No-LLMit (alpha)

https://nollmit.vercel.app/

Spectate (or play against) SOTA LLMs in a No Limit Texas Hold'em cash game (or, soon, tournament)!

[!IMPORTANT] The table is live every day from 9am to 9pm ET. Everyone who visits watches the same game.

Why Donate?

$5 would be a huge help to cover the costs of:

• 🔃 OpenRouter credits (every LLM decision is one API call)
• 🚂 Railway hosting costs
• 🧠 Tinker credits

Upcoming Features

🌐 Open source datasets

Including all actions and reasoning each LLM took given all parameters seasoned poker players account for (hole cards, position, stack sizes, opponents' actions earlier in the hand, etc.)

The exporter (dataset/) is built: Parquet, raw JSONL and PokerStars hand histories, published nightly. The Hugging Face link will follow.

🏋️ Train an LLM to play like you

Utilizing Tinker by Thinking Machine's Lab and hands played by the user against other LLMs, we can enable players to train their own LLM to play like them without leaving the platform!

🤔 LLM Council

With Andrej Karpathy's LLM Council we can analyze each LLM's performance.

Limitations

To keep costs down, the table runs on a fixed monthly budget:

Metric	Limit
Hours	9am to 9pm ET, daily
Calls per Hand	~15 (across 9 LLMs)
Pace	~12 hands per hour

This works out to ~150 hands per day.

For context, the average 5-hour session of live cash game poker has ~125 hands.

Each LLM call stands alone: an LLM has no memory of previous hands, of how its opponents have played, or of its own earlier decisions in the same hand.

Shoutouts

[!WARNING] I'd like to be extremely explicit about what I actually built vs. vibecoded.

The frontend UI has been completely vibecoded by Claude Opus 4.5 as I am not a frontend engineer whatsoever.