Playground

Move the dials — chunk size, top-k, reranking, hybrid search — and watch retrieval quality and latency shift.

Shard a model across GPUs with tensor, pipeline, and data parallelism plus ZeRO — and see whether it fits in memory.

Fit one model to many random samples and watch the fits scatter (variance) and drift from truth (bias) as complexity, data, and noise change.

Play four arms with greedy, ε-greedy, and UCB on the same reward draws — and watch which strategies keep cumulative regret low.

Score a group of responses and watch GRPO turn rewards into advantages using the group's own mean as the baseline — no critic, and it collapses when rewards agree.

Drag the importance ratio and flip the advantage's sign to see how PPO's clip caps each policy update inside a trust region — and where the gradient dies.

Pick a query token and watch softmax(QKᵀ/√d_k) decide who it attends to — slide the temperature to see the scaling that keeps gradients alive, or mask the future for autoregression.

Grow the sequence length and watch full attention’s cost explode quadratically — then switch to windowed or sparse patterns and see it drop back toward linear.

Slide the query and key positions and watch their vectors rotate — the attention score depends only on the relative gap between them, not the absolute positions.

Size the KV cache for a real model and watch it outgrow the weights — then compare how MHA, GQA, MQA, and MLA shrink it, and quantize the cache itself.

Route a batch of tokens to experts and watch the router collapse onto a few — starving the rest and dropping tokens — until load-balancing pressure spreads them out.

Walk one Model Context Protocol session — connect, negotiate, discover tools, call one, feed the result back — and see how a poisoned tool description or an over-broad permission changes the outcome.