01 / Show notes
What happens when a model does not chat, but makes the same small decision over and over? And why does working with agents suddenly make an ultrawide monitor sound tempting?
We begin with Jev: filtering memories, tagging contacts, semantic queries and spotting advertising. These are workflows built around frequent judgments rather than conversation. A key distinction runs through the discussion: matching an output type does not guarantee a correct decision.
We then turn to multimodal tools, local models and creative work: using GPT-6 Astra to help produce training data, making promotional videos and music with code, and the very practical problem of projects filling up a hard drive.
Browsers, chats and task lists compete for screen space. From that everyday friction, we explore voice input, multiple accounts, persistent computers, cloud Macs and enterprise data boundaries. Better capabilities do not remove the need for a workable environment and human oversight.
Recorded online via Tencent Meeting on September 19, 2026. Experiences, opinions and predictions reflect the participants’ views, not independent verification of product capabilities.
Questions from the episode
- Where do high-frequency decisions fit better than chat?
- Why are structured output, accuracy and “no hallucinations” different claims?
- How do small models and multimodal tools fit into creative work?
- Do agents make an ultrawide necessary—and what about storage?
- What boundaries should govern a persistent agent’s accounts and data?
Chapters
- 00:00 Intro
- 00:49 Highlights
Jev: beyond chat
- 01:30 Opening
- 01:41 Jev and structured decisions
- 03:47 Contact labels and frequent decisions
- 07:22 Semantic search and database queries
- 08:47 Valid outputs are not necessarily correct decisions
- 17:18 Filtering disguised advertising
Models, tools and multimodality
- 20:57 Connecting new models to agents
- 25:07 Understanding audio and video together
- 27:51 Quantization and local models
- 31:43 SVG, icons and design workflows
- 35:19 Using GPT-6 Astra to create training data
- 40:59 Realtime voice and computer use
- 43:44 Embodied AI and zero-shot tasks
Creating with more capable AI
- 48:07 When vibe coding fills the disk
- 56:56 Making promotional videos and music with code
- 01:09:39 Voice input and realtime feedback
Agent environments and boundaries
- 01:20:07 Multiple accounts and workspaces in Codex
- 01:24:29 Persistent agents and content distribution
- 01:26:42 Coding and office workflows
- 01:29:52 Cloud Macs and agent environments
- 01:30:49 Enterprise data and logging boundaries
Content creation and research automation
-
01:36:25 AI series and content production
-
01:38:06 AI-assisted research and human oversight
-
01:40:04 Outro
References
People and terms
Guizang moderates this episode, joined by co-hosts Xiangyang Qiaomu, Orange and Yang Pan. All four are the show’s co-hosts, not external guests.
Jev’s type constraints and decision accuracy are separate questions. Computer Use means performing actions through a computer interface; it is not itself a determination of AGI.
Production credits
Next Token Weekly is a weekly roundtable hosted by Yang Pan, Guizang, Orange and Xiangyang Qiaomu, exploring AI technology, products and their real-world effects.
Music: Electric Dreams / Origami by Scott Buckley, used under CC BY 4.0. Scott Buckley Music Library.
Community
Join Next Token Weekly’s second WeChat listener group to discuss AI tools and real workflows. Scan the QR code below before October 4, 2026.
Join Next Token Discord to share your experiences and perspectives.
