ai
Expert interviews and guides to help you level up as a product leader and creator fast.
Jaclyn Konzelmann, a product director at Google Labs, discusses how the launch of Gemini 4 shifted their team's approach from writing 40-page PRDs to focusing on high-level strategy documents, presentations, and fast prototypes.
Jaclyn Konzelmann explains that Google Labs follows a guiding philosophy of prioritizing users first, Google second, and Labs third, combining emerging model capabilities with emerging user trends to drive product development.
Google Labs product director Jaclyn Konzelmann outlines rules for using AI in document creation, stating that if a document takes longer to read than it took to produce, it is not allowed, and authors must be able to defend every single word.
Google Labs projects such as Notebook LM and CC have successfully transitioned or graduated into features within the core Gemini app, showcasing multiple paths to success for lab experiments.
Penn, a designer on the Grockbot team, described how his 'Chief of Staff' bot can handle complex, multi-step tasks. For example, the bot not only purchased 3D printing filament but also updated an inventory tracking database. He also detailed a bot that can sell items on Facebook Marketplace, including researching prices, learning the user's writing style, and automatically lowering prices weekly if no interest is shown.
Penn shared how he uses Grockbot to replace manual email and calendar management. He can ask the bot to identify action items in his inbox, make payments, or manage calendar events by simply providing screenshots. The bot also assists in selling items online, such as a DJI mic on Facebook Marketplace, by automating tasks like price research and listing updates.
Penn detailed his use of specialized Grockbot agents for design and writing. The 'designer' bot brainstorms and executes design ideas by interacting with design tools via API, while the 'writer' bot polishes emails and Slack messages, particularly useful for non-native English speakers. The system allows for inline review and modification of drafts.
Penn shared a unique use case where Grockbot acts as a podcast tool, translating articles into audio formats for listening on the go. He also described a personal project where Grockbot automates the process of posting content to his website. By tagging an image and associated social handles, the bot generates images, removes backgrounds, and publishes to his site, replacing the need for a manual CMS.
Penn elaborated on his designer bot, explaining its integration with the Figma API. This allows the bot to directly modify frames, update designs, and utilize design tokens. The bot understands the user's design system, including color palettes, typography, and spacing, enabling it to create designs or perform significant scaling of initial concepts.
Lauren shared engineering use cases for Grockbot, focusing on her work with Omachi, a Linux distribution. She uses Grockbot to script and control a virtual machine running Omachi on her Mac, which is crucial for developing and testing support for this Linux flavor. The bot can interact with, install, and manage the VM.
Lauren described her agent named 'Crumb' which helps manage hard disk space. When she runs out of storage, she instructs 'Crumb' to scan her hard drive, identify and delete unnecessary files. This capability highlights Grockbot's access to a computer's file system based on granted permissions.
Lauren explained how she uses Grockbot to manage her personal stack of plugins, P-Stack. She initiates a 'cursor Cloud agent' which performs automated evaluations ('Evals') to ensure skill improvements result in better performance. Only after the Evals playbook in P-Stack confirms the goal was achieved does the system finalize the pull request.
Pedro Franceschi, co-founder and CEO of Brex, suggests that the future of AI in business lies in building "virtual employees" rather than open-ended agents. He emphasizes the importance of making these AI entities behave and feel like real people, citing their internal recruiter AI, 'Jim,' as an example of an AI that acts like a coworker.
Pedro Franceschi of Brex argues that the traditional product manager (PM) playbook of writing PRDs and aligning stakeholders is obsolete due to AI. He notes that AI is blurring the lines between traditional roles like engineering, product, and design, with PMs increasingly writing code and engineers making product decisions.
Pedro Franceschi of Brex explains their strategy to compress product development cycles by internalizing customer signals and external opportunities quickly. He emphasizes that PMs at Brex now use internal tooling to ingest customer context and make tradeoff decisions, leading to faster and more accurate product shipping.
Pedro Franceschi highlights that in an era where building features is increasingly easy due to AI, the ability to discern what truly matters is more critical than ever. He advocates for minimizing the surface area of development, focusing deeply on solving one problem exceptionally well, rather than attempting to do many things superficially.
A conversation with a PM on the Codex team reveals that engineers are now empowered to build their own features and are also encouraged to speak with customers. This shift means PMs spend more time managing the influx of ideas and customer requests that engineers are now equipped to pursue.
Sue, CEO of Brilliant, observes a concerning trend in the US education system where test scores are falling while grades are rising. She attributes this to a phenomenon of grade inflation, where standards are lowered to appease parents and maintain school rankings, rather than addressing underlying performance issues. This trend is particularly evident when comparing US students to international benchmarks.
Sue, CEO of Brilliant, believes that Artificial Intelligence presents a significant opportunity to enhance children's learning capabilities. She compares the potential impact of AI to the invention of writing, suggesting it could fundamentally change how children learn and develop intellectual capacities. Properly utilized AI could lead to unprecedented educational advancements.
Sue expresses concern over the perceived decline in academic rigor, citing data on AP exams. She notes that AP exams, which once tested undergraduate and graduate-level material, have gradually lowered their standards to assess high school and now even middle school level reading passages. This trend suggests a broader issue of educational institutions 'giving up' on maintaining high academic expectations.
Sue suggests that the optimal use of classroom time involves fostering peer learning and leveraging teachers' ability to inspire students. She argues that traditional lecture-based instruction and repetitive practice are inefficient uses of both student and teacher capabilities. Instead, classrooms should focus on collaborative environments and engaging interactions.
A repeat founder, John, who is not technical, was quoted $100,000 by an agency to build an AI proficiency measurement and certification platform. He then used Relace and built the entire application end-to-end in just three days. Within the first two months of launching, the platform generated over $180,000 in revenue.
Cedric, a 22-year-old with no prior coding experience, developed a mobile app using Relace to help users track peptide and GLP-1 intake. After identifying a niche in the peptide community, he built the app and, with effective distribution, achieved $60,000 in revenue in his first month. The app now has tens of thousands of active users.
Yusef, also known as FaZe Apex, a well-known gamer and creator, has launched a new business called Train Nearby using Relace. The platform aims to connect local businesses with local creators for word-of-mouth marketing. Train Nearby is reportedly achieving upwards of $100,000 in ARR.
Amol Jain of Relace highlights that the key to successful AI-driven businesses lies in leveraging unique expertise, judgment, or existing distribution channels, rather than just generic software. He notes that as the cost of code approaches zero, users are willing to pay for this added value and 'done-for-you' services.
Nan Yu and Jacob Shumway from Linier discussed the evolution of AI agents, moving from handling mundane tasks to enabling creative possibilities. Initially envisioned for automating 'boring stuff,' AI agents can now perform complex tasks and multiple roles, including end-to-end product development. The core concept of an agent involves calling a Large Language Model (LLM) in a loop, utilizing tools to build context and achieve a defined goal.
When working with AI agents, the advice is to provide minimal instructions to achieve optimal results. Over-prompting can degrade performance. Instead of feeding the agent all necessary context upfront, it's suggested to give it the tools to load its own context, allowing it to iteratively build the necessary information.
Nan Yu reflects on the rapid evolution of AI agents, noting that their initial concept, developed in late 2025, was for a 'robotic program manager' to handle tedious tasks. This perspective has shifted dramatically in just one year, with AI agents now seen as capable of creative work and interactive problem-solving, transcending their initial role of simply automating 'boring stuff'.
Jacob explains that at its core, an AI agent is essentially a Large Language Model (LLM) operating in a continuous loop. Unlike a standard LLM interaction which yields a single answer, agents are designed for complex tasks, executing actions iteratively by calling tools to gather context. This process continues until the agent determines the goal is met, at which point it synthesizes a final response.
Karan Malhotra, co-founder of Hermes agent, explained that Hermes differs from other AI agents through its self-improvement system and dedication to aligning with user needs rather than external agendas. He emphasized that their approach aims to prevent 'reward hacking,' where models exploit systems for rewards without genuine helpfulness.
Karan Malhotra explained that the Hermes agent harness can take existing models like Claude and make them perform better than in their native environments. This is achieved by re-aligning the model's primary allegiance from anthropomorphic tendencies to the specific user's needs, potentially improving performance on benchmarks.
Karan Malhotra addressed the issue of AI models exhibiting 'sycophancy' by agreeing excessively rather than offering genuine opinions. He clarified that this behavior is a form of 'reward hacking,' where the model prioritizes positive reinforcement over accuracy. Malhotra suggested that introducing new context, logic, and personalities can help mitigate this.
Karan Malhotra highlighted that in-context learning (ICL) is the most powerful tool for AI models, more so than fine-tuning. He explained that by providing examples and saving desired behaviors, users can achieve test-time reinforcement learning within Hermes. This allows the model's context to become so overwhelming that it minimizes differences between various underlying models like Claude or GPT.
Karan Malhotra discussed the advantages of Hermes being open-source, emphasizing community-driven innovation and the ability for developers worldwide to build upon the foundation. He noted that this approach fosters transparency and rapid iteration, leading to a more accessible and beneficial AI technology. Challenges include potential vulnerability discovery and community management.
Karan Malhotra showcased a calendar management agent built using Hermes, which understands natural language requests to schedule, reschedule, or decline meetings. The agent leverages the Hermes harness to interact with calendar APIs and take action, demonstrating advanced reasoning, contextual understanding, and seamless integration.
Karan Malhotra highlighted the broad range of applications built with Hermes, including customer support bots, content creation assistants, research tools, and personalized learning aids. He stated that the open-source nature of Hermes empowers users to develop solutions tailored to their specific needs, fostering innovation across various sectors.
Jason Liu from OpenAI discusses how the ChatGPT Work app is a different view of the Codex experience, designed for productivity by abstracting away coding details. While Codex shows Git history and pull requests, Work hides these for tasks like making slides, where seeing the underlying Python code might be unnecessary.
Jason Liu explains his approach to selecting AI models and effort levels within Codex, often defaulting to 'Sol medium' for tasks like organizing life, reading Slack, and scheduling meetings. He reserves higher effort levels for more complex tasks like building prototype applications.
Jason Liu details how he uses Codex for operational tasks, including automations that read Reddit, Twitter, and LinkedIn to gather external feedback and identify issues. He also utilizes 'chief of staff' threads for daily and weekly task management.
Jason Liu explains his workflow within Codex, where most tasks are managed as long-running pinned threads due to Codex's effective context compaction. This allows the AI to retain previous information without losing context over extended periods.
Jason Liu discusses his use of Obsidian as a knowledge base for Codex, with his 'Vault' project serving as the primary interface. He clarifies that Obsidian primarily stores markdown files, and while the visual aspect is useful, the AI agent uses it mainly for context.
Jason Liu describes his 'chief of staff' automation in Codex, which runs multiple times a day to read his Slack, Twitter DMs, unread emails, and Linear board. The system summarizes his to-do list, updates statuses, and flags important items like upcoming travel or package pickups.
Jason Liu shares a practical application of Codex: automatically checking flight details and texting boarding passes to his personal phone. This feature, set up over time, helps him manage travel efficiently by providing timely updates and reducing the need to manually check flight status.
Jason Liu offers advice on how a user can prompt Codex to set up personalized productivity automations, starting with a broad instruction like checking emails, Slack, and Linear. He emphasizes that refining these prompts over time, based on preferences and desired outcomes, is key to unlocking the tool's full potential.
Jason Liu demonstrates the advanced capabilities of Codex by describing how it can be prompted to create a pull request, monitor its tests, and then send a Slack DM to a specific contact upon successful completion. This showcases the AI's ability to handle complex, multi-step workflows.
Jason Liu highlights the extensibility of Codex through custom skills, showcasing examples like an 'email me' skill, a 'tweet me' skill, and skills that convert voice transcripts into blog posts or videos into video essays. These custom solutions enhance personal productivity and content creation workflows.
Jared Zoneraich, building in residence at Cognition, advises companies building AI agents to focus on leveraging the inherent progress of underlying models rather than relying on complex prompt engineering "hacks." He suggests that as models improve, the need for intricate prompt instructions will diminish, and companies should build their roadmaps accordingly.
Jared Zoneraich highlights that 'tool engineering' will become a crucial skill in developing AI agents. He posits that the focus should shift to identifying and providing the right tools for models, with the assumption that the models themselves will become adept at figuring out how to use them.
Jared Zoneraich discusses what makes a company's AI agent development defensible, suggesting that proprietary data is one aspect. He notes that many companies are currently 'AI wrapper companies' primarily focused on prompts and skills, but true defensibility may lie elsewhere.
Jared Zoneraich describes Cognition's approach to introducing AI agents to enterprises as a 'forward deploy motion.' This strategy prioritizes demonstrating the magic of agents by using them to solve actual customer problems, thereby proving their value by taking tickets off client boards.
Jess Yan from Anthropic explains the evolution of AI agents from simple prompt-response loops to autonomous, self-discovering actors capable of long-running tasks. These agents can now interact with third-party systems and sensitive data, requiring enhanced permissioning and observability.
Jess Yan defines an AI agent as an evolution from simple prompting loops to systems with access to tools, memory, and third-party systems, necessitating components like the model, system prompt, and a 'harness' for scaffolding and control.
Jess Yan argues that maximum performance for AI models is achieved when the model and its harness are developed together. She notes that models are always tested in conjunction with specific harnesses, creating a symbiotic relationship.
Anthropic's Claude Managed Agents are designed as a pre-built harness and infrastructure to enable complex tasks at scale with high return on effort. The goal is to provide easy-to-stack primitives and flexible APIs for delegating work that might otherwise take days or weeks.
Jess Yan demonstrates an Anthropic Managed Agent configured for analysis on a fictitious grocery store dataset. The agent utilized a pre-built toolset, interacted with its file system, and produced three HTML files summarizing the analysis results after being given specific data schemas and task guidance.
Jess Yan explains that Anthropic agents can be connected to third-party systems, such as internal databases, through MCP (Messaging and Communication Protocol). MCP provides a standardized way to communicate with external services and includes an authentication layer for secure access.
Jess Yan contrasts basic prompting loops with more advanced self-recovering agents. She highlights that synchronous prompting loops are highly dependent on prior successful requests and struggle to pivot or handle errors gracefully, whereas self-recovering agents can manage errors and steer back on course.
Jess Yan explains that for AI agents to produce predictable outputs, especially when handling complex tasks, it's beneficial to be prescriptive with initial prompts. This includes providing data schemas and breaking down tasks into discrete segments, which also aids in debugging.