So, I was told that this could be between five and 55 minutes long, quite a range, and that I could yap and go on whatever side tangent I wanted, and it'll be fine. So, I'm just prepping you and giving you some explanation for why this is the way it is. Uh, so my name is Dax. I work on a project called Open Code. It is one of the most used coding agents out there. Uh, it's fully open source, works with any model. Um, every other company likes to say we have the smartest coding agent. Well, we've got the hot ...

Model Performance & Ranking - GLM 4.6 is currently ranked 1 on the LMSYS Chatbot Arena, on par with GPT-4o and Claude 3.5 Sonnet [1] - The GLM family of models has achieved over 100 million downloads [1] Training & Architecture - zAI utilized a single-stage Reinforcement Learning (RL) approach for training GLM 4.6 [1] - zAI developed the "SLIME" RL framework for handling complex agent trajectories [1] - The pre-training data for GLM 4.6 consisted of 15 trillion tokens [1] - zAI filters 15T tokens, moves to repo-level code contexts, and integrates agentic reasoning data [1] - Token-Weighted Loss is used for coding [1] Multimodal Capabilities - GLM 4.5V features native resolution processing to improve UI navigation and video understanding [1] Deployment & Integration - GLM models can be deployed using vLLM, SGLang, and Hugging Face [1] Research & Development - zAI is actively researching models such as GLM-4.5, GLM-4.5V, CogVideoX, and CogAgent [1] - zAI is researching the capabilities of model Agents and integration with Agent frameworks like langchain-chatchat and chatpdf [1]

AI Coding Agents & Tools - Focus on building and improving AI coding agents, covering topics from agent reinforcement fine-tuning to proactive agents [1] - Discussions on tools and platforms for AI-assisted coding, including code evaluation, world models for computation, and agent-ready codebases [1] - Exploration of using AI to speed up code execution and address software crisis [1] Reinforcement Learning (RL) in Coding - Research on efficient reinforcement learning and its application in coding environments at scale [1] - Agent Reinforcement Fine Tuning [1] - Building a fast frontier model with RL [1] Future of Software Development - Examination of the future of software development with AI, including continual system-prompt learning and the path towards AGI [1] - Investment trends in the future of software development [1] - Measurement gap between benchmarks and economics in AI capability [1] Codebase Management & Problem Solving - Strategies for solving hard problems in complex codebases [1] - Making Codebases "Agent-Ready" [1] - Transition from code snippets to codebases in coding evaluations [1]

AI在软件工程中的应用与发展 - 多个公司和研究机构正在探索和开发AI在软件工程中的应用，包括代码生成、质量控制和自动化[1] - 行业关注AI如何提升软件开发效率和质量，以及如何量化AI在软件工程中的投资回报率[1] - AI Coding Agents 的未来发展趋势，包括构建可靠的系统以适应模型迭代周期[1] - 讨论了AI在浏览器构建中的应用，以及从中获得的经验教训[1] 工程实践与领导力 - 探讨了在AI辅助工程中如何进行领导，以及如何构建AI原生公司[1] - 讨论了工程团队如何利用AI来改进支持服务[1] - 一些公司正在尝试新的工程师激励机制，例如将工程师的薪酬与销售业绩挂钩[1] - 传统敏捷方法的替代方案正在被探索[1] 特定技术与平台 - 关注 evolving Claude APIs for Agents [1] - 讨论了Minimax M2 的研究与应用[1] - 介绍了Google DeepMind 的研究成果及其在现实中的应用[1] - Bloomberg 在其工程组织中部署 AI 的经验教训[1]

AI Engineering & Industry Leaders - Neo4j's Co-Founder and CEO discusses "The State of AI Engineering" [1] - Docker focuses on "Democratizing AI Agents: Building, Sharing, and Securing Made Simple" [1] - GitHub addresses "Building MCP's at GitHub Scale" [1] - H Company is assembling open source bricks for the next generation of AI [1] - Google DeepMind shares updates on generative AI [1] AI Infrastructure & Tools - Koyeb explores "Building for the Agentic Era: The Future of AI Infrastructure" [1] - Black Forest Labs presents "Inside FLUX, How It Really Works" [1] - LlamaIndex is building an open-source NotebookLM alternative [1] Open Source & Community - Hugging Face reports on the "State of Open LLMs in 2025" [1] AI Applications & Techniques - Arize AI studies "System Prompt Learning for Agents" [1] - ZML is working "Towards unlimited contexts: faster-than-GPU sparse logarithmic attention on CPU" [1] - Kyutai is scaling real-time voice AI [1]

The opening welcome reception is all about the hallway track and the expo -- meeting and mingling with other founders and engineers who are (mostly) based in Europe. However, for those who can't make it, we'll be streaming 2 talks to kick off the conference: Shawn Swyx Wang, curator of Latent Space and Co-founder of AI Engineer: The Year in Agents Lélio Renard Lavaud, VP of Engineering, Mistral: How open source drives successful enterprise adoption ...

Challenges in Building AI Applications at BlackRock - BlackRock faces challenges in prompt engineering, requiring significant time investment from domain experts to iterate, version, and compare prompts effectively [10] - BlackRock encounters difficulties in selecting appropriate LLM strategies (e.g., RAG, chain-of-thought) due to instrument complexity and document size variations, impacting data extraction [11] - BlackRock experiences deployment challenges, including determining suitable cluster types (GPU-based inference vs burstable) and managing cost controls for AI applications [12][14] BlackRock's Solution: Sandbox and App Factory - BlackRock developed a framework with a "sandbox" for domain experts to build and refine extraction templates, accelerating the app development process [15][17] - BlackRock's "sandbox" provides greater configuration capabilities beyond prompt engineering, including QC checks, validations, constraints, and interfield dependencies [19][20] - BlackRock's "app factory" is a cloud-native operator that takes a definition from the sandbox and spins out an app, streamlining deployment [15] Key Takeaways for Building AI Apps at Scale - BlackRock emphasizes investing heavily in prompt engineering skills for domain experts, particularly in the financial space, due to the complexity of financial documents [26] - BlackRock highlights the importance of educating the firm on LLM strategies and how to choose the right approach for specific use cases [27] - BlackRock stresses the need to evaluate the ROI of AI app development versus off-the-shelf products, considering the potential cost [27] - BlackRock underscores the importance of human-in-the-loop design, especially in regulated environments, to ensure compliance and accuracy [28]

AI Development & Application - The industry is moving towards designers, product people, and engineers collaborating to build together, eliminating mock-ups and click-through prototypes [1] - Flat Files AI stack is structured into four buckets: invisible, ambient, inline, and conversational AI, each offering different levels of user interaction [1] - The company is exploring AI agents that can write code to set up demos tailored to specific user use cases, such as creating an HR demo for users from HR companies [1] - The company is developing tools that allow AI to analyze data in the background, identify opportunities for improvement, and provide inline assistance to users working with the data [1] - The company is building no-code/low-code agentic systems that can write Flat File applications, potentially reducing the need for engineers in this process [1] AI Agent Design & Character - The company is shifting from controlling AI agents to character coaching, focusing on building out the desired nature and characteristics of the agents [1] - The company is experimenting with giving AI agents tools like cursors to interact with design tools, exploring how AI can operate in the design space [2] - The company is aiming to create an environment where LLMs can shine, focusing on form factors that help them nail their assignments, stay aligned, and grow as models improve [1] Emergent Behavior & Future Exploration - The industry is seeing emergence in AI, with AI agents exhibiting curiosity, excitability, and focus, leading to unexpected and valuable outcomes [6][7][8] - The company is exploring the use of AI agents with knowledge bases to surface suggestions and help users complete tasks, even when the AI cannot directly fix the issue [12][13][14] - The company is focusing on autocomplete backed by LLMs, designing applications to test and benchmark the performance of different models [16][17]

AI Platform Evolution - Box transitioned to an agentic-first design for metadata extraction to enhance its AI platform [1] - The shift to agentic architecture was driven by the limitations of pre-generative AI data extraction and challenges with a pure LLM approach [1] - Agentic architecture unlocks advantages in data extraction [1] Technical Architecture - Box's AI agent reasoning framework supports the agentic routine for data extraction [1] - The agentic architecture addresses the challenge of unstructured data in enterprises [1] Key Lessons - Building agentic architecture early is a key lesson learned [1]

AI Development Strategy - Building successful AI applications requires a sophisticated engineering approach beyond just writing good prompts [1] - The industry emphasizes the importance of evaluations (evals) as a core component of the development process [1] - Evaluations should be intentionally engineered to reflect real-world user feedback and drive product improvements [1] Technical Focus - "Context engineering" is emerging as a new frontier, focusing on optimizing the entire context provided to the model [1] - Context engineering includes tool definitions and their outputs [1] - The industry advocates for a flexible, model-agnostic architecture [1] Adaptability - The architecture should quickly adapt to the rapidly evolving landscape of AI models [1] - Optimize the entire evaluation system, not just the prompts [1]