Chapter 11 — Reps
Ten reps to turn the cloud-vs-local fork from a table you read into a judgment you own. They move from naming your own tools, to classifying your own data, to running the same task in both venues and seeing the tradeoff with your own eyes. The keyboard is the gym. Don’t read these — do them.
Ground rules
- Type it yourself. Write your prompts and your reflections in your own words. Your judgment learns what your fingers do, not what your eyes skim.
- Run/try everything. Every prompt, every venue, every measurement. A rep you only imagined is a rep you didn’t do. (If you can’t install a local app, use the free-small-cloud stand-in named in each rep — that path is real work, not a cop-out.)
- Predict before you measure. Before you run a prompt in two venues, write down which one you expect to be better, faster, more private — then compare. The gap between your guess and the result is the learning.
- AI usage note (every rep that uses AI): end with one honest line — what you asked, which venue you used and why, what you verified. The human owns the verdict.
- The privacy rule that never turns off: no real confidential or regulated data goes into any cloud tool in these reps. Use fictional stand-ins. Practicing the habit is the point.
- Keep a
reps.txt. One short written reflection per rep. This is graded thinking, not busywork. A plain doc (Google Docs, Word, Notepad) is fine.
Reps 1–3: Map your venues and your data
Rep 1 — Name your two doors
List every AI tool you currently use (ChatGPT, Claude, Gemini, a phone assistant, a note-taker, anything). For each, write one word: cloud or local. Then count. Most people find all of theirs are cloud.
Reflection: If every tool you own is cloud, which of your real work tasks are you currently unable to safely bring to AI at all? Name two. Those are the tasks local AI is about to unlock.
Rep 2 — Classify five real snippets
Open code/data-sensitivity-guide.txt. Pick five things you actually wrote or handled at work this month (paraphrase them so they’re fictional). Sort each into a tier: Public / Internal / Confidential / Regulated. Then write the venue each is allowed to use.
| Snippet (fictional) | Tier | Allowed venue |
|--------------------------------------|--------------|----------------------|
| "LinkedIn post about our open house" | Public | free or frontier cloud|
| "review of [named employee]" | Confidential | local / vetted tool |
| ... | | |Reflection: Which snippet was hardest to classify, and why? The gray zone between Internal and Confidential is where most real mistakes happen — describe your gray zone.
Rep 3 — The one-line test in the wild
Take the hardest-to-classify snippet from Rep 2 and apply the monitor test: “If this were on tomorrow’s front page with my name attached, who gets hurt?” Write the answer as a specific sentence — an actual person or consequence, not “it’d be bad.”
Reflection: Did naming a specific victim change your tier? The test works because it forces a vague “kind of sensitive” into a concrete harm. Say whether it moved your answer.
Reps 4–6: The napkin math and the two venues
Rep 4 — Will it fit? (no install required)
Using the §11.4 rules (~2 GB/B at FP16, ~0.5 GB/B at 4-bit), compute the memory footprint for a 7B, a 13B, and a 70B model at 4-bit. Then look up how much memory (RAM) your own laptop has. Which of the three fit?
7B at 4-bit ≈ 7 × 0.5 = ~3.5 GB (+ ~15–20% overhead)
13B at 4-bit ≈ ?
70B at 4-bit ≈ ?
My laptop has ___ GB of memory. Fits: ____Reflection: What’s the largest model your laptop could realistically run? Why does the 70B model belong in a data center and not on your desk? Tie it back to the FP16-vs-4-bit compression.
Rep 5 — Set up one local runner (or the stand-in)
If you can install: download GPT4All or LM Studio, grab a small model (a 7–8B, quantized to Q4_K_M), and send it one message. Note how long the download was and how the first reply felt. If you can’t install (no-admin path): open a free-small-cloud model (a “mini/nano/flash” tier, or free DeepSeek/Qwen) in your browser. It will stand in for local for the capability comparisons — just remember it’s still cloud, so the privacy benefit is simulated, not real.
Reflection: What surprised you — the setup, the speed, or the quality? If you used the stand-in, state clearly which lessons this week you’ll have to reason about rather than prove (hint: the privacy ones).
Rep 6 — Same prompt, two venues
Pick one public task (e.g., “summarize this published article in five bullets” or “draft a friendly reminder email”). Run the exact same prompt on (a) a frontier cloud model and (b) your local model / free-small stand-in. Predict the winner first. Save both outputs.
Reflection: Was the frontier output actually better, or just longer? Name one specific thing the bigger model caught that the smaller one missed — or admit it was a tie. For this task, which venue would you really use, and why?
Reps 7–9: Route it, time it, unplug it
Rep 7 — Time the difference
Re-run Rep 6’s prompt and roughly time each venue (a phone stopwatch is fine). Record seconds-to-first-word and seconds-to-finish for both.
| Venue | To first word | To finish | Felt like... |
|----------------|---------------|-----------|--------------|
| Frontier cloud | | | |
| Local / stand-in| | | |Reflection: Was the speed gap big enough to matter for how you’d actually work? Speed is one axis of three — would you trade this much speed for the privacy local gives you on a sensitive task? Say when yes and when no.
Rep 8 — Route a sensitive task correctly
Invent a fictional confidential task — e.g., “rewrite this blunt review of Jordan P. into kind, specific feedback” (make up the details; keep them fake). Decide its venue using the guide, then actually run it in the venue you chose. If that venue is local and you can’t install, run it in a vetted/enterprise stand-in you describe, or do it by hand and state where it would have gone.
Reflection: You just refused to send confidential text to the cloud on purpose. Describe the pull to just paste it into the convenient tool. That pull — convenience vs trust — is the whole chapter. Which won, and why?
Rep 9 — Pull the plug
Turn off your Wi-Fi. Try to use (a) your cloud assistant and (b) your local model (skip this half if you’re on the stand-in — and note that the stand-in would also be dead offline). Record what happened.
Reflection: Name one real situation in your work or life where you’d be offline and still need AI. What does that tell you about keeping a local model installed as a fallback?
Rep 10: The bridge, in your own words
Rep 10 — Explain the “OpenAI-compatible endpoint” to a coworker
Without jargon, write three sentences explaining to a non-technical colleague why a tool built for ChatGPT can sometimes be pointed at a model on your own laptop instead — and why that’s good for privacy. Use the §11.7 one-sentence idea as your seed, but say it your way.
Reflection: Could a smart coworker who’s never heard the phrase follow your explanation? If not, cut a word until they can. Teaching it is how you prove you own it.
Done? One Last Thing.
A miniature of Project 11, end to end. Pick three real (fictionalized) tasks from your work spanning at least two sensitivity tiers. For each:
- Classify its tier and write the allowed venue (from
code/data-sensitivity-guide.txt). - Run it in a frontier cloud model and your local/stand-in venue.
- Score quality (1–5), speed (fast/ok/slow), and privacy (data left / stayed).
- Write one sentence: for this task, going forward, I will use ___ because ___.
Drop the four columns into a copy of code/cloud-or-local.csv. If you can do this for three tasks tonight, you can do the whole project: run the showdown, score it, and defend the routing. That’s the job in miniature — AI drafts, you decide where the data is allowed to go, and you sign the verdict.
Up next: Project 11 — Project P11: Frontier vs Local Showdown.