Recap: Mural’s 30-minute GPT-Live-1 sessions cost about $3–$5

Today's top 18 insights for PM Builders, ranked by relevance from YouTube, X, and GitHub.

Recap: Mural’s 30-minute GPT-Live-1 sessions cost about $3–$5

#1 ▶️

1,000 days of Duolingo didn't work. So he built his own app with Codex.

How I AI Podcast

William created Mural with Codex, a native iOS and Android Norwegian conversation app using Swift for iOS, Kotlin for Android, and GPT-Live-1 for real-time voice tutoring. GPT-Live-1 voice sessions cost approximately $3 to $5 per 30 minutes; William allocated a $1,000 OpenAI budget for users, with about $100 remaining at the time of the interview.

#2 𝕏

Mistral AI demonstrated Voxtral 3 Transcribe at AI Engineer Paris, using it to capture every main-stage talk in real time. Context biasing with speaker names and technical terms from the agenda improved accuracy, and the technology is available for other events and applications.

#3 𝕏

NVIDIA AI announced new NVIDIA Dynamo deployment recipes for SKtelecom’s A.X K2, LG AI Research’s K-EXAONE 2.0, UpstageAI’s Solar Open 2, and Motif Tech’s Motif-3. The recipes deliver up to 2x inference performance on NVIDIA B200 GPUs, though the baseline and methodology were not specified.

#4 ▶️

Why PMs are suddenly valuable again | Tulsee Doshi (Head of Product, Gemini)

Lennys Podcast

Tulsee Doshi shared her frontier-AI product framework: a 60-20-20 feature mix spanning 60% that works with current models, 20% that is usable but imperfect, and 20% targeting early capabilities. Doshi recommends starting initial evaluations with five prompts and expanding to roughly 20 that distinguish between two models. Google’s evaluation updates included replacing the six-tone Fitzpatrick scale in some evaluations with Professor Monk’s 10-tone skin-tone scale.

Also covered by: @Lenny Rachitsky

#5 𝕏

NVIDIA AI shared that Hyundai Motor Group is exploring NVIDIA NemoClaw and OpenShell, while @AIMSafety_ is integrating OpenShell to help businesses use AI agents securely. @LG_AI_Research is opening API access to K-EXAONE-2.5, which was trained on NVIDIA Nemotron datasets.

#6 ▶️

OpenAI VP: How AI Changed My Family’s Fight Against Cancer | Ashley Alexander

Peter Yang

Peter Yang shared how Ashley Alexander used a project-only-memory ChatGPT project to organize her father’s medical records, interpret treatment information in real time, and prepare focused questions during appointments that she says typically last about 15 minutes in the U.S. The siloed setup kept her father’s health details separate from her own conversations, highlighting a product pattern for managing sensitive, context-specific information.

Also covered by: @Peter Yang

#7 🐙 garrytan/gbrain

garrytan/gbrain - **garrytan/gbrain** is Garry Tan’s MIT-licensed, self-hostable memory system for sharing sourced, correctable facts across AI agents, with optional search, synthesis, and knowledge-graph features. The repository includes agent and plugin directories, documentation, tests, and security materials; it has 31,917 GitHub stars and 215 daily trending gains. Worth trying if you need portable, user-controlled memory across multiple agents.

#8 🐙 ZhuLinsen/daily_stock_analysis

ZhuLinsen/daily_stock_analysis - ZhuLinsen/daily_stock_analysis is an MIT-licensed LLM stock-analysis app for six markets and ETFs, with API, bot, strategy, evaluation, web/desktop, backtesting, portfolio, and notification components. It has 68,526 stars and 505 daily trending gains. Worth trying if you need starter code for a multi-market investment research and alert product.

#9 🐙 homeassistant-ai/ha-mcp

homeassistant-ai/ha-mcp - homeassistant-ai/ha-mcp lets Claude, ChatGPT, and other AI assistants configure and automate Home Assistant. The MIT-licensed repository includes an in-process integration, an app/add-on, and Docker and PyPI/stdio options; it has 5,617 GitHub stars with a daily gain of 31. Worth trying if you need an AI interface for Home Assistant; the recommended integration requires Home Assistant 2026.8.0 or newer.

#10 𝕏

Yann LeCun commented that the “P” in JEPA stands for “predictive,” and that the key distinction is between joint embedding and generative architectures—not JEA and JEPA. He added that whether a predictor is necessary depends on the task.

#11 𝕏

Garry Tan said he makes his own evals to run in Capy, Claude Code, or Codex as needed, sharing his garrytan/gbrain-evals GitHub repository as an example.

#12 𝕏

claire vo 🖤 commented that custom furniture—even if AI-designed—will be very expensive, except perhaps small 3D-printed trinkets. Drawing on her small San Francisco home, she analyzed custom-furniture costs, AI design limitations, and why IKEA remains an optimal option.

#13 𝕏

Jason Zhou shared a post about using @treg_ai to build an in-house go-to-market agent.

#14 𝕏

Kevin Yien commented that tolerance for rapid product changes is uneven: Codex could change its navigation hourly and be “fine,” while companies with “important” user bases may move more cautiously than newer players—a version of the innovator’s dilemma.

#15 𝕏

Garry Tan said his gstack-evals project replicated an unspecified result from a Jev Engineering article and reported that particle-1 performed better, though he provided no criteria or measurements.

#16 𝕏

Udi Menkes 🚢 shared that Intuit’s AI Business Advisor, which he leads, includes peer benchmarking similar to NADA 20 Groups, where car dealers compare P&Ls and learn from one another’s numbers. The product compares a business with similar businesses, highlights strengths and areas for improvement, and advises on next steps.

#17 𝕏

clem 🤗 shared @victormustar’s observation that more individuals are uploading small fine-tunes on Hugging Face, citing dale, a 0.8B Qwen-related fine-tune, as an example.

#18 ▶️

Could AI Have Predicted The November 5 Stock Market Crash?

All About AI

In a retrospective demonstration using archived news and market data available as of midnight before November 5, 2026, the Sikt Intelligence model forecast a roughly flat S&P 500 close. The video reports that the index instead fell 4.3% and the Nasdaq fell 5.2%. After being told of a 4% S&P 500 drop, the model ranked an AI and big-tech sell-off as the likeliest explanation, despite having assigned such a one-day decline about a 0.7% probability.

Recap: Mural’s 30-minute GPT-Live-1 sessions cost about $3–$5

Get tomorrow's brief first