Labs
Ideas you can operate
One idea from the curriculum per lab, with the arithmetic actually running. Change an input and the numbers change, because they were computed rather than drawn.
- Labs
- 8
- Open now
- 8
- Ideas made operable
- 24
- Network calls
- 0
Every lab runs entirely in your browser. No account, no API key, nothing sent anywhere - they keep working with the network off.
Open now
8 labs, finished and honest about what they simulate.
- Inference EngineeringOpen
Byte-pair tokenizer
Train BPE on your own text and watch the merges formPaste any text and run real byte-pair encoding over it, one merge at a time. The merge table, the resulting tokens and the characters-per-token ratio all update as you turn the dial.Subword vocabulariesWhy token counts are not word countsMerge orderRun it - Inference EngineeringOpen
Attention heatmap
Type a sentence, see which words attend to whichA full single-head self-attention pass computed in the page: scores, the scale factor, the causal mask and the softmax, drawn as a heatmap you can read row by row.Scaled dot-product attentionCausal maskingSoftmax temperatureRun it - DeploymentOpen
Consistent hashing ring
Remove a node and measure exactly how many keys movePlace servers on a hash ring, give each one virtual nodes, then drop one and see the real percentage of keys that had to move - next to the number a plain modulo would have given you.Virtual nodesRebalance costLoad skewRun it - DeploymentOpen
Latency budget
Assemble a request path from real hardware numbersBuild a request out of cache hits, disk reads, network hops and cross-region calls, and see where the budget actually goes - serial versus parallel, p50 versus tail.Latency numbersSerial vs parallel costTail amplificationRun it - DeploymentOpen
Cache eviction simulator
Run the same access trace through LRU, LFU and FIFOOne workload, three eviction policies, one hit-rate chart. Includes the scan pattern that makes LRU look foolish and the skewed one that makes LFU look brilliant.Eviction policiesHit rate vs cache sizeScan resistanceRun it - Everything DataOpen
Bloom filter sizing
Pick bits per element and watch the false positives arriveInsert keys, query keys you never inserted, and compare the measured false-positive rate against the one the formula predicted for your chosen bit count.Bits per elementHash countFalse-positive rateRun it - Inference EngineeringOpen
Gradient descent on a surface
Change the learning rate and watch the run divergeDrop a starting point on a loss surface and step downhill with plain SGD, momentum, or Adam, with the trajectory drawn over the contours.Learning rateMomentumWhy Adam adapts per parameterRun it - Inference EngineeringOpen
Quantisation memory budget
Work out whether that model fits in your GPUParameter count, precision, KV cache and context length, turned into the number of gigabytes you actually need - and the precision you would have to drop to.Precision formatsKV cache growthVRAM budgetingRun it