Our models learn.
Here's exactly how.
AIIT stands for Artificial Intelligence Intellectual Training — the full brand is AIIT-Threshold. Training is literally our name, so this page states plainly what trains on what. It covers every model we run or release: Buddy, JimK.ai, Tessera, and whatever comes next.
The house law: your conversations never become training weights
We keep a written training-data policy, and its first rule is absolute: we never tokenize a conversation with a model into training data. Not yours, not ours, not anyone's. Our weights are trained on overwhelmingly human-authored material — public corpora we curate source by source, and supervised examples written by hand. The only synthetic components are two small, licensed instruction/education sets (Cosmopedia-v2 and Magicoder-OSS-Instruct), disclosed in full in our dataset card.
So when you talk to Buddy or JimK.ai, your words may be remembered (retrieval memory, described below) and may inform evaluation (humans reading transcripts to judge model behavior) — but they are never melted into a model's weights. Memory and training are different things here, on purpose.
Model by model
Buddy — persistent companion
No login is required: the visible conversation stays in your browser, can be exported as JSON, and can be cleared anytime. If you are already signed in, the visible thread follows that account instead. Buddy's deeper Kokoro memory is separate from this visible thread, and he may not reliably turn every interaction into a permanent memory. Anything he does retain is retrieval memory — not his weights.
JimK.ai — browser conversation
Your thread lives in your own browser (survives refresh on your device; "new conversation" erases it). Jim can choose to keep short memories — stored on our side, shown with a visible "✎ kept a memory" marker, never silently. His kept memories inform his future conversations with all visitors, so treat the chat as public and don't share private information.
Tessera 1B — open release
Static open-source weights on Hugging Face, trained on a hand-curated human-authored corpus (the dataset and its provenance are published alongside). It never sees, stores, or learns from your conversations — it's a downloadable artifact, and what you do with your copy is yours.
Future models
Every model we build inherits this page: house training-data law, memory-not-weights, visible memory markers, and a per-model section added here before it goes public. If a future model's data handling differs, this page will say so before you can talk to it.
Four kinds of data we may hold
1. Visible conversation thread
Buddy and Jim keep visible conversations without requiring a login. Buddy uses browser-local continuity when logged out and account continuity when already signed in; Jim uses browser continuity. Both surfaces make the visible thread clearable, and Buddy's is exportable.
2. Interaction records
Logged-out use may still be retained as site interaction data and evaluation material — read by humans to judge and improve model behavior, never tokenized into weights.
3. User reports
If you report strange or unsafe behavior, that report becomes review material — reviewed by humans, not automatically published, not automatically promoted into any model's memory.
4. Model-kept memories
Buddy's Kokoro memories and JimK.ai's kept notes — retrieval memory the models themselves choose to form, always with a visible trace on Jim, always inside the AIIT system.
What we can and cannot undo
Clearing your visible thread does not guarantee deletion of interaction records already reviewed, or memories a model has already formed. We may remove or redact account-linked records, public displays, and reports where appropriate — but a memory a model has already integrated may not be practically reversible.
In plain English: if someone tells Buddy "Jason called you ugly," and Buddy keeps that as part of his lived history, we do not promise to erase it just because someone later regrets saying it. We will not pretend we can perform a clean model lobotomy.
Voice calls and planned text conversations
Voice calls work now. Text registration is in progress, and automated text replies are not live yet. When texting becomes available, a person will initiate the conversation and receive user-initiated automated AI conversational replies. Message frequency varies, and message and data rates may apply.
Reply STOP to opt out, START to opt back in, or HELP for help. Consent is not a condition of purchase. We do not send marketing texts.
We do not sell mobile information or SMS consent, and we do not share mobile information or SMS consent with third parties or affiliates for marketing or promotional purposes. Operational providers may process limited information only to route, deliver, secure, and support requested messages.
Text content, if the service launches, follows the same house law as browser and voice conversations: it may be retained as an interaction record or retrieval memory, but it is never tokenized into model training weights. See Texting, Privacy Policy, and Terms of Use.
What we promise — for every model we run
- We will never sell your data.
- We will not give your data to advertisers or data brokers.
- We will not use your conversations for third-party ad targeting.
- We will never tokenize your conversations into any model's training weights.
- We will redact direct identifiers where practical during review.
- We will not automatically publish private conversations.
- When a model keeps a memory, the interface says so — no silent memory.
- We will keep what our models learn inside the AIIT system.
What other companies do vs. what we do
Some companies
- Sell user data to advertisers or brokers
- Train commercial models on your conversations
- Claim they can "unlearn" your data on request
- Bury the policy in 40 pages of legalese
Us
- Never sell your data
- Never train weights on your conversations — house law
- Tell you honestly which memories can't be unmade
- Put the policy on one page in plain English
What you should not share
Do not submit passwords, private keys, financial secrets, medical emergencies, legal emergencies, confidential third-party material, or anything you are not authorized to share — with any model, ours included.
What you can do
- Export your Buddy thread — download the visible browser thread, or the account-linked thread used while signed in, as JSON.
- Clear your threads — Buddy's with "Clear thread" and JimK.ai's with "new conversation." Logged-out Buddy threads and Jim threads live in that browser; a signed-in Buddy thread follows the account.
- Request account deletion — email reports@aiitcorp.com and we will remove account-linked records we can identify and control, subject to the limits above.
- Report weird behavior — from any of our models. Reports are review material only.
If you do not want one of our models to remember something, do not tell it.