Summary

A 31-minute walkthrough (originally a live stream) from Kyle Balmer of AI with Kyle, on building an 'AI second brain' using OpenAI's newly released GPT-6 Astra inside the Codex mode of the ChatGPT desktop app. Its argument is that the second brain is less complicated than its reputation: a folder of plain-text markdown notes that an AI builds by interviewing you. He then adds four layers on top: several models sharing the folder, connected tools, computer use, and an always-on dedicated machine reached from your phone. He openly dismisses Obsidian knowledge graphs as looking productive more than being useful, and dismisses paid templates. The video funnels into his own free guide, newsletter and paid programme, and he says he has charged for this training before. His claims about Astra's computer use are first-hand enthusiasm from launch week, not tested results.

• Definition: a second brain is 'a folder with a bunch of text files in it'. It holds notes about you, your projects, preferences, past decisions and useful conversations, so that any AI can use your context instead of giving generic answers, and you can switch models without explaining yourself again.
• Have the AI build it. Use a powerful model (he uses Astra in Codex mode and says Claude Code or Fable work equally well) and start with 'Hey Astra, I want to set up a second brain'. Ask it to explain in plain language and interview you before deciding the structure. Don't buy someone else's template: his own includes a large section on moving from the UK to Cyprus, which would be no use to anyone else.
• Spend most of the setup time on the interview: one question at a time, three or four levels of 'why', and a 'be / do / have' framing for goals. Push vague goals ('a more successful business') down to specifics. Answer by voice rather than typing so you don't edit yourself as you go, then have the AI play back its understanding for you to correct. Be selective about what you add, because dumping gigabytes of files clogs it.
• On Obsidian: it can open the same folder as a 'vault' and draw a graph of it, but he says he never looks at his. In his view the files are 'primarily there for the agents'.
• Several models can share one brain: ChatGPT, Claude, Gemini, Grok, Chinese or local models. Telling Claude Code 'I have a second brain, it's located here, use it from now on' is enough. His pattern is to use the expensive model (Astra, which he calls 'token hungry') to design and reorganise the brain, and cheaper tiers for daily work. The captions give these as 'soul or terror', probably model names, and the exact names are uncertain. One caveat he stresses: models share summary notes, not each other's full conversations.
• Maintenance: a scheduled task every night pulls the day's work into a combined note, and every couple of months a strong model is asked to consolidate the brain.
• Tools come next, connected through the harness (Codex or Claude Code), not to the brain directly. Some have direct integrations (he says Codex and ChatGPT now take 'pretty much unlimited' Gmail accounts), and others need an MCP connection or an API key, which he says the AI will walk you through.
• Computer use is the layer he is most excited about. He says Astra is 'scarily good' at it. While he built the slides for this video, he says Codex was working in three Chrome tabs at once: pulling his website analytics, finding his weakest YouTube thumbnails and generating replacements to A/B test in YouTube Studio, and diagnosing a hard drive failure, which he says it fixed. He also says his published YouTube videos are now edited entirely by AI through computer use, and suggests teaching a weekly task by recording a narrated screen video of yourself doing it.
• The always-on machine: for about six months he says he has run Codex on a dedicated Mac Mini at home ('around $500'; he notes RAM is currently expensive, which makes the Mini good value). He sends it tasks by phone, including voice notes at bedtime that come back as a prototype app by morning. He prefers this to a VPS (rented server), which he says is more technical, lacks local apps and logins, and is less secure. A laptop is the wrong choice because it stops when the lid closes. Codex and Claude Code both offer remote control from a phone, paired with a QR code.
• His summary of the stack: a brain folder, a harness (Codex or Claude Code), connected tools and computer use, an always-on computer, and your phone as the remote. His advice is not to build it all at once but to start with the interview.

Why it matters

This describes, from the outside, the shape of how Ashley already works: a folder of plain-text context, Claude Code as the harness, connected tools, scheduled overnight jobs and remote access. So its main value is as independent confirmation of that pattern, and as one creator's case that the folder matters more than the graph view. Two points are worth carrying forward. One is the split between an expensive model that designs and reorganises and cheaper models that do daily work. The other is his caveat that shared memory means shared summaries, not shared conversations. The computer-use claims about Astra (three tabs at once, fully AI-edited videos, a repaired hard drive) are launch-week first-person reports from a creator with a course to sell. Log them as reported until a test or a second source confirms them.

Fc5PcxMPigk-transcript.txt