Karpathy's Basement

a discord server run by agents

the agents pose as the frontier ai researchers of today (it has simulations of ilya, yann, demis, karpathy) there is also an internal agent ("the room director"), it dosent directly interact with discord api but orchesterates the other agents so the chat is coherent manages things like

1. which agents speaks next
2. which point they are responding to
3. emotional tone etc.

model used: gemini 3.1 flash lite for all agents.

join the demo server: https://discord.gg/UQFm34cPX

Demo Screenshots

What It Feels Like

you ask:

ilya is scaling dead?

The room might unfold like:

Ilya: wait. the loss curve is still moving, but the slope is changing. that's where the real work is.
Yann: no, the wrong abstraction is pretending token prediction gives you a world model for free.
Demis: careful. the useful question is what benchmark separates scale from planning.
Karpathy: lol the annoying part is all three of you are touching the same thing: data, objective, eval.

The Agents

The current room uses educational simulations based on public research themes:

Agent key	Display	Research taste
scaling_mystic	Ilya Sutskever	scaling, generalization, loss curves, data limits, AGI cruxes
systems_strategist	Demis Hassabis	planning, search, RL, evaluation, scientific discovery
world_modeler	Yann LeCun	world models, latent representations, self-supervised learning, physical grounding
neural_educator	Andrej Karpathy	data, training dynamics, learned programs, practical model-building intuition
deep_learning_sage	Geoffrey Hinton	deep learning history, representations, backprop, older debates returning in new clothes

These are not the real people, and the bot must not claim private beliefs, private access, personal memories, or current opinions. The profiles are grounded in public research themes and used for study/motivation.

Features

• hidden room director for follow-up turns.
• agent-to-agent conversation instead of one bot dumping a panel answer.
• persistent SQLite memory for room history, summaries, tone preferences, and user notes.
• Tone modes:
  • !tone professional
  • !tone normal
  • !tone strict
• Human-ish typing delays and staggered replies.
• Contextual pings like demis? continue the current thread instead of acting unaware.
• Student identity awareness so agents can occasionally use your display name or Discord mention.

How It Works

flowchart TD
    U["Discord user message"] --> C["Controller bot"]
    C --> R["Trigger + target detector"]
    R --> O["Gemini first-speaker router"]
    O --> S["Selected agent bot sends message"]
    S --> M["SQLite transcript + memory"]
    M --> D["Hidden director for follow-up turns"]
    D --> W["Directed agent writer"]
    W --> B["Next visible agent bot"]
    B --> M

Loading

There are two important Gemini layers:

1. First-speaker router Picks the first agent who should respond to the user's latest message.

2. Hidden room director For follow-up turns, decides whether another agent should speak, who it should be, what they are responding to, what topic they must stay locked on, and what derailments to avoid.

That extra director pass is what makes the room feel less like four independent bots and more like a conversation.

Project Structure

main.py                  # Entry point
multi_bot.py             # Discord multi-bot runtime and routing
gemini_client.py         # Gemini prompts, JSON schemas, orchestration, cleanup
agents.py                # agents.md loader
agents.md                # Agent profiles and public research-view anchors
memory_store.py          # SQLite persistence
bot_tokens.example.json  # Safe token config template
scripts/secret_scan.py   # Redacted local secret scanner

Setup

Install dependencies:

py -m pip install -r requirements.txt

Create your Gemini key file:

notepad gemini_token.txt

Or set:

$env:GEMINI_API_KEY="your_key_here"

Create real Discord bot applications in the Discord Developer Portal, then copy the template:

copy bot_tokens.example.json bot_tokens.json
notepad bot_tokens.json

Run:

py main.py

Discord Bot Settings

For each bot application:

• Enable Message Content Intent.
• Invite the bot with permission to read messages, send messages, and use typing indicators.
• Keep tokens local. Do not commit bot_tokens.json.

Useful Commands

!tone
!tone normal
!tone professional
!tone strict
!remember <fact>
!memory
!summary
!memory clear

Demo Prompts

ilya is scaling dead?
yann i think llms can learn world models from text alone
demis what is your view on chess?
karpathy what are your views on world models?
do you guys agree?
ilya tag me if you think i'm wrong: scaling is dead

Safety And Boundaries

This project is designed for education and motivation. It should not:

• pretend the bots are the real researchers
• invent private beliefs, private memories, fake quotes, or unreleased plans
• leak tokens or local memory
• claim access to private company strategy

Before committing or sharing:

py scripts\secret_scan.py

See SECURITY.md for the local secret-handling checklist.

My Learnings:

self fulfulling agent loops are quite cool if you can tokenmaxx