OpenAI Launches ChatGPT Pulse
Today's curated insights on AI product management from X/Twitter across 60+ expert sources and YouTube channels.
OpenAI Launches ChatGPT Pulse
From X
AI Product Launches & Updates
ChatGPT Pulse launch: Sam Altman @sama introduced ChatGPT Pulse, a proactive feature delivering personalized daily updates from your chats, feedback, and connected apps like your calendar, rolling out to Pro users on mobile today.
GDPval evaluation unveiled: OpenAI @OpenAI announced GDPval, a new benchmark measuring AI performance on real-world, economically valuable tasks, grounding progress in evidence rather than speculation.
Gemini Robotics 1.5 introduction: Sundar Pichai @sundarpichai unveiled Gemini Robotics 1.5, enabling robots to reason, plan ahead, use digital tools like Search, and transfer learning across embodiments as a step toward general-purpose robotics.
AI Tools & Applications
Perplexity Search API: Arav Srinivas @AravSrinivas launched the Perplexity Search API to deliver millisecond-latency search results for grounding LLMs and agents with real-time web data via a custom search index.
Thought-to-post pipeline: Rowan Cheung @rowancheung shared his favorite AI workflow—using an AI-powered voice dictation app during walks to capture ideas and then auto-generate optimized posts in his writing style, boosting productivity.
Gemini CLI integration: Philipp Schmid @_philschmid noted that the Gemini CLI is now included in Google AI Pro & Ultra subscriptions at no extra cost, centralizing command-line access.
Product Management Insights & Strategies
Building AI evals: Lenny Rachitsky @lennysan highlighted how @HamelHusain and @sh_reya teach the world’s most popular AI evals course, demonstrating live how to design model evaluations—a critical skill for AI PMs.
LLM persona for career growth: Madhu Guru @realmadhuguru advised setting up an LLM with your target persona, feeding it raw ideas to critique and periodically extracting its refined prompts to accelerate learning.
Cart-before-horse development: Aakash Gupta @aakashg0 argued that with prototyping costs plummeting, PMs are flipping the traditional sequence—building early prototypes before defining detailed requirements.
AI Industry Developments & News
Psychotherapy app ban: DeepLearningAI @DeepLearningAI reported that Illinois became the second U.S. state to ban AI apps from administering psychotherapy without direct doctor participation under the Wellness and Oversight for Psychological Resources Act.
Wayfinding AI research: Google Research @GoogleResearch shared user insights from Wayfinding AI, a prototype agent guiding users to better health information through proactive conversational guidance and goal understanding.
AI agents in education: NVIDIA AI @NVIDIAAI forecasted that the next generation of students will be learning, creating, and leading with their own AI agents, highlighting collaborations between @iamwill, @ArizonaState, and NVIDIA technology.
From YouTube
Why AI evals are the hottest new skill for product builders | Hamel Husain & Shreya Shankar
Lennys Podcast • September 25, 2025
Hamel Husain and Shreya Shankar walk through their data-driven framework—starting with manual error analysis on sampled AI app traces, then open and axial coding, and finally LLM-as-judge evaluators—to help product teams build, prioritize, and automate evals that actionably improve AI applications.
Key Takeaways:
- Begin with open coding: review ~100 randomly sampled traces, note the first visible error per trace (e.g., hallucinating a “virtual tour”), and stop when “theoretical saturation” yields no new failure modes.
- Use an LLM to cluster open codes into axial codes (actionable categories like “tour scheduling errors” or “formatting errors”), then pivot and count occurrences to prioritize the highest-impact issues.
- Automate checks via code-based evaluators for simple, verifiable outputs and LLM-as-judge evals for nuanced failure modes—each defined as a binary pass/fail prompt aligned against human labels before integration into CI and production monitoring.
AI Avatar Models Are Getting INSANE Powerful - Testing Workflows
All About AI • September 25, 2025
The video demonstrates character-swap tests using NVIDIA’s van animate-replace and ByteDance’s Omnihuman 1.5 avatar models—applying static images to video clips with lip-synced speech—and presents an end-to-end workflow employing OpenAI speech-to-text and 11 Labs voice synthesis, with code in the nano_banana repository.
Key Takeaways:
- NVIDIA’s 14B van animate-replace model performed a 15-second character swap in under 20 minutes, preserving tight lip-sync and realistic head movements when overlaying a target image on source footage.
- ByteDance’s updated Omnihuman 1.5 rendered a high-fidelity female avatar holding the NVIDIA coffee mug and exhibited context-aware animations like coordinated sipping motions on the “take a sip” cue.
- The presenter outlined a fully automated pipeline that uses OpenAI’s speech-to-text to transcribe audio, 11 Labs for voice cloning, and the van avatar API for video replacement, with a plug-and-play script hosted in the nano_banana GitHub repo for members.