aggregate intellect

so, you have an agentic ai product idea, should you pursue it? while the answer is almost always "yes", knowing more about the challenges you should expect is a good way to improve your chances of success. 1100 people have signed up for this week's aggregate intellect "AI Agents Series", where we will discuss this topic with Jason Dea. Friday, Nov 21, 12pm ET You should expect to hear: 🟢 Grounded Problem Validation - How to ensure your AI idea solves a genuine problem, not just a shiny use case. 🟢 Recognizing Qualitative Shifts - Why generative AI removes old barriers (e.g., data normalization, rigid APIs) and how that expands viable product ideas. 🟢 Evaluating Agentic Potential - How to decide if an “agent” really adds value - or if a simpler automation would suffice. Jason Dea is a product leader and startup advisor with over 15 years of experience scaling SaaS companies from early traction to acquisition. He has led product strategy, marketing, and growth at organizations like Uberflip, Honeybee Benefits, and TELUS Health, driving multimillion-dollar ARR growth and successful exits. Currently Venture Chief Product Officer at Koru (funded by Ontario Teachers' Pension Plan), Jason helps founders discover, validate, and launch new ventures. With a background spanning engineering, product marketing, and executive coaching, he specializes in aligning teams, building scalable roadmaps, and unlocking predictable growth for startups. ps. comment or dm if you'd like to join

you know how you listen to a talk and think that you'd like to have this deck so that you can look at it later? so this Friday, at aggregate intellect "AI Agents Series" we will do an experiment where we do exactly this with a twist. I will provide a prompt that will regenerate my slides for you in a way that is personalized to your technical level and area of interest, and you can go through the slides reading the notes prepared for you while listening to me and the topic happens to be context management for agents, if that matters let me if you'd like to join

The information firehose around AI is overwhelming. All of us subscribe to 15 newsletters, follows 50 accounts, bookmarks 100 articles. Still can't find what's actually relevant to our work when we need it. The irony? The solution to information overload is an AI problem we can actually solve today. This Friday (Nov 7, 12pm ET), Alireza Darbehani will show how he's gone about this in aggregate intellect "AI Agents Series". You'll learn: 🔹 How to design pipelines that combine retrieval, summarization, and ranking agents 🔹 How to build feedback loops that improve content quality over time 🔹 The infrastructure choices that matter for reliable, maintainable systems This is about practical engineering for a problem you probably have right now. Comment or DM me if you want to join.

Some of the most important moments in our careers are the ones we don't recognize until years later. For me, one of those moments was meeting David Scharbach back in 2017. I started using AI in my academic work in 2015-2016 and in 2017 I got a job as a data scientist which meant that I started getting paid to do AI. To get my bearings in the industry, I was attending every meetup I could find. David was running one of those groups, and we hit it off immediately. Our paths crossed more frequently when he launched his well-known Machine Learning Summits, where I helped on the steering committee. But it was a smaller side project of his that changed everything for me: a paper reading group. As an academic-turned-industry person, it was exactly what I was looking for. When he got too busy to run it consistently, I took it over. That 30-40 people mailing list, nurtured with David's advice and concrete help (he even set us up with a formal email address in 2018), became the seed for our regular paper meetings. That seed eventually grew into "Toronto Deep Learning Series (TDLS for those who remember that), and later my company and its community "AI Socratic Circles (AISC)" which we still run as a global community of over 6,000 AI researchers, engineers, and founders. Before I decided to incorporate aggregate intellect, we even briefly considered launching it as a joint effort - envisioning it as the product side to his summits. He was a role model and an advisor who set me on the path I'm still on today. He's my unsung hero. Send a thank you note to someone that quietly changed the course of your career

"We can't standardize on one AI assistant across our team because everyone has different preferences but also the landscape is shifting so fast." Sound familiar? This is something we've been increasingly experiencing at aggregate intellect and also hearing from clients and prospects. Here's what's happening: teams are creating isolated AI workflows. Sarah uses Claude for strategy docs, Mike relies on Cursor for code reviews, and Lisa has her own Gemini setup for research. Each person becomes an expert in their chosen tool, but knowledge doesn't transfer. And these are the ones who adopted a tool and don't even get me started on the laggers. The bigger problem isn't the tool fragmentation itself - it's that we're building institutional knowledge that disappears when someone leaves or when the AI landscape shifts (which it will, constantly, like just now, or now). Think about it this way: if your team's AI workflows are locked into specific platforms, you're essentially creating technical debt in your knowledge management system. This, of course, translates into an efficiency debt in your business workflows! 🟢 What interoperability looks like in practice: • Standardized prompt libraries that work across different AI models • Shared context systems that don't depend on proprietary platforms (or at least can be easily exported / imported) • Workflow documentation that translates between tools 🟢 Why consistency matters beyond efficiency: • New team members can onboard faster • Institutional knowledge survives platform changes or employee turnover • Quality becomes predictable across the organization The companies getting this right are thinking about AI workflows the same way they think about version control - tool-agnostic, transferable, and built to last. I've been experimenting with this approach in my own work, building systems that travel seamlessly between different AI assistants. While I primarily use Claude Code for almost everything, I have been setting things up so that I can easily switch to Gemini CLI or vs code + copilot or chatGPT web version etc without loss of continuity. This could be because another tool is better at a certain job or for cost reasons or sometimes to be able to work more effectively with other team members. Friday, 12pm ET I'll walk through the setup I'm using if anyone's curious about the details. DM / comment if you'd like to join and I'll send you the link. What is your setup?

What should AI builders consider about the human side of their craft? We've gotten so good at optimizing for specific skills, but the ground is shifting under our feet. The "proven" career paths for talent are being disrupted by the very tools we're creating. It feels like we're handing out maps for a territory that's being redrawn in real-time. This is the tension that I'm excited to explore this Friday with Ian🤗 Yu in aggregate intellect "AI Agents Series": "Building AI Systems: Architecture, Team, and Trade-offs." We're going to be digging into the messy, human side of building AI: 🟢 How the culture of your team shapes the architecture of your systems (and vice-versa). 🟢 The unglamorous but critical work of building for maintainability and long-term flexibility. 🟢 The very real trade-offs you have to make when you're building for both today's needs and tomorrow's unknowns. This isn't a session about finding the "right" answers. It's about asking better questions and learning to navigate the uncertainty together. It's a conversation for anyone who's in the trenches, trying to build robust, effective AI systems in a world of constant change. If you're also wrestling with these questions, I'd love for you to join us. DM or comment if you want to join.

You know that feeling when your AI just doesn't quite get it? When it pulls in irrelevant info, or misses the nuance you were hoping for? That's a common challenge in building truly intelligent systems. This Friday, at aggregate intellect "AI Agents Series" we're diving deep with Marc Pickett, co-author of the paper "Better RAG using Relevant Information Gain". His work introduces a fascinating new way to optimize Retrieval Augmented Generation (RAG) – not just for relevance, but for diversity in the information retrieved. Think about it: what if your LLM could organically find the most useful context, even with limited window sizes? Marc's research offers a fresh perspective, moving beyond traditional relevance metrics to achieve state-of-the-art performance in question answering. It's about making AI more effective and reliable in real-world applications. DM / comment if you'd like to join

🚀 We’re kicking off our two-part series with Dr. 🟢 Amir Feizpour, CEO of aggregate intellect, AI advisor, and community builder with 5,000+ practitioners worldwide. In this session, you’ll gain: ✅ Practical insights on working with LLMs and Agentic AI ✅ Operational benefits of advanced AI integration, from productivity to workflow ✅ A clear view of organizational constraints and real-world impacts 📅 Thursday, Sept 25 | 11:00 AM – 12:30 PM EST 🎟️ Limited seats available, don’t miss out! 👇 Register today by following the link in the comments below!

does your "deep research" implementation actually work? that's what Suhas Pai will speak about at aggregate intellect "AI Agents" series tomorrow, Fri 12pm ET this will be part 2 following session 1 where he spoke about architecture of these systems: 🟢 Definition of a Deep Research System: It's a system that uses an arbitrary number of retrieval and reasoning steps to fulfill a detailed information need. Unlike traditional search engines, it doesn't just provide links or summaries, but a detailed report by synthesizing information from multiple sources. 🟢 Core Components: Deep research systems consist of a retrieval system (for searching), a reasoning system (for connecting information), a report generation system (for creating the final output), a context orchestrator (for managing long reasoning chains and context length), and a depth and breadth balancer (for controlling the scope of the search). 🟢 Interplay of Search and Reasoning: A crucial aspect is the iterative loop between searching and reasoning. The system doesn't search first and then reason; instead, it reasons based on current information, which then informs further search queries, potentially leading down "rabbit holes" or broader investigations. 🟢 Challenges in Development: ◌ What to Search: Determining what exactly to search for is a major bottleneck, as the system needs to be smart enough to generate precise queries to get desired results, going beyond simple automated Google searches. ◌ When to Search/Stop: Knowing when to initiate a search and, more importantly, when to stop searching or reasoning, is a complex and active area of research. ◌ Report Generation: Structuring the vast amount of synthesized information into a coherent and detailed report that is still digestible and doesn't miss key "nuggets" is challenging, especially given potential context window limitations. 🟢 Internal vs. External Control: ◌ Internal Decision: The ideal scenario is for the model to be trained to intrinsically know when to search, reason, and stop. ◌ External Decision (Scaffolding): In practice, external heuristics and guardrails are often used to control the system's behavior (e.g., limiting the number of documents searched, reasoning budget, or context length thresholds). While effective, this approach can be brittle. 🟢 Context Engineering / Orchestration: Managing the context window dynamically is vital, especially when dealing with hundreds of documents. This involves knowing the original task, what has been searched, how previous searches relate to the task, and the current state of answer generation, while avoiding information overload. 🟢 Depth and Breadth Balancing: The system needs to manage both in-depth "rabbit hole" exploration and broad-based research. This can be controlled by setting user-defined or dynamic ceilings for depth and breadth, although accurately defining and assigning these levels can be difficult for the model. 🗓️ comment/dm to join

how is a Deep Research system built? that's what Suhas Pai will talk about at aggregate intellect "AI Agents Series" later this week. how deep research systems retrieve and generate reports what components power agentic research workflows which design patterns shape current system architectures "In the last few months, several pioneering AI labs have launched powerful 'Deep Research' features that search extensively across a large number of data sources and produce comprehensive reports in response to user queries. In this talk, we will discuss the anatomy of such a system, focusing on the key components involved, and architectural patterns." Fri, Aug 8 2025,12pm ET comment or dm if you'd like to join