What 100 Days of Daily Local AI Use Taught Me About the Hype

A hundred days ago I committed to using local AI on my Mac every working day and — the part that matters — keeping a log. Not a journal of feelings; a log: what I used, for what, whether it helped, in a plain-text file appended via a Raycast hotkey. The whole experiment cost me maybe four minutes a day of logging, and it produced the most honest dataset I have about what AI actually does to my work.

This is the retrospective. Numbers included, hype audited, embarrassments admitted.

What stuck — the boring 80%

The headline finding is anticlimactic, which is itself the finding: the AI use that survived 100 days is the unsexy use. Four workflows account for about 80% of my logged sessions.

Transcription — the undisputed champion: 94 of 100 days. Whisper (large-v3-turbo, via MacWhisper and whisper.cpp) transcribing meetings, voice memos walked into my phone, and Czech interviews. It is the only AI tool I’d call genuinely flawless in daily use, and the only one my non-techie friends ask about after seeing it.

Summarization — 81 days. Meeting transcripts, long PDFs, the daily skim of newsletters compressed into bullets. Qwen 2.5 14B via Ollama handles this so well that I stopped sending these to cloud models entirely around day 20.

Code assistance — 76 of the ~70 days I wrote code (some days had multiple distinct sessions). The mix shifted over the hundred days: less “write this for me,” more “review this diff,” “explain this error,” “write the tests I’m too lazy to write.” The reviewer-and-explainer role aged better than the author role.

Translation — 64 days, my Czech↔English bridge. Drafting English from Czech thinking, translating client material back, and the niche superpower: checking that my English emails don’t carry Czech sentence structure (“word order sounds translated — fix only that”).

Notice what this list is: drudgery compression. Nothing in the top four is “AI does my job.” All of it is “AI does the part of my job a diligent intern with infinite patience would do.”

What I abandoned — the graveyard

The log is just as valuable for what stopped appearing in it.

AI-generated social posts: dead by day 23. I generated LinkedIn and X posts from my articles for three weeks. They were grammatical, on-message, and utterly fungible — they sounded like everyone else’s AI posts, because they were everyone else’s AI posts with my nouns swapped in. Engagement on AI-drafted posts ran visibly below my hand-written ones. Verdict from the log, day 23: “reads like a press release about a human.” Now AI suggests angles and I write the words.

The over-automated pipeline: dead by day 41. I built a beautiful chain — RSS → summarize → categorize → auto-file into Obsidian with tags. It broke silently twice: once when a model update changed output formatting and the parser quietly filed garbage, once when a feed URL died and the pipeline happily processed nothing. I discovered each failure more than a week later. A workflow that fails loudly is a tool; a workflow that fails silently is a liability. I rebuilt it with one manual confirmation step, and honestly, the 30 seconds of human glance was the missing component all along.

The autonomous agent: dead by day 58. An agent that triaged my inbox into drafts and to-dos. On paper, an hour saved daily. In practice, I reviewed every draft anyway (it was my name on the sends), corrected a meaningful fraction, and the reviewing took longer than writing — because verifying someone else’s judgment is slower than exercising your own. The log entry that killed it: “supervision cost > replaced labor.” That sentence generalizes further than I’d like.

The surprises

Two things I did not predict at day 0.

The biggest local-AI win was psychological, not technical. I expected the case for local models to be privacy and cost. The actual daily difference was the absence of metering anxiety. With API pricing, some quiet accountant in my head evaluated whether each question was “worth a call.” With a local model, that accountant is unemployed — and I experiment shamelessly. Dumb questions, half-formed ideas, running the same prompt five ways to see what changes, feeding it whole books. A meaningful chunk of what I learned in these 100 days happened only because trying things cost zero. Flat-rate beats metered for learning, even when metered is objectively cheap.

Privacy mattered more over time, not less. I assumed privacy was a starting principle that would fade into background. The opposite happened: as local AI proved trustworthy, I fed it progressively more sensitive material — contracts, financials, candid meeting audio, half-baked business ideas — things I simply would not have pasted into a cloud box. The trust ratchet only turns one way, and it unlocked the most valuable use cases last.

The hype audit — claims vs. log

The fun part. I took the loudest claims of the AI discourse and checked them against 100 days of logs.

“10x productivity” → measured: 1.3–1.5x, sometimes ~1.0x. On writing-heavy and admin-heavy weeks, my output (posts shipped, emails cleared, transcripts processed) ran roughly 30–50% above my pre-experiment baseline — a genuinely big deal compounded over a year, and nothing like 10x. On deep-thinking weeks — architecture decisions, strategy, anything where the bottleneck was me understanding the problem — the gain was approximately zero. AI cannot think for you faster than you can think; it can only clear the rubble around the thinking. Anyone reporting 10x is either doing work that was 90% rubble or measuring vibes.

“AI will replace your apps” → it replaced exactly 4. A paid transcription subscription, a translation tool, a text-summarizer subscription, and a grammar checker — roughly €25/month of subscriptions, gone. It did not replace my editor, terminal, notes app, mail client, or browser; it embedded into them. The model is an ingredient, not a meal.

“Agents will run your life” → not this year, not for me. See the day-58 graveyard entry above. Constrained, supervised, single-purpose automation: yes, daily. Autonomous open-ended agents: the supervision tax still exceeds the labor saved for my work. I’ll re-test in the next hundred days; trajectory matters more than snapshots.

“Local models are toys” → also false. The boring 80% — transcription, summarization, translation, code explanation — runs entirely on a 14B model and Whisper. The hype is wrong in both directions at once: AI is less transformative than advertised and local AI is more capable than dismissed.

The durable lessons

If I compress 100 days into three sentences I’d defend in court:

AI compresses drudgery, not thinking. Every surviving workflow removes mechanical work — transcribe, summarize, translate, reformat, explain. Every abandoned workflow tried to outsource judgment. The boundary between those two is the most useful line in applied AI, and you find it empirically, not from a keynote.

Workflow design beats model choice. I spent the first twenty days benchmark-chasing — comparing models, quants, runtimes. Switching models changed my results by single-digit percentages. Redesigning how a task flows — where the human checkpoint sits, how output lands in the right app, the prompt’s failure-proofing — changed results by multiples. The model is the engine; nobody’s commute improves from engine swaps alone.

The best AI setup is the one you stop noticing. Around day 70 the log entries got shorter — not because I used AI less, but because using it stopped being an event. Hotkey, transcript, done. That invisibility is the actual finish line. If your AI setup still feels like operating impressive machinery, it isn’t done.

The next 100 days

The experiment continues, with a new hypothesis list: a fine-tuned small model on five years of my own writing (can it kill the day-23 social-media problem?), RAG over my full Obsidian vault as a daily driver rather than a demo, a re-test of supervised agents on narrow workflows with loud failure modes, and vision models on my photo library — the one major modality my log barely touched.

Same method: every day, four minutes of honest logging. Because the real lesson of the first hundred days is meta: in a discourse running on demos and hot takes, a plain-text log of what you actually did is an unfair advantage. Day 101 starts tomorrow.