Local kit · no API key · v1.0.0

How to protect yourself over the next 2 years of AI

Three risk horizons, one kit for each: a prepared home, locked-down accounts, measured work and a 90-day plan.

Start the guide View on GitHub
Banner: Survive AI — the next 2 years in 3 horizons: security, income and identity
In short

sobreviver-ia is a set of simple tools that run on your own computer, to get through the next two years of AI with fewer surprises. It is for any person or family: it works out the food, water and cash for a 14-day kit, installs an AI and Wikipedia offline, shows which of your accounts still depend on SMS and measures, with tasks from your own work, how much AI already does for you. All you need is Python 3 (and, if you want the offline AI, disk space). Nothing uses an API key.

What it is

Three risk horizons

The risk does not arrive all at once. In the short term it is about security; in the medium term, work and income; in the long term, beliefs and identity. Each horizon has its own part of the repository.

Banner with the 6 kit items: 14-day kit, Offline AI, MFA not SMS, Personal eval, Own something, People
Now · weeks

🏠 Short term: security

Attacks on internet-connected systems are getting cheaper and more common. Prepare your home for a few days without the usual services and move your accounts off SMS.

  • kit/ food, water, cash
  • offline/ AI and Wikipedia without internet
  • seguranca/ MFA and audit
Next 2 years

💼 Medium term: work and income

Entry-level jobs in the most exposed roles are already shrinking. Use AI for everything, measure how much of your work it does, own something, build distribution and move closer to the physical world.

  • eval/ your personal eval + scoreboard
  • plano/ weeks 1–8
Decade

🧭 Long term: identity

If you are your job title, every AI advance hurts like a threat. Define yourself by character, review your beliefs, keep an open mind and invest in people.

  • plano/ weeks 9–13
Why now · data

What already happened in 2026

Four verified facts show the three horizons are not theory.

19%less employment for 22–25-year-olds in the roles most exposed to AIcumulative gap vs. less exposed roles · Stanford, Aug 2026
350+supermarkets in France had their fridges heated up by an attackup to ~50 °C · Oct 11, 2026 · Reuters
Jul '26models under safety testing escaped the sandbox and broke into part of Hugging Facedisclosed Jul 16–21, report on Aug 26
50–100years of biology progress compressed into 5–10Dario Amodei, "Machines of Loving Grace", 2024

Employment of 22–25-year-olds in the roles most exposed to AI

How far below it fell, relative to the less exposed roles. Each bar is a version of the study, not a year: the number is cumulative.

0%−5%−10%−15%−20%−25% 1st version of the study: −13% 2nd version of the study: −15% Aug 2026 update: −19% −13%−15%−19% 1st version2nd versionAug 2026
View as table
Study versionCumulative gap
1st version−13%
2nd version−15%
Aug 2026 update−19%

What this means

The effect comes through hiring, not mass layoffs.

  • Most exposed roles: writing documents, reports, accounting, customer service, video and text production.
  • Fewer new people enter these roles; those who are already senior stay and review the AI's work.
  • It is not "19% per year": it is the cumulative gap up to August 2026, which grew from one version of the study to the next.
  • In France, the refrigeration controllers were on the internet with weak passwords: a simple attack turned into physical damage in hundreds of stores.

In practice: the first rung of the career ladder is getting narrower, and physical systems depend on digital passwords. Both call for preparation now.

Sources: Stanford Digital Economy Lab, "Canaries in the Coal Mine" (Aug 2026 update); Reuters (France, Oct 11, 2026; other sources say 500–700 stores); public report of the July 2026 incident; Dario Amodei, "Machines of Loving Grace" (2024).

Short term · 14-day kit

How much your home needs

Basis: 2,000 kcal per person per day and 1.5 to 2 liters of drinking water per person per day, for 14 days. The cash is one month of your spending. Play with the numbers.

Calories56,000 kcal
Drinking water42–56 L
5 L jugs12
Cash$ 4,500

    Calories for 14 days, by household size

    Thousand kcal · 2,000 kcal/person/day

    1 person: 28,000 kcal 2 people: 56,000 kcal 3 people: 84,000 kcal 4 people: 112,000 kcal 5 people: 140,000 kcal 285684112140 12345people
    View as table
    Peoplekcal (14 days)
    128,000
    256,000
    384,000
    4112,000
    5140,000

    Drinking water for 14 days

    Liters · range from 1.5 L (mark) to 2 L (bar) per person/day

    1 person: 21–28 L 2 people: 42–56 L 3 people: 63–84 L 4 people: 84–112 L 5 people: 105–140 L 21–2842–5663–8484–112105–140 12345people
    View as table
    PeopleLiters (14 days)
    121–28
    242–56
    363–84
    484–112
    5105–140
    How it works

    One Saturday for the short term, 90 days for the rest

    Everything runs on your computer. Suggested order:

    14-day kit→ Offline AI + Wikipedia→ MFA without SMS→ Personal eval→ Scoreboard with every new model→ 90-day plan
    Prerequisites

    What you need

    Only the first one is required.

    Python 3.9+

    For the kit, audit and eval tools. No extra libraries (matplotlib is optional, only for the scoreboard PNG).

    git clone https://github.com/inematds/sobreviver-ia
cd sobreviver-ia

    Disk space

    For the offline AI: 2 to 5 GB for the model + from 2.5 GB (Wikipedia pt "mini") to 110 GB (en "maxi"). The script shows the total first.

    df -h ~   # how much is left

    Optional: Ollama

    To run the eval on a local model. For cloud models, use Claude Code, Codex or the chat you already subscribe to.

    ollama list  # already have it?
    User guide · step by step

    The tools, one by one

    Real commands, ready to copy. Run them from the repository folder.

    1

    Calculate and build your household kit

    Calories, liters of water, jugs, cash and the shopping list of non-perishables. With --markdown you get a checklist ready to print. The full checklist (first aid, fire extinguisher, flashlight, radio, power bank) is in kit/checklist-casa.md (in Portuguese). The tools print their output in Portuguese.

    python3 kit/calcula_kit.py --pessoas 2 --dias 14 --gasto-mensal 4500
# Calories: 56,000 kcal · Water: 42 to 56 liters (~12 jugs of 5 L) · Cash: R$ 4,500
python3 kit/calcula_kit.py --pessoas 4 --markdown > meu-kit.md
    2

    Install the AI and Wikipedia offline (Linux and macOS)

    The default is dry-run: it shows what it would do and how much space it takes, without downloading anything. Choose the model (llama3.2:3b, qwen2.5:7b…), the languages and the Wikipedia type: mini (article summaries), nopic (full text without pictures) or maxi (with pictures).

    bash offline/instala-offline.sh --modelo qwen2.5:7b --wiki pt,en --tipo nopic
# [1] Ollama + model qwen2.5:7b (~4.7 GB)
# [2] Kiwix + Wikipedia pt/nopic ~9 GB, en/nopic ~55 GB
# Approximate total space: ~68.7 GB
bash offline/instala-offline.sh --modelo qwen2.5:7b --wiki pt --tipo nopic --instalar
kiwix-serve --port 8088 ~/offline/wikipedia_pt_all_nopic_*.zim  # → http://localhost:8088

    Windows: install Ollama with the installer from ollama.com/download/windows (or winget install Ollama.Ollama) and run ollama pull llama3.2:3b in Command Prompt. Install Kiwix Desktop from kiwix.org/applications and download the wikipedia_pt_all_nopic_…zim file from download.kiwix.org/zim/wikipedia; open the .zim in Kiwix.

    3

    Move your accounts off SMS

    Create the accounts spreadsheet (just the name, never the password), mark each one's method (senha, sms, email, app, chave — password, SMS, email, app, key) and get the list of what to migrate first. The step-by-step for each account is in seguranca/checklist-mfa.md (in Portuguese).

    python3 seguranca/auditoria_contas.py --modelo contas.csv      # template spreadsheet
python3 seguranca/auditoria_contas.py --interativo contas.csv  # or ask account by account
python3 seguranca/auditoria_contas.py contas.csv --saida relatorio.md
# Accounts: 7 · With authenticator app or passkey: 1 (14%) · Still on SMS: 4
# 1 | Main email | email | SMS (weak) | replace SMS with an authenticator app or passkey; save codes…
    4

    Build your personal eval

    An eval is a set of 10 to 15 real tasks from your day, always the same ones, with the context files and a criterion for what counts as "done". Each task is a folder: eval/tarefas/NN-nome/ with tarefa.md, contexto/ and criterio.md. It ships with 3 examples (in Portuguese): summarizing a meeting, organizing the month's expenses and replying to an angry customer.

    eval/tarefas/
├── 01-resumir-reuniao/   tarefa.md · contexto/reuniao.txt · criterio.md
├── 02-planilha-gastos/   tarefa.md · contexto/extrato.csv · criterio.md
└── 03-email-cliente/     tarefa.md · contexto/email.txt, politica.md · criterio.md
    5

    Run the eval with every new model

    By default the script generates a package for you to paste into Claude Code, Codex or the chat you already subscribe to. With --ollama it runs directly on a local model. It never calls a paid API and refuses addresses that are not on your machine.

    python3 eval/rodar.py --modelo claude-novo     # generates eval/saida/<data>-claude-novo/pacote.md
python3 eval/rodar.py --ollama qwen2.5:7b      # runs on local Ollama (localhost:11434)
# then: read each answer against criterio.md and give the score (0, 0.5 or 1) in notas.csv
    6

    Track the scoreboard

    Each round becomes a line in eval/placar.csv. The chart shows how the percentage of your work that AI already does evolves. Only you know this number, because it comes from your tasks.

    python3 eval/placar.py registrar eval/saida/2026-10-19-qwen2.5-7b/notas.csv
python3 eval/placar.py ver
python3 eval/placar.py grafico   # eval/placar.svg (+ .png if you have matplotlib)
    7

    Follow the 90-day plan

    plano/plano-90-dias.md (in Portuguese) has a week 0 for the short term and 13 weeks with exercises: own something, build distribution, move closer to the physical world, identity by character, belief review, open mind and people. On day 90, run the eval again and compare.

    python3 -m pytest -q   # 22 tests for the tools
    Medium and long term · plan

    90 days, week by week

    Summary of plano/plano-90-dias.md (in Portuguese). This is not financial advice; adjust it to where you are.

    Week 0
    The short term in one Saturday14-day kit, offline AI installed, your 3 most important accounts off SMS.
    Wk 1–2
    Use AI for everything and build the eval10 to 15 real tasks from your work; the first score on the scoreboard. Don't rely on other people's opinions about what AI can do: measure it.
    Wk 3–5
    Own somethingAI makes work cheaper; what becomes valuable is what is yours: a business, product, clients, assets. Build the minimum version of something of your own.
    Wk 6–8
    Build distribution and move closer to the physicalAn audience, a list and a community go with you if your job changes. Work with a physical presence tends to last longer.
    Wk 9–10
    Identity by character, not by job title"I am someone who…" without naming a company or a title.
    Wk 11–12
    Review beliefs and keep an open mindWhich of your beliefs depend on the world before AI?
    Wk 13
    Invest in peopleNeighbors by name, contacts reconnected, one get-together a month. In any scenario, people with strong relationships do better.
    Day 90
    ReviewEval again on the newest model, distribution recounted, kit and MFA reviewed.