Turn your computer
into an AI hotspot.

Each marker is a tool used the way people use it in 2026, placed at the model work one task typically gets. The gray dot is what that task ends in: one answer, one draft, one app, one result. ChatGPT Work takes one request to one finished deliverable (OpenAI launch). A chat averages 2 turns in LMSYS-Chat-1M, a public chat service like LM Studio’s. Lovable apps take 5 to 30 prompts in community prompting guides. A blog post in Claude chat or Cowork takes a median of 13 rounds (Anthropic Economic Index). Agent tasks average about 20 tool calls on OpenClaw and Hermes Agent in WildClawBench; OpenWork, not measured there, runs the same one-path agent loop. Why those single answers look alike: different models’ answers to the same open question averaged 71% to 82% similarity in Artificial Hivemind (NeurIPS 2025), and a 2026 study of 6,875 essays found AI assistance cut the variance of essay structure by 70 to 78%, concluding that homogenization is a function of interaction design (arXiv 2603.21228). Exploration row: Midjourney returns four images per prompt (Midjourney); Codex cloud runs best-of-N up to 4 attempts (OpenAI, limit); Cursor 2.0 runs up to 8 agents on one problem (SD Times); the thread page count is modeled from a 32K to 54K context at about 4K tokens per page; Claude Pro’s roughly 45 messages per 5 hours is Anthropic’s older published floor. Coral dots are estimates, not measurements: distinct directions when one tool is asked for N versions, from one model’s sampling curve in NoveltyBench (GPT-4o: 2.88 distinct answers in 10), extended as d(k) = 1 + 0.82 ln k, and at half rate past about 10 for a single thread. Product badges are stand-ins, not official logos.