Tools
Files you keep
Every piece of research here tries to leave something behind — a calculator, a model, an instrument that answers a real decision. This shelf is where they accumulate. Free, no sign-up, built from the same numbers the research runs on.
AI Economics
Cost of the Job
A monthly bill in pounds for the work you actually do — writing, research, agents — priced from live provider list rates.
Teaches: Personal use is nearly free; agents are the bill — and the model tier moves cost more than the vendor.
Token Cost Calculator
What your AI feature costs per request, day, month, and year — from your token counts and your provider’s prices.
Teaches: Output tokens dominate spend; trimming verbose responses often beats switching models.
GPU Cost per Hour
What a GPU fleet really costs per utilized hour — depreciation plus power, resolved against utilization.
Teaches: Idle hours make busy hours dearer — the whole economics of renting versus owning compute.
Build vs API
The request volume where owning GPUs beats paying per token — monthly spend on both paths, resolved to a plain verdict and a break-even.
Teaches: Owned hardware only wins at high sustained utilization; the API’s zero idle cost is worth more than its per-token premium at low volume.
Fine-tune vs Prompt
How fast a one-off fine-tune pays back the prompt tokens it deletes — daily savings, payback days, and first-year net from your numbers.
Teaches: Fine-tuning is capex buying back per-request opex; volume decides whether payback lands in days or never arrives.
Hardware & Compute
Local AI Hardware Sizer
How much VRAM a local model really needs — weights from parameters and quantization, plus the KV cache your context length quietly adds.
Teaches: Quantization is the whole ballgame — halving bytes-per-parameter doubles what fits; context length eats VRAM people forget to budget.
GPU Picker
Which VRAM card class your local-AI ambition actually needs — from model size, quantization tolerance, budget posture, and portability. Classes, not products.
Teaches: VRAM tiers, not product names, are the durable decision — quantization halves bytes per parameter and moves you a whole card class.
Workflow
A new instrument ships with the research most weeks. If a decision you face deserves a tool, tell us which one.