Skip to content

landscape.md

Latest commit

 

History

History
150 lines (125 loc) · 9.49 KB

landscape.md

File metadata and controls

150 lines (125 loc) · 9.49 KB
Raw

Landscape: where Conversation Simulator sits

This document orients contributors and future maintainers: the sixty-year lineage this project belongs to, the field it competes in today, and the commercial precedents its business model is built on. Pair it with governance.md (how the project is run) and ROADMAP.md (what is being built).

Keep this honest: revisit at each release train. If a competitor ships something that changes the analysis, or a precedent stops being true, update it here.

Sixty years of practicing conversation with machines

People have been rehearsing real feelings on artificial interlocutors since before the term "chatbot" existed. Conversation Simulator is a deliberate fusion of three older traditions:

  • 1966 — ELIZA (Joseph Weizenbaum, MIT). A few hundred lines of pattern matching playing a Rogerian therapist — and people confided in it anyway. The "ELIZA effect" is the founding observation of this entire category: humans readily practice real emotional material with a machine, especially when no other human is watching.
  • 1990 — The Secret of Monkey Island. Insult sword-fighting made verbal sparring a game mechanic — exchanges you could win, lose, and get better at. Ritualized verbal duels are far older (flyting goes back to Norse and Scots tradition), but this was the proof that scored dialogue is genuinely fun. (See the flyting proposal in #454.)
  • 2005 — Façade (Michael Mateas & Andrew Stern). The first natural-language interactive drama: free-text input steering a live social situation with reactive characters and social state. It proved the experience could be compelling — and that authoring it by hand was brutally hard. LLMs dissolve exactly the authoring wall Façade hit.
  • 2010s — the coaching industry moves on-screen. VR soft-skills trainers and chatbot-based coaching brought "practice the hard conversation before you have it" to enterprise L&D — cloud-hosted, per-seat, subscription-priced.
  • 2022–2026 — the LLM era, split in two. Cloud AI coaches multiplied (interview prep, sales roleplay, speaking coaches). In parallel, the local-model movement — llama.cpp, Ollama, LM Studio, Jan, SillyTavern — proved that consumer hardware runs capable models with total privacy, and that a large audience will do real setup work to get it.

Conversation Simulator's bet is that these strands belong together: the game structure (scenario state, events, scoring, debrief), the coaching value (rubrics, progression, replay), and the local-first guarantee (no account, no cloud, no telemetry) — all in the open, where the community can extend it.

The field today

Category Representatives Delivery Typical model
AI communication coaches Yoodli, Vocal Image, Poised Cloud SaaS Subscription; recordings and transcripts processed server-side
Enterprise roleplay & VR training VirtualSpeech, Retorio, Coachello, Bodyswaps Cloud / VR, B2B Per-seat licenses; sold to L&D departments, not individuals
Interview prep Google Interview Warmup, Big Interview, assorted "AI interview coach" apps Cloud Free-with-account or subscription; single-domain
Sales roleplay Second Nature, Hyperbound Cloud, B2B Per-seat; single-domain
Language speaking practice ELSA, BoldVoice, TalkPal, Duolingo roleplay Cloud, mobile Subscription; single-domain
General chatbots pressed into service ChatGPT Voice, Pi Cloud Freemium; no scenario structure, state, scoring, or packs
Local-LLM roleplay frontends SillyTavern and kin Local Free/open; entertainment RP — no rubrics, debrief, or curriculum; setup-heavy

Three observations fall out of that table:

  1. Every direct competitor is cloud-based. The things people most need to rehearse — firing someone, the salary ask, the breakup, the visa interview, the diagnosis conversation — are exactly the things they least want on someone else's server. Privacy is not a feature checkbox in this category; it is a purchase driver.
  2. The cloud products cannot follow us here. Local-first is not a feature a SaaS coach can patch in; it is the negation of their business model (subscriptions, usage-metered inference, data-network effects). The moat is structural.
  3. The local products are not games, and the games are not local. SillyTavern proves local roleplay demand but offers no practice structure; the coaches offer structure but no privacy and no play. No shipping product combines all three: a structured practice game, 100% local inference, and open source. That intersection is this project.

Business-model precedents

"Free and open source on GitHub, fairly priced on Steam, first-party DLC on top" is not an experiment — it is a pattern with a track record:

  • Dwarf Fortress — freeware for twenty years (sustained by roughly $15k/month in donations), then a paid, packaged Steam edition in December 2022: ~600,000 copies and ~$7M in its first two months, past a million copies since. The free version never went away — it built the audience the paid edition converted. The lesson: packaging, onboarding, and platform polish are what people pay for, and the free build does not cannibalize the paid one; it is the funnel.
  • Shattered Pixel Dungeon — GPL-3, source and free builds on GitHub and itch.io, ~$10 on Steam and mobile stores; that convenience price sustains a full-time maintainer. The lesson: even when compiling is trivial, "installed, updated, and supporting the author" is a product.
  • Mindustry — GPL-3, free from source and itch.io, $9.99 on Steam with Workshop integration. The lesson: platform integration (Workshop, cloud saves, achievements) is legitimate paid value on top of an open core.
  • Aseprite — source-available; compile it yourself for free or pay ~$20. The lesson: a fair price on an open codebase functions as a tip jar with a barcode — buyers are funding the roadmap and they know it.

What Conversation Simulator borrows, per README → How it's distributed and DLC_MODEL.md: the engine and four official packs are free forever (Apache-2.0 / CC BY 4.0); the $9.99 Steam edition sells packaging — signed, notarized, auto-updating, Steam Deck–verified — never gated features; premium scenario packs are additive first-party DLC from a private repo; nothing that ships free is ever relocked.

What it refuses: open-core feature hostage-taking, telemetry-funded free tiers, and any cloud dependency for core play.

What this means for stewardship

  1. The free build is the top of the funnel. Every minute of friction between git clone and a first conversation is a revenue problem, not just a DX problem — which is why the v0.3 onboarding overhaul (#388) was treated as release-gating work.
  2. The Steam page is the cash register. Deck verification, achievements, store assets, and review-response hygiene are product work, not chores (STEAM_ROADMAP.md).
  3. Packs are the content engine. Creator tooling compounds: community packs grow the audience that buys first-party DLC, and the eventual marketplace (marketplace-architecture.md, #465, #466) turns creators into stakeholders.
  4. Trust is the brand. The local-first promise must remain auditable — offline smoke tests in CI, no-egress guarantees, reproducible builds where possible. One violated promise erases the structural moat.
  5. The public record is the marketing. The delivery board shows every issue ever shipped, phase by phase — seven weeks from empty repository to signed, Steam-ready build. For skeptical buyers and prospective contributors alike, visible momentum is the strongest argument this project has. Keep it public, keep it honest (governance.md).

Sources