents for Everything Agents excel in complex, high-value tasks (e.g., coding), not simple ones. Use workflows for low-cost, predictable scenarios. 2. Keep It Simple Agents are models with tools in a loop. Overcomplicating risks cost and latency spikes. 3. Think Like Your Agent Anticipate how agents interpret tasks and environments to avoid errors. “Coding agents succeed because they’re verifiable and valuable,” Jang noted. Sumith Chintala: Personal Local Private AI Agents Sumith Chintala, co-founder of PyTorch at Meta, explored building personal local AI agents for deep augmentation. Why Local and Private? Chintala argued for local agents due to: Context: Agents need full personal data (e.g., Gmail, WhatsApp) for reliability. Control: Cloud services risk unpredictable actions or monetization biases. Privacy: Avoid “thought crime” risks with sensitive queries. Technical Challenges Running agents locally (e.g., on a Mac Mini) faces hurdles: Slow Inference: Local models lag behind cloud APIs. Multimodal Limits: Open models struggle with nuanced tasks (e.g., shopping). Catastrophic Actions: Lack of classifiers for irreversible errors. Bullish Outlook: Open models (e.g., LLaMA, Grok) are compounding intelligence faster than closed ones, promising a local agent future. Conclusion Day 2 of the AI Engineer Summit 2025 showcased the promise and pitfalls of agent engineering. From Google’s practical research tools to Anthropic’s simplicity mantra and Chintala’s vision for local agents, the event underscored 2025 as a pivotal year for AI builders. Catch the full talks on YouTube and join the conversation! Stay tuned for more AI insights as we head toward the 2025 World’s Fair!
