Under the Hood of Generative AI (LLMs)
7 demos
-
Embedding SpaceVisualize word embeddings in 2D — explore clusters, nearest neighbors, and vector arithmetic like king − man + woman = queen.
-
Single NeuronInteractive single-neuron visualization — adjust inputs, weights, and bias, then see how activation functions shape the output.
-
Attention MechanismExplore how the attention mechanism works through an intuitive cocktail-party analogy.
-
Base Model vs Instruct ModelCompare how a raw base model and an instruction-tuned model respond differently to the same prompts.
-
Illusion of ChatToggle between a polished chat UI and the raw token stream underneath — see how special tokens create the "conversation."
-
The Context Window X-RaySplit-screen X-ray of a chat app: the chat a user sees vs. the payload the model receives — memories, skills, documents, search results, and silent compaction.
-
Interpreting LLMsExplore how researchers peer inside language models — from polysemantic neurons to interpretable features and attribution graphs.
GenAI Development
3 demos
GenAI Business Value Creation
6 demos
-
Meeting Cost CalculatorQuantify the true cost of your meeting culture with hidden cost multipliers, then triage meetings using the Decide / Learn / Bond / Do framework.
-
GenAI Pricing LabStrategic pricing calculator — navigate the ChatGPT Tax, compare three delivery options (API, open-source, local), and generate data-driven pricing tiers.
-
Build vs. Buy TCO12-month total cost comparison for self-hosted open-source vs. vendor API, with custom canvas charts and a pragmatic hybrid option.
-
What Should You Build?6-step wizard: validate a GenAI idea with the Three Pillars, Four Validation Tests, and get a concrete MVP prototype recommendation.
-
Close the DealPitch a GenAI solution to a skeptical CTO through 4 branching dialogue scenes. Build trust, handle objections, and close the deal.
-
Where Does the Money Go?Interactive capex-to-model timeline: see how each year's Big 5 investment splits into inference, training, and research — each arriving 2–4+ years later.
The Human-in-the-Loop Lab
7 items
"Don't worry, we'll just put a human in the loop." These simulations let you feel why that sentence hides so much work: automation bias, alert fatigue, anchoring, and the moral crumple zone, played out across the sectors the EU AI Act calls high-risk. Companion suite to the booklet LLM-Human Interaction Design Patterns for Operations.
-
The Operator's DilemmaFive acts of being the human in the loop of an IT operation: rubber-stamp triage under a timer, an anchoring trap, confidence theater, a 3 AM kill-switch scenario, and designing the seam yourself. Includes a personal debrief.
-
The Credit DeskBanking: six loan applications, an AI credit score, and an ordering experiment. Does seeing the model's number first change your judgment? Two of the scores are deliberately wrong.
-
The Triage WardMedical: work a night-shift alarm queue where one alarm out of eighteen matters, then work the same night after an evidence-based alarm redesign.
-
The WatchlistLaw enforcement: verify facial-recognition watchlist matches against the record under a countdown. Some of the AI's matches are wrong, and one of its rejections is too.
-
The ShortlistHiring: an AI-ranked candidate shortlist with two seeded injustices. Measures your Weight-of-Advice score: how much of the final ranking was actually yours?
-
The Border QueueBorder control: the same risk assessments packaged three ways (raw decimals, calibrated categories, bare verdicts) and what each framing does to your referrals.
-
The DocketJustice: a pre-trial docket with an "advisory only" AI risk score. One case goes wrong three months later, and the inquiry knows exactly whose signature is on the order.
The EU AI Act Lab
2 items
Tools for reasoning about the AI Act the way an engineer should: grounded in the actual text, honest about what is settled and what is still pending. Educational, not legal advice.
-
The NavigatorWhere does your AI system land under the EU AI Act? Walk the real classification method — the Article 3(1) test, both Article 6 routes, the derogation filter, the profiling override — and get your risk tier, obligations with article citations, and post-Omnibus dates.
-
The Paper TrailYou classified your system high-risk — now survive conformity. A five-act provider simulation: the runway decision, the Annex IV evidence pack, the assessment route, a white-label trap, and a post-market incident on a real reporting clock. Scored, with citations.