How Stanford & MIT Researchers Use AI Agents

Introduction and Welcome

Opening Remarks

• Jin Konan Wang welcomes attendees, expressing gratitude for their participation.

• Attendees are encouraged to share their locations in the chat, fostering community engagement.

• Acknowledgment of participants from various cities, including Kansas City and NYC.

Purpose of Founder Coho

Founding Motivation

• Jin explains the inception of Founder Coho, emphasizing the need for rapid learning in AI startups.

• The importance of collaborative learning is highlighted as a core reason for establishing the cohort.

Overview of Activities

Growth and Engagement

• Founder Coho has evolved from offline events to online platforms, including podcasts and live streams.

• The organization boasts over 200,000 members across various media channels and hosts 50 events annually.

Introduction to Today's Topic

Leveraging AI in Research

• Jin introduces Shannon and Yi Jia from MIT and Stanford, who will discuss their research using AI.

• They have created the "Augmented Mind Podcast," which integrates AI throughout its production process.

Presentation by Shannon

Podcast Creation Insights

• Shannon expresses gratitude for being part of the event and introduces her co-presenter Yi Jia.

• Discussion begins on how they utilize AI in creating their podcast as a self-case study.

Challenges in Disseminating Research

Exploring New Formats

• As PhD students, they reflect on innovative ways to share technical information beyond traditional papers.

• Podcasts are identified as an effective medium to convey both technical content and personal stories behind projects.

Podcast Development Journey

Progress Overview

• The team has released six episodes covering diverse topics related to human-centered AI work over six months.

• Notable guests include founders discussing significant issues like privacy and technology applications.

Workflow Challenges with AI Integration

Key Areas of Focus

• Three main branches identified: content preparation/writing, recording/post-production, design/coordination.

• Emphasis on maintaining a unique aesthetic while ensuring human touch amidst increasing automation.

Human Collaboration with AI Tools

Balancing Technology with Humanity

• Shannon uses a "canned food" analogy to illustrate that while AI can produce outputs efficiently, human creativity remains essential.

• She emphasizes that collaboration between humans and AIs enhances quality without losing personal touch.

This structured markdown file captures key insights from the transcript while providing timestamps for easy reference.

Challenges in Human-Agent Interaction Scalability

Overview of Current Limitations

• The scalability of human-agent interactions is limited, particularly when considering extended usage beyond one hour.

• While agents perform well for short tasks, issues arise with prolonged engagement, leading to user frustration and inefficiency.

Interaction Friction and Context Management

• Users face significant interaction friction due to fragmented contexts across multiple chat windows, which can degrade agent performance over time.

• Managing numerous agent sessions creates chaos; users struggle to track conversations and context, complicating the workflow.

Learning and Coordination Issues

• Agents do not retain learning from previous interactions, resulting in repeated mistakes that waste user effort.

• Current agents are designed for individual use rather than team collaboration, making it difficult to share work effectively among teammates.

Goals for Enhancing Agent Collaboration

Scaling Human-Agent Collaboration

• The aim is to improve organizational efficiency by increasing the number of agent sessions a user can manage simultaneously without wasting resources.

• There’s a focus on enhancing the utility of agent hours so they contribute meaningfully rather than just consuming tokens.

Team Collaboration Dynamics

• Effective scaling requires addressing how collaboration occurs among team members; current systems limit efficiency to single-user scenarios.

Building a Self-Evolving Repository

Components of the Repository

• The self-evolving repository consists of three parts: harnessing agents on GitHub, continuous learning components, and an orchestration layer for task management.

Functionality on GitHub

• Agents can be invoked directly within GitHub environments to answer questions or implement code by tagging them in issues or pull requests.

Benefits of Integrating Agents into GitHub

Zero Learning Cost Advantage

• Utilizing agents within familiar environments like GitHub eliminates learning costs associated with new tools or platforms.

Structured Sessions and Seamless Workflows

• By operating directly on GitHub, agents streamline workflows without requiring local setups or code cloning.

Enhancing Team AI Collaboration

Real-Time Updates and Feedback Loops

• Teams can collaborate more efficiently as updates made by one member are immediately visible to others through shared threads in GitHub.

Continuous Learning Mechanism

• All interactions are recorded within the codebase allowing for continual learning; feedback from past interactions helps avoid repeated mistakes in future tasks.

Enhancing Code Review with Automated Agents

Agent Sessions and Rubric Implementation

• The agent sessions are designed to evaluate code implementations against newly added rubrics, identifying violations and automatically fixing them to ensure the final version meets all standards.

Orchestrating Long-Form Tasks

• A key feature allows agents to manage long-form tasks effectively, addressing the challenge of reviewing large code chunks that can be time-consuming compared to manual coding.

Focused Pull Requests Creation

• The agent can split tasks into specific pull requests (PRs), creating issues and implementing them sequentially after each passes review, streamlining the development process.

Continuous Improvement Loop

• The implementation loop involves the agent writing code, switching to review mode, checking rubrics, and making necessary adjustments until all quality checks are satisfied.

Importance of Dual Reviews

• Both rubric reviews and code quality assessments are crucial for ensuring that the final product not only meets technical requirements but also aligns with specific project needs.

Goals of Agent Collaboration

Speeding Up Team Interactions

• The ultimate aim is for teams to interact more efficiently with agents, allowing these automated systems to take on more responsibilities in coding tasks.

Reducing Repetitive Feedback Loops

• By automating feedback through established rubrics, repetitive issues can be minimized, enhancing overall productivity in coding practices.

Understanding Rubrics

Definition and Purpose of Rubrics

• Rubrics serve as guidelines or checklists that dictate how code should be structured or systems architected; they help maintain consistency in coding practices.

Example of a Design Rubric

• An example rubric emphasizes flexibility and trust in agent-derived decisions during infrastructure design, guiding how agents should approach coding tasks.

Ensuring Quality Work from Agents

Challenges with Agent Autonomy

• There are concerns about ensuring that agents produce high-quality work without close supervision; this necessitates a robust framework for monitoring their outputs.

Iterative Refinement Process

• To combat potential pitfalls like hacky solutions from agents, an iterative refinement process is implemented where initial implementations undergo quality checks before progressing further.

Visibility and Accountability of Agent Actions

Addressing Human Oversight Concerns

• Questions arise regarding how humans perceive agent actions—whether they view it as a "file-and-forget" approach or if there’s room for reflection on self-improvement by the agents themselves.

Ensuring Proper Functionality

• It’s essential to confirm that agents correctly identify files and execute tasks effectively; this includes maintaining oversight on their operations within team dynamics.

Providing Guidance to Agents

Inputs for Effective Collaboration

• Teams provide guidance by defining issues for improvement while also instructing agents on preferred working methods aligned with team standards.

Learning from Feedback

• After completing tasks, agents receive specific feedback which is distilled into future learning sessions—this continuous feedback loop enhances their performance over time.

Real-Life Application Example

Case Study: Collaborative Development Effort

• A recent collaborative effort involved multiple team members working together using an agent; this showcases practical applications of automation in real-world scenarios such as website updates.

How Sepple Streamlines Development Processes

Overview of Sepple's Functionality

• The speaker discusses the repetitive nature of their work and how Sepple simplifies issue creation by allowing users to paste necessary information directly into a new issue.

• After creating an issue, Sepple facilitates communication between team members, leading to seamless implementation and pull requests without excessive manual intervention.

Bug Fixing Process

• During implementation, a bug was identified on the website involving overlapping text, prompting collaborative problem-solving among team members.

• The speaker contributed a proposal for fixing the bug within the same thread, which was implemented and merged into the final version of the website.

Collaborative Environment

• The shared workspace on GitHub allows all contributors to engage without unnecessary handoffs or delays in communication.

• This example illustrates how Sepple can streamline workflows and enhance collaboration across tasks.

Human-AI Collaboration in Podcast Production

Role of AI in Content Creation

• Questions arise about how agents are used to automate processes like finding topics for podcasts; human input remains crucial in topic selection.

• The importance of defining roles between AI and humans is emphasized; critical decisions about content remain entirely human-driven.

Manual Outreach Efforts

• The podcast team manually reaches out to potential guests, underscoring the value placed on personal interaction over automation.

• A balance is maintained where certain tasks are delegated to AI while retaining essential human oversight for creative aspects.

AI's Limitations in Open-ended Tasks

Non-engineering Applications of AI

• Discussion highlights that many tasks performed with Sepple are non-engineering related but still benefit from coding knowledge.

• While AI can assist with various inquiries, its effectiveness is most pronounced when tasks relate closely to coding activities.

Concerns About Using AI for Interview Preparation

• There are concerns regarding using AI for open-ended programming tasks due to their subjective nature; guidance through rubrics is suggested as a solution.

• Emphasis is placed on ensuring that while AI can generate ideas, it may not produce final versions suitable for specific needs.

Case Study: Human Collaboration in Animation Creation

Challenges in Video Production

• A case study is introduced focusing on human collaboration during animation creation within video production pipelines.

• The complexity of integrating AI into creative processes such as animations showcases potential advancements in automating intricate tasks.

Understanding AI Workflows and Human Involvement

The Role of Visuals in AI Presentations

• The discussion emphasizes the importance of visuals in conveying information, suggesting that animations can effectively illustrate concepts even without audio.

• A focus on the complexity of creating animations with multiple elements, such as tables and desks, is introduced.

Generating Complex Structures with AI

• It is noted that generating intricate designs requires detailed prompts for AI models like ChatGPT to produce satisfactory results.

• The speaker highlights a specific example involving a central robot, showcasing how AI can assist in complex design tasks.

Workflow Integration and Image Enhancement

• The process involves using Gemini 3 for initial image generation but notes issues with background quality and texture.

• To improve the output, the speaker utilized ChatGPT for texture enhancement and manual software for background removal.

Human-AI Collaboration

• Emphasizing that AI should be viewed as a tool rather than an endpoint, the speaker advocates for human involvement in combining various tools to achieve optimal results.

• Reflecting on automation's role, it’s suggested that humans must determine what aspects should remain automated versus those requiring personal touch.

Questions from the Audience

• An invitation is extended to viewers to ask questions during the session, indicating an interactive format.

Choosing Models: Insights from Experience

Model Selection Criteria

• A common question arises regarding how different models are chosen based on their unique traits observed through usage.

Preference for Latest Models

• The speaker prefers using more models simultaneously while opting for the latest versions available to ensure optimal performance.

Differences Between Models

• Notable differences between Cloud Code (more collaborative and human-like interaction) and Codex (task-oriented execution without much dialogue).

AI Code Review Practices

Multi-Agent Review Process

• For code reviews, three separate agent sessions are employed: one from Cloud Code, one from Codex, and a final session to synthesize findings into one cohesive piece.

Importance of Diverse Perspectives

• Highlighting that agents can make mistakes just like humans; thus having different agents review each other's work enhances accuracy and reliability.

Building Effective Teams in AI Development

Prioritizing Roles in Agent Teams

• A discussion on "loop engineering" emerges as essential when building teams capable of autonomous action within AI frameworks.

Understanding Agent Workflows and Capabilities

Importance of Process Understanding

• The key to effective agent architecture lies in understanding workflows and processes, which is crucial for generating meaningful outputs rather than just following typical AI standards.

• Emphasizing the importance of these skills reflects a growing need for efficiency in working with agents.

User vs. Model Capabilities

• Distinction between user capabilities (e.g., managerial thinking) and model capabilities is essential for effective teamwork within agent systems.

• As overall model capabilities improve, specific individual capabilities become less critical, suggesting a focus on holistic development.

Accuracy in Chaining Agent Sessions

• Example illustrates that chaining multiple agent sessions can lead to significant accuracy degradation if not managed properly; a 0.9 accuracy per session results in only 0.3 final accuracy when chained.

• Improved overall agent capability leads to lower degradation rates, enhancing the reliability of outcomes from chained sessions.

Code Review Processes

Decision-Making in Code Reviews

• The code review loop involves determining when an agent should stop its process; typically capped at 10 runs but can extend to 20 or 40 based on performance.

• Agents assess their readiness for handoff after reviews, indicating a level of self-management within the system.

Tools Used for Reviews

• Codex or Cloud are utilized depending on available credits; however, the choice of tool is deemed less critical compared to reconciling different reviews from agents.

• Most issues flagged during reviews range from warnings to blocking levels, emphasizing the summarization aspect of this task.

Evolving Rubrics and Self-Evolving Bots

Research Context and Implementation

• The concept of evolving rubrics stems from previous research aimed at improving model performance through reinforcement learning (RL).

• In production systems, direct training with advanced models like Codex isn't feasible; instead, an inference-time algorithm scores generated outputs against established rubrics.

Iterative Refinement Process

• Even without direct training capabilities, iterative refinement based on rubric feedback allows continuous improvement in output quality.

Addressing Hallucination Concerns

Clarifying Hallucination Sources

• Questions about hallucinations must differentiate between training setups and real-world applications; observing all model rollouts poses challenges during training.

Evaluation Metrics

• Confidence in gradual improvements relative to rubrics indicates progress despite potential inaccuracies; absolute scoring may be less important than relative rankings among outputs.

Continuous Improvement Monitoring

• Regular audits ensure that agents are genuinely improving code quality over time rather than producing erroneous outputs labeled as "hallucinations."

Connecting with Experts

Follow-Up Opportunities

• Participants are encouraged to connect via podcast websites for further insights into human-agent collaboration dynamics discussed during the session.