Approach

How we work

  1. Choose a task. Tell me what you want done and how you'll know the result is good enough. If you're not sure, we'll work that out together on the call.
  2. Work through it together. Online, on your own computer, for about 100 minutes. You know the work. I help you pick the AI tools and guide each step.
  3. Check and continue. We check the result against what you need. If it helps, we compare it with how you do the task now. Faster only counts if the result is just as good. Then I write a short record of what we did, where your work is saved and how to do it again. A 20-minute follow-up about a week later is included.

What the AI is doing

You don't need to know this to use AI, and you can skip it. It shows one useful thing: AI writes by picking a likely next word. Likely isn't always right, so every session includes a check.

A toy neural network guessing the next word Three words, cat, sat and on the, feed 3 middle nodes through weighted links. 2 of the 3 middle nodes pass their totals on as pulses; the rest stay quiet. Middle totals 1.6, 0, 1.2. The outputs become odds for the next word: mat 77%, floor 22%, moon 1%. 1 Words in2 Links3 Add up4 Next-word odds 1cat 1sat 1on the 1.6 0 1.2 mat77% floor22% moon1% A toy neural network guessing the next word Three words, cat, sat and on the, feed 3 middle nodes through weighted links. 2 of the 3 middle nodes pass their totals on as pulses; the rest stay quiet. Middle totals 1.6, 0, 1.2. The outputs become odds for the next word: mat 77%, floor 22%, moon 1%. 1 Words in2 Links3 Add up4 Next-word odds 1cat 1sat 1on the 1.6 0 1.2 mat77% floor22% moon1%
How a neural network guesses the next word

A toy network finishing “The cat sat on the …”, read from 1 to 4.

  • thicker link, stronger weight
  • dashed or hollow: pushes down
  • bigger pulse, more signal
  1. Words in. Each word becomes a number; here, 1.
  2. Links. Each link multiplies what it carries by its weight, set in training.
  3. Add up. Each middle node adds what arrives. Above zero, it passes the total on (filled); otherwise it stays quiet (empty).
  4. Next-word odds. The totals become odds that add to 100%. The model picks from them: usually “mat”, not always.

The sums here are real: these odds come from these links. The weights were set earlier, in training, and don't change while you use the model. Real models have billions of them.

Project help

For someone directing a project built with AI. The drawing shows one way to keep that work checkable: one AI builds, a different AI reviews, and you decide what is accepted. I don't write or review the code. I help you set the brief and the check. I don't offer project help as a package yet.

You give the brief. One AI builds and sends a draft to a different AI, which sends notes back; they go back and forth. Then you check and decide. Only accepted work goes out; anything else goes back to be revised. revise 1 Your brief 2 AI builds draft notes 3 A different AI reviews draft, notes, questions 4 You decide Acceptedwork You give the brief. One AI builds and sends a draft to a different AI, which sends notes back; they go back and forth. Then you check and decide. Only accepted work goes out; anything else goes back to be revised. revise 1 Your brief 2 AI builds draft notes 3 A different AI reviews draft, notes, questions 4 You decide Acceptedwork
Build, review, decide

One AI does the work and a different AI checks it. They pass the draft and notes back and forth. Then you decide what to accept. Nothing goes out until you say so.

Why two different AIs? Research finds an AI struggles to fix its own mistakes without outside feedback. It also tends to rate its own work above others' (Huang et al., ICLR 2024; Panickssery et al., 2024). So a different AI does the review. It has no draft of its own to defend.

Contact

Book a discovery call to discuss your task or project.

Probably not a fit if you want an app built or fixed for you, legal, tax, financial, medical or security advice, AI acting on your accounts without you checking, or a promise of time or money saved.

Book a discovery call Free 30-minute call, by Google Meet · no payment needed · booking via Calendly

Prefer email? Write to andrew@quietforces.com with a few words about your task and some times that suit you. I'll reply to arrange a Google Meet call, or we can keep the conversation to email. Please don't include passwords or confidential material.