2025-09-18 AI Mastermind
Table of Contents
Session Overview
The September 18th session had two high-value threads: the introduction of Perplexity Spaces as a searchable, conversational knowledge base for AIMM session content, and a deep dive on GPT-5 prompting techniques — specifically the newly exposed reasoning effort and verbosity parameters that allow users to control the model’s internal compute allocation and output length.
Lou had uploaded all available AIMM transcripts into a shared Perplexity Space, demonstrating live that it could retrieve specific technical recommendations from past sessions with strong accuracy and full citation. This positioned Perplexity Spaces as a preferred alternative to Google NotebookLM for community knowledge management — more conversational, better at follow-up queries, and capable of combining knowledge base content with live web search.
The second thread was a guided tour of OpenAI’s GPT-5 prompting guide and the Playground’s new prompt optimizer and versioning features. Lou demonstrated how to use the prompt optimizer to restructure a prompt according to GPT-5’s preferred template, and showed the version history and evaluation capabilities that allow prompts to be managed as evolving assets rather than static text strings. The session introduced the concrete technique of setting reasoning_effort and verbosity as the first two lines of any system prompt.
High-Signal Moments
- Perplexity Spaces demonstrated live: query “what open source software is recommended for running LLM?” returned Llama.cpp, LM Studio, JAN, LiberChat, Chatbot UI with full transcript citations — “a stellar job of figuring out making sense of the content”
- Perplexity Spaces preferred over Google NotebookLM: “NotebookLM seems to be very perfunctory and concise. This one seems to be more conversational.”
- GPT-5 prompting key finding: the model has an internal routing layer that determines reasoning intensity; this is now user-controllable via reasoning effort parameter (minimal / low / medium / high)
- Verbosity is separately controllable from reasoning effort — “even though it doesn’t return a lot of context through its reasoning, it still may infer and generate a whole lot of context for its response”
- Stop conditions introduced as a prompt technique: “once you have something within 85% confidence level, stop” — prevents the model from over-searching and filling context with low-confidence material
- Prompt optimizer in OpenAI Playground demonstrated: converts rough prompts into GPT-5’s preferred template (goal, plan, instructions, completion summary, constraints, stop conditions, output format)
- Prompt versioning in Playground: each prompt version has a specific API-callable ID — “when you change it here, all of your agents now have the new functionality”
- Reminder to move to OpenAI Responses API (replacing Completions API) for any API work — estimated deprecation of Completions in under a year
- Community event: Jamie thanked Elizabeth and Kasimir for peer support on survey/assessment work — strong community collaboration noted
Open Questions
- The 50-file limit in Perplexity Spaces means content will need to be amalgamated after approximately one year’s worth of sessions — what is the best compression strategy (summaries vs. merged transcripts)?
- Does the reasoning_effort parameter work through the chat UI, or only reliably through the API?
- How should existing custom GPT-5 instructions (set a year or more ago) be updated to incorporate reasoning effort and verbosity defaults?
- What is the right reasoning effort level for different coaching use cases — client intake analysis (high?), newsletter drafting (low?), strategic planning (high?)?
- Will Perplexity Spaces remain available to free-tier users, or will it require Pro?
Suggested Follow-Through
- Lou: Share the Perplexity Spaces link in Telegram with instructions for joining and how to use the knowledge base for session retrieval
- Lou: Update personal GPT-5 custom instructions to include reasoning effort: medium and verbosity: medium as defaults
- Lou: Share the GPT-5 prompting guide link (platform.openai.com/docs/guides/gpt-5) in Telegram
- Members: Review existing custom GPT-5 instructions in settings — update with reasoning effort and verbosity parameters; consider whether any prior instructions are now superseded by GPT-5’s improved instruction following
- Members: Try the Perplexity Spaces knowledge base on a real question from a prior session and compare the quality of retrieval to what you remember from the session
Additional Resources
Links & Tools Shared in Chat
- [Posha — AI-powered personal chef robot system] — https://www.posha.com/ (shared by Donald Kihenja, in response to discussion of personal chef bots)
- [Comet browser — AI-native browser (group recommendation)] — (no URL shared in chat; referenced by name — Elizabeth Stief sent invite to Ri Ca via Telegram)
- [Adobe Firefly — image editing tool recommended as alternative to nano banana for image editing] — (referenced by Elizabeth Stief)
Books & Articles Mentioned
- None
Ideas from Chat
- Donald asked whether nano banana (likely Deepseek) could edit images it generated — group found it couldn’t reliably; Elizabeth Stief recommended Adobe Firefly as an alternative for image editing
- Donald: “The future of health is happening already — they are discovering and repurposing drugs, micronutrients WAY faster with AI” — AI in drug discovery as a high-signal theme
- Ri Ca: investing in companies that use AI to explore theoretical proteins and select candidates for human receptor bonding — early indicator of AI in biotech investment
- Lou shared his full production system prompt with slash command library in chat — see Insight - Slash Command System Prompts as a Personal AI Interface
Derived Artifacts
- sourdough-excavator (Sourdough Excavator — excavating differentiated expertise)