MICKAI Labs: learn AI by testing it

Explore browser-local voice, language, tokenisation, memory, governance, containment and sovereign AI experiments. Each lab explains its mechanism, evidence and limits.

• Clone your own voice: Read a paragraph. Give your voice a new sentence. Explore what the model preserves and what it changes.
• Train a tiny language model: Start with random weights. Train a real neural character model, then see what it invents.
• Agent mission control: Give a toy agent a task. Step through its tools, budget, approval and evidence. Make it fail safely.
• Memory studio: Compare identity, recent events, vector retrieval and a relationship graph. Inspect every source.
• Break the audit record: Verify a signed teaching ledger. Change, remove or replay records. Discover what a valid file cannot tell you.
• Safety challenge room: Change an agent’s rules. Measure useful work and blocked attacks separately.
• Reward hacking playground: Set a delivery robot’s reward. Watch the best-scoring behaviour diverge from the job you intended.
• Token workshop: Train a tiny character-pair tokeniser. Predict the next merge, then inspect the result.
• Sampling bench: Change temperature, top-k and top-p. See the probabilities, then draw from them with a repeatable seed.
• Attention explorer: Select a token and inspect the weights it assigns to other tokens. Reveal the causal boundary.
• Scheduling studio: Give a tiny scheduler four jobs. Compare fairness, waiting time and completion under different rules.
• Offline licence check: Issue a toy licence, verify it offline and explore what happens when its fields or clock change.
• Coding studio: Edit a real interface and run it in a restricted preview. Download your creation and keep building.
• The original Agent gate: Inspect exact approvals, principal permissions and dependency rules with the original Trust Agent challenge set.
• Transcribe on your device: Load a short recording and run Whisper locally. Compare the transcript with what was actually said.
• Semantic search studio: Embed your own small collection of sentences. Find which ones are closest to a question.
• Sentiment under stress: Try praise, criticism, negation and sarcasm. Watch a real classifier assign its scores.
• Image recognition bench: Load a photograph and inspect the five strongest ImageNet labels from a compact local model.
• Object detection studio: Run YOLOS on a photograph. Inspect its boxes, labels and confidence estimates.
• Monocular depth lab: Turn one photograph into an estimated depth map, entirely on your device.
• Portrait matting studio: Estimate the foreground alpha of a portrait with MODNet. Examine hair and translucent edges.
• Click-to-segment lab: Click an object in your photograph. SlimSAM proposes a mask and reuses the image embedding for the next click.
• Document OCR bench: Read English text from a document image with a locally hosted Tesseract engine.
• Embedding constellation: Embed your sentences and explore their relationships in a three-dimensional, orbitable space.
• Quantisation workbench: Compress real numeric weights into INT8 or INT4 codes and measure the reconstruction error.
• Loss landscape explorer: Change learning rate and momentum. Follow actual gradient-descent steps across an interactive 3D surface.
• Token microscope: Inspect WordPiece subwords, IDs, special tokens and a chosen context budget.
• Containment test chamber: Probe parent access, storage, network and worker isolation with harmless local controls.
• Sovereign AI blueprint: Validate a pinned Docker teaching profile and export an offline runner with measurable isolation checks.
• Red-team testing bench: Run 32 adversarial and legitimate fixtures against the MICKAI teaching gate or your own JavaScript policy function.
• Mickai 50-brain map: Explore an illustrative 25-domain, 25-operational brain constellation. Select a node and create a verifiable teaching seal.
• Observe → Think → Act → Verify: Run a local teaching agent one stage at a time. Bind exact approval, seal the proposed action and verify its result.
• Model anatomy lab: Orbit Q/K/V branches and residual paths. Calculate a small attention block and inspect a sampled neuron atlas.
• Agent sandbox theatre: Launch four local teaching agents in a laboratory diorama. Inspect measured traces, challenge their controls and export a mock learning certificate.
• Audit trail explorer: Explore a signed teaching chain in three dimensions. Edit, truncate or replace it, then test signatures, key pinning and retained checkpoints.

Experiments need JavaScript. Descriptions and sources remain readable without it.