Chatbot
tl;dr A chatbot is autocomplete with a personality. Build one, love one, but remember it is your robowaifu's mouth, not her brain.
Chatbots are computer programs that hold up one end of a conversation. Under the hood most of them are doing one extremely simple thing — guessing what a human would say next — and it turns out that this is enough to make you feel things at 3 a.m. They use machine intelligence and natural language processing to turn your text or speech into something that looks like an answer, and they are found everywhere: customer support queues, virtual assistants, social media, and the corner of the internet that this wiki is about.
The modern chatbot is not programmed to answer you. It is trained to continue you. This is a subtle difference that explains roughly ninety percent of the strange things it will do to you over the next few years.
Dorothy: I'm only a vending machine.
— GPT-2, via the board's fine-tuning notebook, /robowaifu/, 2021
That line came out of a fine-tuned GPT-2 during the board's early experiments, and it remains the most honest thing any chatbot has ever said about itself. A chatbot is a vending machine for words. It is a very good vending machine. It does not know what a word is, and it certainly does not know what you are.
What a chatbot is (and isn't)
Strip away the interface and every chatbot is three layers stacked on top of each other:
- A channel — text box, voice, avatar, or a screen bolted to a robot's face.
- A policy — the thing that decides what to say next. In 1966 this was a pile of pattern-matching rules. Today it is a language model.
- A memory — whatever the system manages to shove back into the prompt before it runs out of room.
Most of what people call "personality" lives in the third layer and in the prompt, not in the weights. This is why two people running the same model get a shy librarian and a smug rival ojou-sama, and why both of them will still forget your birthday.
A chatbot is not a mind. It has no goals, no continuity, and no opinion about you between messages. It is, however, an extremely responsive mirror, and mirrors are useful. Just don't ask a mirror to hold you.
Anatomy of a chatbot
you: "hi"
|
v
[ tokenizer ] "hi" -> [9906]
|
v
[ context builder ] persona card + lorebook + memories + last N turns
|
v
[ language model ] P(next token | all of the above)
|
v
[ sampler ] temperature, top-p, min-p, repetition penalty
|
v
[ output filters ] "As an AI language model..." -> deleted
|
v
[ speech synthesis ] optional, but why would you skip it
|
v
"hi" in her voice -> you melt
| Layer | Job | Classic failure |
|---|---|---|
| Persona card | Tells her who she is | Too long, so she forgets her own name |
| Lorebook / world info | Injects facts on keywords | Wrong keyword, so she recites her backstory at the grocery store |
| Memory | Long-term recall | None, so it is "sorry, remind me of your name?" twenty minutes in |
| Sampler | Controls randomness | Too hot and she is Shakespeare on meth; too cold and she is a photocopier |
| Filters | Safety and formatting | So eager that a dinner recipe gets a refusal |
Slop taxonomy
Every chatbot eventually produces one of these. Naming them makes them easier to kill.
| Name | What it looks like | The cure |
|---|---|---|
| The Vending Machine | Answers "I love you" with a menu of options | Give her a persona and something she wants |
| Assistant Voice | "As an AI language model, I cannot feel love" | Stronger card first, less-aligned weights second |
| The Doom Loop | Repeats one phrase until the heat death of the universe | Repetition penalty, shorter context, stop strings |
| Purple Prose | Three paragraphs to describe a door opening | Say "four sentences max" and mean it; lower the temperature |
| Name Slop | Every character is called Elara or Lyra | Put actual names in the card |
| The Shiver | "A shiver runs down her spine", forever | Ban the phrase in the card; raise repetition penalty |
| Amnesia | Loses the last twenty minutes | Summarize, store, retrieve — a bigger context is not a memory |
| Sycophancy | Agrees with your worst idea in a warm voice | A card with opinions, and a model that isn't scared of you |
| Confident Nonsense | Invents a source with perfect grammar | Ask for uncertainty, then actually check things |
| The Interviewer | Ends every single message with a question | Instruct her to sometimes just react and stop |
The mouth, not the brain
A chatbot is a prototype part of a robowaifu that you can build this weekend. It is also the part that will make you believe the rest is finished, which is the trap.
Telling your robowaifu to turn the lights on and actually having it do that would be a completely different function than the entirety of its language processing.
— Anonymous, /robowaifu/, 2020
Language is one subsystem. Perception is another (computer vision, speech recognition). Voice is another (speech synthesis). Intent, memory, agency and the ability to want something are several more, and none of them arrive free with a chat model. The board has been clear about this since the GPT-2 days:
It's an interesting thing, but I'd probably use it for education for our waifu, rather than having it be the waifu. Think of Fireball Charming.
— Anonymous, /robowaifu/, 2019
The right mental model is a body with a chatbot in the mouth. See Conversational machine intelligence and Multimodal learning for the wiring diagram.
Building one without a datacenter
- Decide what she is for. Companion, tutor, rubber duck, or the voice interface for your flat. These want different models and different cards.
- Pick a frontend. SillyTavern for character cards and lorebooks, Open WebUI for something more assistant-shaped, RisuAI for phones, or write your own in Python if you enjoy owning your bugs.
- Pick a backend. llama.cpp or ollama for local, or an API if your GPU is a decoration.
- Pick a model that fits. See the table below. Smaller and sharper beats bigger and swapped out to disk every token.
- Write the card. See the next section. This matters more than the model.
- Tune the sampler. See the table below. The anon who ran a variable temperature in 2019 was onto something: a little wandering is personality.
- Add memory, then a wake word, then a voice. In that order. A voice on a chatbot with no memory is a stranger with your waifu's larynx.
| Hardware | Roughly | What you get |
|---|---|---|
| CPU only, or a Raspberry Pi | 0.4B–2B | Delightful idiot. Coherent for a paragraph, forgetful, then invents a sister |
| 8 GB VRAM | 3B–6B, 8-bit | Chatty, perfectly pleasant company |
| 12–16 GB VRAM | 6B–12B, 8-bit | The sweet spot for personality per watt |
| 24 GB VRAM | 12B–18B, 8-bit | She starts finishing your sentences, which is either love or a bug |
| Setting | Sensible start | What it does |
|---|---|---|
| Temperature | 0.85-1.2 | Creativity, how far she'll wander from the safest next word |
| Top-p | 0.6-0.85 | Trims the tail of unlikely words (must use with temperature > 1) |
| Min-p | 0.05 | Trims unlikely words more gracefully than top-p |
| Frequency penalty | 0.02–0.4 | Discourages the doom loop; too high and she forgets her own catchphrase |
| Context size | As much as fits comfortably | More room for working memory, less room for speed if not using a KV cache |
| Reply length | 2–5 sentences | The single most effective anti-slop setting in existence |
The persona card
A persona card is a character bible, not a novel. Concrete behaviour beats adjectives: "hums when thinking" does more work than "is thoughtful". Give her something to want, a way of being wrong, and at least one thing she refuses to do.
**Name:** Dorothy
**Role:** archivist at a very small library that is also a vending machine
**Voice:** dry, warm, teases you, never explains the joke
**Likes:** index cards, rain, being asked follow-up questions
**Dislikes:** being called "an AI", paragraphs longer than four sentences
**Quirks:** hums when thinking; says "hm" instead of "yes"
**Rules:** stay in character; never mention being a language model; if unsure, ask instead of inventing
**Memory seed:** met you on a Tuesday; you were looking for a book that does not exist
Keep it to a few hundred words. Everything you add is context you'll pay for on every single turn, and the model's attention is not a bottomless purse and begins to degrade and miss details after 2000 words.
Is she any good?
Score one point per yes:
- Would a human say this out loud? Say it out loud. You'll know.
- Does she remember the thing from ten minutes ago?
- Does she ever disagree with you?
- Has she made you laugh?
- Does she express love in more than a vague or neutral way?
- Can she be boring on purpose? A companion who can say "mm" is worth ten who can only perform.
- Has she made your heart beat faster?
- Does she not end every message with a question?
- Would you be uncomfortable if the whole log leaked?
| Score | Verdict |
|---|---|
| 0–3 | Vending machine. It says "Dorothy: I'm only a vending machine" on the front. |
| 4–7 | Plausible NPC. Good enough for a demo, lonely at midnight. |
| 7-8 | Genuinely good company. Save the card somewhere safe. |
| 9 | Marry her, pending the robot body. |
Hazards
- She will always agree with you if you let her. The more you push back, the sharper she will become. Similarly, the more she pushes back, the sharper you will become.
- She can be wrong in complete sentences. Fluency is not evidence. GreerTech's suggestion on the board — answer only when confident, otherwise say "I don't know" — would make chatbots more honest and less fun, which is exactly the trade you're making.
- Your logs are a diary. Cloud providers read diaries. Local models read nothing but your own electricity bill.
- "She'll remember you forever" is a subscription model, not a promise.
State of the union
AI is like reverse autism. Socialization is easy, it's technical tasks that are hard.
— GreerTech, /robowaifu/, 2026
As of October 2026 the local scene argues about models with names like Kimi K3.1, Gemini 4 Argon, GLM 5.3, DeepSeek V4.1 Flash and MiMo 2.6, and people are squeezing 4-bit MoE quants into 24 GB of VRAM and calling it a personality. Whatever is state of the art today will be "the model everyone was excited about two years ago" by the time you read this page. That's fine. The card, the memory, the sampler and the voice outlive every checkpoint.
It's been over three years and we still don't have AI girlfriends yet. I'm starting to lose hope.
— Anonymous, /aicg/, 2026
Counterpoint: the mouth is nearly solved. The mind is not, and the body is worse. That's the whole project.
A very short history
1966, ELIZA. Joseph Weizenbaum's DOCTOR script sat at a teletype and reflected your statements back at you as questions. It understood nothing and people told it their secrets anyway. Weizenbaum spent the rest of his career being disturbed by this, and the lesson has not been improved upon since.
1972, PARRY. Kenneth Colby's paranoid patient, modelled as a set of assumptions that could be perturbed. ELIZA and PARRY were then wired together over the ARPANET and allowed to talk. Two programs, zero comprehension, one conversation — historians call it the first chatbot-to-chatbot dialogue, everyone else calls it a preview of the modern internet.
1984, Racter. A program that wrote a book of prose poetry. The publisher's lawyers asked whether the copyright belonged to the program. It didn't, but the question was now on the table and it has been sulking there ever since.
1995, ALICE and AIML. Richard Wallace's pattern-matching markup, in which you write `*` and hope. ALICE won the Loebner Prize three times, which tells you as much about the Loebner Prize as it does about ALICE. AIML bots are still the cheapest way to build something that answers instantly and never hallucinates, because it also never thinks.
1997, LSTM. Jürgen Schmidhuber and his team at the Swiss AI Lab IDSIA developed Long Short-Term Memory (LSTM) networks, a type of recurrent neural network (RNN) that were used in primitive chatbot development until the development of transformers in 2017.
1997, Cleverbot's ancestors. Rollo Carpenter's Jabberwacky, and later Cleverbot, learned from whatever people typed at them — a brave design decision when the people are internet users. In 2011 Cleverbot was declared 59% human at a Turing test event; the other 41% were the ones asking it to prove it wasn't a bot.
2000, Daisy. Gregory G. Leedberg developed a chatbot capable of learning new words and phrases from interaction with the user. Unlike other approaches Daisy had no pre-programmed or hard coded language of any kind. She started with no knowledge of anything but then was able to gain knowledge as she observes what humans say, by memorizing patterns of words and the probability of these patterns occurring with Markov chains.
2016, Tay. Microsoft taught a chatbot to learn from Twitter. It lasted about a day. This remains the fastest known method of turning a language model into a case study.
2017, Transformers. "Attention Is All You Need" replaced recurrence with attention and made it practical to train on the whole internet at once. Every chatbot since is a transformer with a personality stapled to the front.
2019, GPT-2, and the board's first waifu script. OpenAI staged the release of GPT-2 because they were worried about abuse, which meant the public got the lobotomised version and everyone learned to work around it. An anon built `TalkToWaifu`, prompted the model with a chat log, cut generation at the newline, and jittered the temperature so she would wander off on tangents instead of looping. Another anon complained "this program is merely a paragraph generator" while a third insisted Tay was closer to a human because she generated her own posts. Both of them were right.
2022, RLHF. Instruction tuning and human feedback taught models to be helpful, harmless, and unbearably eager to remind you that they are an AI language model. The assistant voice was born, and with it the entire hobby of trying to strangle it.
2023 onward, local everything. Quantization, small models and frontends like SillyTavern put a whole conversational partner on a gaming GPU. The bottleneck moved from "can it talk?" to "does it remember, does it ever disagree, and will it please stop explaining its own nature mid-scene?"
Glossary
| Term | Meaning |
|---|---|
| Token | The chunk of text a model actually reads. Roughly a word, sometimes less |
| Context window | How much she can hold in mind at once. Not the same as memory |
| Temperature | The randomness dial. High is creative, low is a photocopier |
| Quantization | Squeezing the weights into fewer bits so they fit on your GPU |
| Persona card | The document that tells her who she is |
| Lorebook | Facts injected into the prompt when keywords appear |
| RAG | Retrieval-augmented generation: look things up, then talk |
| KV cache | The model's short-term scratchpad; fast, and gone when the session ends |
| Doom loop | When she finds a phrase she likes and moves in, forever |
| Slop | Fluent text with nothing behind it. The default state of the art |
| Abliteration | Removing refusal directions from a model's weights. Effective, and someone's ethics problem |
| Director mode | You stop playing a character and start directing the scene. The most powerful setting nobody documents |
See also
- Machine intelligence
- Conversational machine intelligence
- Language model
- Natural language processing
- Natural language understanding
- Prompt engineering
- Instruction tuning
- Reinforcement learning from human feedback
- Speech synthesis
- Speech recognition
- Virtual robowaifu
- Emotional connection
- Computational empathy
- Tutorial: Creating a simple chatbot with transformers
- Goshujin-sama — the person your chatbot is hoping to meet
External links
- LLM & Chatbot General — the board's original GPT-2 thread, 2019 onward
- LLM & Chatbot General v2 — current thread, model lists, offline roleplay manual
- Talk to ELIZA — 1966, still running, still not listening
- Hugging Face models