# Inner Speech Essay - Planning and Structure Planning document for potential Substack essay on [[inner speech]] and LLM dialogue. ## Essay Structural Options ### Option A: Two Separate Essays (Series) **Essay 1: "Language Helps Us Think" (The Inner Speech Piece)** - Autobiographical hook: [[Claude Code]] experience - Engage with Riley/Verge piece: "Yes, language isn't intelligence. But language helps us think." - Inner speech literature: [[Keith Frankish]]'s cyclical model, [[Lev Vygotsky]], [[Andy Clark]] - LLM dialogue as "hybrid inner speech" - End with: "So what do we *do* with this? That's for next time." **Essay 2: "LLMs and Understanding" (Existing Draft)** - Picks up where the first left off - Oracle vs. understanding - [[cognitive traction|Cognitive traction]], comprehension stretcher - The aesthetics of understanding ([[Elisabeth Schellekens]]) - Practical implications **Pros:** Clean separation; each essay has one core move; builds readership anticipation **Cons:** Readers might not read both; some repetition ### Option B: Combined Essay (Ambitious) **Part I: The AGI Skeptics Are Right (But Missing Something)** - [[Benjamin Riley]]'s argument: language ≠ intelligence - The hype correction (MIT Tech Review) - Concede the point—then pivot **Part II: Language Helps Us Think** - Clark, [[Gary Lupyan]], Frankish, [[Peter Carruthers]] - Inner speech as cognitive tool - The cyclical model of reasoning **Part III: LLM Dialogue as Hybrid Inner Speech** - The [[double-extension structure]] - Reciprocal prompting and loopiness - Unauthoredness as feature - Cognitive traction **Part IV: From Theory to Practice** - Oracle vs. understanding - Comprehension stretcher - The aesthetics of understanding **Pros:** Full argument in one place; more intellectually satisfying **Cons:** Long; might lose casual readers; harder to write ### Option C: The Personal Essay Frame **Hook:** "I've been using Claude Code for a week. It's changed how I think." **The Observation:** Not claiming it's intelligent. Not claiming it understands me. Something weirder: it feels like an externalized version of talking to myself. **The Puzzle:** Why would a language model feel like inner speech? **The Investigation:** (weave together theory and personal experience) - What is inner speech? (Vygotsky, Frankish) - What does it do? (Trigger autonomous processes, decompose problems) - How does LLM dialogue do something similar? **The Insight:** LLM dialogue is "hybrid inner speech"—privacy of self-talk, novelty of conversation. **The Implication:** We're not talking about AGI. We're talking about a new cognitive tool. Like writing. Like calculation. Like inner speech itself. **Pros:** Engaging; accessible; voice comes through **Cons:** Less rigorous; might feel lightweight for philosophy audience ### Option D: The Counterintuitive Lede **Open with:** "The AI skeptics are right. Language models will never be intelligent." **Beat:** "But that's not the interesting question." **Pivot:** "The interesting question is: what happens when you couple a language model with a mind that *is* intelligent?" **Body:** Develop the inner speech parallel **Closing:** "LLMs aren't artificial minds. They're cognitive prosthetics. And like all prosthetics, what matters isn't what they are—it's what you can do with them." **Pros:** Grabs attention; clear positioning; memorable framing **Cons:** Might alienate some readers who want to hear about AGI; oversimplifies --- ## Autobiographical Hook Options ### Option 1: The Claude Code Discovery "A week ago I started using Claude Code. I expected a fancy autocomplete. What I got was something stranger: conversations that felt like talking to myself, except the other side could surprise me." ### Option 2: The Philosophy Angle "I've been doing philosophy with LLMs since 2022. Not asking them for answers—more like thinking out loud with a very patient interlocutor who never judges my half-formed ideas." ### Option 3: The Contrast "When I use ChatGPT for recipes or code, it feels like a search engine with manners. When I use it for philosophy, it feels like something else entirely—like an externalized version of the voice in my head." ### Option 4: The Specific Moment "Tuesday afternoon, I was stuck on a problem about [X]. I started typing into Claude, not expecting an answer, just wanting to articulate the problem. Three exchanges later, I had the insight. But here's the thing: I couldn't tell you which part came from me and which from the model." --- ## Potential Titles - "Hybrid Inner Speech: What LLMs Actually Do for Thinking" - "Language Helps Us Think (And So Do Language Models)" - "The Double Extension: LLMs and the Medium of Thought" - "Not Intelligence, But Traction" - "Talking to Yourself, Externally" - "Cognitive Prosthetics: What AGI Skeptics Miss" - "The Unauthored Interlocutor" --- ## Conceptual Moves ### 1. The Double-Extension - First extension: Language is cognitive technology (Clark, Lupyan) - Internalization: Inner speech is contracted outer speech (Vygotsky) - Second extension: LLM coupling extends through the same medium ### 2. The Loopiness as Mechanism - Frankish's cyclical model: Produce → Perceive → Interpret → Respond → Repeat - In LLM dialogue, the loop is split between user and model - The LLM plays the role autonomous processes play in inner speech ### 3. Unauthoredness as Phenomenological Feature - Both inner speech and LLM text have attenuated authorship - Similar interpretive stance for both - "Enough otherness to break cognitive ruts, enough nothingness to prevent social anxiety" ### 4. Cognitive Traction as Original Contribution - The metaphor: gravel gives tires purchase on road - LLM words provide surface for thought to push against - Not adversarial—just grit for forward motion ### 5. Riley Engagement Options **Option A: Agree and add** "Riley's right that language isn't intelligence. But he's looking at the wrong question. The question isn't whether LLMs think—it's whether coupling with them changes how *we* think." **Option B: Complicate** "Riley says language isn't thinking. But the inner speech literature suggests it's more complicated. Language *participates* in thinking, even if it doesn't constitute it. So a language model might participate in thinking too—not as a thinker, but as a tool." **Option C: Redirect** "The AGI debate is a distraction. Whether LLMs will ever be 'intelligent' is less interesting than what they're doing right now: functioning as cognitive prosthetics that extend the same mechanism by which language already extends cognition." --- ## Questions to Consider 1. **Audience:** Philosophy colleagues? Substack general readers? Both? 2. **Tone:** How personal do you want to go? The Claude Code experience is compelling but might feel too "tech enthusiast" for philosophy readers. 3. **Scope:** Do you want to stay at the level of phenomenology and cognitive science, or engage with the [[extended mind]] literature more formally ([[Orestis Palermos|Palermos]], Clark, etc.)? 4. **The Schellekens material:** Does the aesthetics-of-understanding angle belong in this piece, or is that a separate argument? 5. **Practical payoff:** Do you want to end with implications for how to use LLMs, or keep it purely descriptive/theoretical? --- ## Related Notes - [[Inner Speech and LLM Coupling - Research Discussion]] - Core theoretical material - [[LLMs and Understanding]] - Existing Substack draft on oracle vs. understanding - [[ChatGPT Extended - Large Language Models and the Extended Mind|Smart, Clowes & Clark (2025)]] - Extended mind literature on LLMs