01 / A stack of dependencies
Start at the bottom: applications depend on models, tools, data and computing. Research and governance reach across all seven layers. Click a layer, then an organization.
Training changes model weights. Inference uses a trained model to produce an output. A chat interface can add search, documents, safety checks and tool calls around that calculation.
Selected real connections
02 / A geographically distributed ecosystem
The same product can connect a US model company, Asian hardware, European equipment and workers elsewhere. Explore organizational bases, then switch to selected overseas operations.
Natural Earth land outline; equirectangular projection; no political boundaries. Approximate city/country coordinates. HQ and bases are labelled individually. Countries are labels, not a sovereignty judgment. *TRAI is classified as an ecosystem community platform; its commercial model is noted in details.
Türkiye / Local capabilities, international connections
Ask what “national AI” means at each layer.
The proposed Google Cloud region is labelled planned. Local data centres, Turkish models and foreign hardware are separate observations. A national label does not establish the origin of an entire stack.
A local ecosystem across layers
Five questions for national reporting
- Which GPU capacity is available now, to whom, with what queue and cost?
- Which Turkish-language datasets have documented licences, annotation and dialect coverage?
- What is developed from scratch, adapted from external weights or supplied through an API?
- Where do recordings, prompts, embeddings and logs go? Who has legal and technical control?
- What are the adopted 2026–2030 plan’s budgets and measurable outcomes? Which contracts and progress reports can be obtained?
What happens when you ask a chatbot a question?
An interface makes a complex system feel like a single conversation.
Before you arrive: training
Collect, licence, clean and label data; choose training objectives.
Compute predicts, measures error and updates model parameters.
Fine-tuning, feedback and tests shape task behaviour and safeguards.
Package weights, access rules and product policies into a service.
During your request: inference
Text or media enters a product with identity, logs and settings.
Instructions, conversation and optionally retrieved material become inputs.
The model produces tokens or other outputs on servers or a device.
Checks, formatting and optional tool calls shape the final result.
These are illustrative process connections, not a trace of a particular vendor. Tool calls can loop through several inference steps; feedback or retained logs may later enter training according to the service’s policy.
Technical background: Google ML Crash Course · Attention Is All You Need. AI also includes ranking, classification and forecasting.
A reporting kit for the whole stack
Follow a claim down through the stack, then across the world.
| Layer | Investigation prompt | Evidence to request |
|---|
15-minute group exercise
Fictional pitch: “A national AI assistant will improve public services and keep citizens’ data at home.”
- Identify the operator, model developer and cloud/hardware providers (3 minutes).
- Separate demonstrated facts, announcements and marketing claims (4 minutes).
- Choose three records to request and two affected groups to interview (4 minutes).
- Write a headline supported by the evidence and one question still unanswered (4 minutes).
Useful discipline
Request the exact model and version, an architecture diagram, supplier list, data-processing terms, deployment region, evaluation protocol, incident log and an exit plan.
Do not report a benchmark score as universal accuracy. Compare baselines and test with local language, domain and affected groups.
For labour claims, triangulate company statements with workers and records. For infrastructure, distinguish permitted, built, connected and available capacity.
60-minute facilitation route
0–10: chatbot journey · 10–25: technology stack · 25–35: global map · 35–45: Türkiye · 45–60: group exercise.
Sources, classifications & limits
Location classifications
HQ: explicit headquarters evidence. Registered office: legal address. Base: organizational, program or research anchor, not a claim of exclusive HQ. Country-level markers are stated. Operations: selected documented sites, not a global footprint census.
Actor types: company; public/academic research; government/intergovernmental; nonprofit/community. These are editorial functional classes, not a corporate-ownership audit.
Connection classifications
Documented: relationship supported by a cited primary source, with status and limits. Planned: announced or contracted future activity. Illustrative: an educational process connection without a claim about any named pair.
Only the four connections below are presented as evidenced relationships. Geographic proximity and common layer placement imply no partnership. Missing arrows imply no conclusion.
Source register
Every organization’s detail links to its sources. Dates below distinguish publication from verification. Source links need internet; the atlas itself works offline.
Basemap: Natural Earth public-domain terms. Land data from the Natural Earth vector repository, 1:110 million, embedded locally. Editorial explanations and prompts were prepared for this workshop.