built throughORANGEBOX·see what it ships·$1 →

::AtomEons · the comprehensive ai gateway · 2026-05-21

AI is rewriting the labor market.
Here's how to be on the winning side of it.

The on-ramp for forty-four million workers whose jobs are exposed to generative AI in the next decade. Named tools. Named builders. Twenty concrete revenue paths. Fifty-plus honest answers. No course to buy. No bootcamp. No upsell.

::TL;DR · the whole page in three lines

  • WHOFor anyone whose job is exposed to AI displacement — or anyone tired of vague hype and ready for named tools, named people, and real numbers.
  • MAPNamed AI tools with free-tier guidance (Claude, ChatGPT, Gemini, Perplexity, Cursor, Claude Code, Ollama, etc.) · named builders to follow · 50+ honest FAQs · 30-60-90 day plan.
  • EARN20 specific revenue paths with realistic income ranges — 8 of them require zero coding.

::00 · why this page exists

We are the on-ramp for the forty-four million.

The International Labour Organization shipped data in January 2026 putting one in four jobs worldwide at exposure to generative AI displacement. The top tier is 3.3% — jobs where most of the daily work can be done today, not someday, by a model you can rent for pennies an hour. In the United States that translates to roughly forty-four million workers in the next decade. The exposure rate in high-income economies is 34%. In low-income economies, 11%.

That is our user base. Not VCs. Not researchers. Not other founders. People who used ChatGPT seven times for a recipe and a birthday poem and felt like an idiot every time. People who got a 4am Singapore email last month and watched fifteen years of service end inside a Slack notification. People whose child just asked "is ChatGPT going to replace teachers" and who don't know how to answer.

The other AI labs aren't building for them. Anthropic writes for the already-fluent. OpenAI ships the next developer keynote. Google has the inventory of the planet and aims at distribution. Every "AI for everyone" announcement comes packaged inside an enterprise sales motion priced at two thousand dollars a seat per year.

This page is the alternative. Free. Specific. Ruthless about which tools are worth your money, which builders to trust, and which bootcamp-bro "AI course" sellers to walk away from. We don't take venture money. We don't gate education behind a subscription. We charge a dollar for our cockpit because anyone with an itch to learn should be able to afford the real tool.

If you read one section: the 20 revenue paths. If you read three: the revenue paths, the tools we trust, and the 90-day plan.

::01 · what ai actually is

A calculator for everything that used to require reading and writing.

In one paragraph: a Large Language Model is software trained to predict the next word in a sequence. That sounds modest. The trick is that prediction at sufficient scale looks indistinguishable from comprehension. The model that powers Claude has read trillions of words. To predict the next word in "the capital of France is  ___," it had to internalize geography. To predict the next word in "the moral of the parable of the prodigal son is ___," it had to internalize the parable. To complete your half-written email persuasively, it had to internalize what makes emails persuasive. We call this AI because the behavior crosses the threshold where pretending otherwise stops being useful.

For the long version with examples, six concrete things you can do tonight, and six honest limits, read /start first — that's our 11-minute on-ramp written for someone who has used ChatGPT under ten times. Then come back here for the tools, the people, the money paths, and the FAQ.

::02 · tools we trust

The tools we trust — with reasoning.

We use everything below directly. No affiliate links. No sponsorships. If a tool drops off this list later it's because the lab stopped trusting it, not because someone paid us to swap it out.

Anthropic Claude

general chat · the careful one

free tier (Sonnet) covers daily personal use

The most thoughtful default. Long context (1M tokens on Opus 4.7). Honest about what it doesn't know. The lab's primary working model. If you only learn one chat AI, learn this one — it transfers to every other surface.

ChatGPT

general chat · the popular one

free tier with GPT-4o-mini

Best free-tier voice mode. Largest user base. Image generation built in. Best 'just give me an answer' interaction for non-technical first-timers. Pair with Claude — use ChatGPT for fast informal tasks, Claude for anything that has to be right.

Google Gemini

general chat · the long-context one

generous free tier · 1.5M context on free tier

Free tier matches paid Claude on context length. Free Gmail / Docs / Drive integration is genuinely useful. Best for processing huge PDFs and long documents on a free plan. Worse on coding than Claude.

Perplexity

AI search · cite-the-source

free tier covers ~5 pro searches/day

Best replacement for Google when you actually want sources. Every answer cites links. Use for any factual question where you need to verify. The lab uses it for current-events lookups.

Claude Code (CLI)

AI coding · the agentic terminal

$20/mo Claude Pro — best price/perf in coding

Reads your whole repo. Edits multiple files. Runs your tests. Commits. The most capable coding agent shipping in 2026. Pairs with our ORANGEBOX cockpit for project memory.

Cursor

AI coding · the IDE

free tier covers light personal use · $20/mo Pro

VS Code fork with Claude / GPT / Gemini baked in. Best for solo developers who want IDE comfort. Cmd+K to edit inline. Cmd+L for the agent panel.

GitHub Copilot

AI coding · the autocomplete

free tier (50 chats / 2000 completions monthly)

Best autocomplete-in-editor experience. Pairs cleanly with VS Code, JetBrains, Neovim, Xcode. Use it for autocomplete, use Cursor or Claude Code for refactoring.

v0 by Vercel

AI coding · the UI generator

free tier covers ~3 generations/day

Type a description, get a working React component. Best for landing pages, dashboards, forms. Output is real shadcn/Tailwind code you can drop into any Next.js project.

Replit

AI coding · the cloud IDE

free tier covers learning + small projects

Browser-based dev environment with AI built in. Best for absolute beginners — no install, no terminal, just code. Replit Agent ships working web apps end-to-end.

Ollama

local AI · run models offline

100% free, no signup, no cloud

Download a model, run it on your laptop. No data leaves your machine. Best for sensitive content, code with NDA concerns, or just learning how LLMs work. Pair with `llama3.3:70b` (Meta) or `qwen2.5-coder:32b` (Alibaba) for coding offline.

Hugging Face

open AI · the model hub

free model downloads, free transformers library

GitHub for AI models. 800,000+ open models, datasets, demos. The lab's research arm pulls models from here. Best if you ever want to look under the hood.

LM Studio

local AI · the friendly UI

free for personal use

Same idea as Ollama but with a GUI. No terminal required. Best for non-developers who want to try local AI without committing to the command line.

Midjourney

image gen · the artistic one

free tier paused as of 2024 · $10/mo basic

Highest aesthetic quality for stylized art. The 10 cinema stills on our /research/lessons-from-sci-fi page were Midjourney v8.1. Best for moodboards, marketing visuals, book covers.

caveat · Subscription-only since the trial ended.

DALL-E (via ChatGPT) & Adobe Firefly

image gen · the safe one

DALL-E in ChatGPT free tier (limited)

Best when you need commercially-safe images. DALL-E (inside ChatGPT) is trained on licensed content. Adobe Firefly explicitly excludes copyrighted material. Use these for client work where IP matters.

ElevenLabs

voice · the realistic one

free tier covers ~10 min/month

Cleanest voice-clone-from-3-minute-sample on the market. Best for podcasters, audiobook narrators, YouTube creators. Voices sound human, not robotic.

Whisper

voice · transcription

open source — runs anywhere

The transcription model. ~95% accuracy on English. Run it locally for free, run it in ChatGPT voice mode for free, run it via OpenAI API for $0.006/min. Used for podcast transcripts, meeting notes, accessibility captions.

Runway

video gen · the cinematic one

free trial credits

Text-to-video, image-to-video, video editing. Best for short clips, social content, music videos. The shortest path from script to moving image.

Vercel

hosting · the Next.js home

generous free tier · hobby projects free

Where atomeons.com lives. Push to GitHub, site deploys in 60 seconds. Best for any AI app you build with Next.js. Free tier covers most personal projects forever.

Supabase

database · the open one

free tier covers small apps · $25/mo Pro

Postgres in the cloud with auth, storage, edge functions, realtime. The lab uses it for the Founder's View archive. Best for any AI app that needs to remember users and data.

Stripe

payments · the standard

free to integrate · 2.9% + 30¢ per transaction

The payment processor every solo founder uses. Connect mode for marketplaces, Checkout for one-shot sales. The lab takes ORANGEBOX + B00KMAKR purchases through Stripe.

Loops & Resend

email · transactional + marketing

Loops 5k contacts free · Resend 3k/mo emails free

Loops for marketing + transactional in one. Resend for purely transactional (cheaper, simpler). The lab uses Loops for the Founder's View list and Resend as a failure-soft fallback.

Æ ORANGEBOX

ours

AI cockpit · ours · turbo-optimize Claude

FREE this week · perpetual after · 30-day Material Failure Guarantee

The lab's own product. v1.0.0-beta shipped 2026-05-30 as the turbo-optimization system for Claude — persistent memory across sessions, 10-80× context compression, reusable skill primers, tamper-evident JSON receipts on every action, 14-department named-role routing (AE0-AE14). Optional fallback to GPT/Gemini/Ollama via BYO key. Local-first. Zero telemetry. License §4A bans us from ever switching to subscription. Source included.

caveat · Windows 10/11 at v1.0.0-beta. macOS + Linux on roadmap.

Claude Desktop

general chat · the desktop app

free with Claude account · same tiers as claude.ai

Native macOS / Windows app. Persistent window, no browser tab clutter, voice mode, MCP server connections, file uploads with proper Finder integration. If Claude is your daily AI, the desktop is a better home for it than a browser tab.

Claude API (via Anthropic Console)

AI building · pay-per-token

$5 free credit on signup · no recurring cost · pay per call

The API surface for building AI products. Same model intelligence as Claude.ai but you control the integration. Pricing: Sonnet 4.7 is ~$3/MTok input + $15/MTok output. Prompt caching cuts that 90%. Used by every founder shipping AI products in 2026.

Mistral

general chat · the European one

free tier covers daily use

French AI lab. Their large models (Mistral Large 2, Codestral) compete with Claude on code and reasoning at a lower price point. Best if you want EU-data-residency (Mistral hosts in France), or for any team that prefers a non-US AI provider for regulatory reasons.

DeepSeek

AI reasoning · the cheap one

free tier · API at ~10x cheaper than Claude/GPT

Chinese AI lab. DeepSeek R1 / V3 deliver strong reasoning at API prices around $0.27/MTok input vs Claude Sonnet's $3. If price is the bottleneck on your AI product, DeepSeek beats the field. Trade-off: hosted in China — not safe for confidential business data.

caveat · Don't send confidential business data — hosted in China.

Qwen (Alibaba)

AI · the open-weights flagship

free tier · open weights available on Hugging Face

Alibaba's flagship. Strong coding performance, multimodal (vision + audio), and open weights (run locally via Ollama with `qwen2.5-coder:32b`). Best open-source coding model in 2026 by most benchmarks. The lab's local fallback when Claude API is unavailable.

Together AI · Groq · Cerebras

inference · run open models at speed

free signup credits at all three

Three competitors that host open-weight models (Llama, Mistral, Qwen, DeepSeek) at API endpoints. Together AI: broadest model catalog. Groq: cheapest, fastest tokens-per-second on the market for ~10 specific models. Cerebras: the fastest inference for narrow models with their wafer-scale chips. Use these when you need open-model inference without running it yourself.

::03 · builders we read

The builders we read. The ones who teach, not sell.

None of these people are taking your money to teach you AI. They publish because they want the field to be better. Some sell related products (books, courses, consulting). None have built their living off "here's how to make $10k/month with AI" pitches. Follow them and you stay in signal.

Andrej Karpathy

ex-OpenAI · ex-Tesla AI · Eureka Labs

What: Education-first. His 'zero to hero' YouTube playlist is the best free intro to how LLMs actually work.

Why: Closest thing to a public AI professor. No marketing voice. Builds working models from scratch on camera. If you want to know what a neural network IS, this is the entry door.

Simon Willison

co-creator of Django · indie LLM blogger

What: Runs simonwillison.net — the most consistently useful LLM blog in the world. Daily-grade signal.

Why: Tracks every model release, every API change, every interesting paper. No hype. Has been writing this beat since GPT-2. The lab subscribes to his Atom feed.

François Chollet

creator of Keras · ARC-AGI benchmark

What: Stress-tests AI claims with hard reasoning benchmarks. Author of ARC-AGI, the puzzle benchmark current LLMs still fail at.

Why: The serious skeptic. When the hype gets loud, Chollet's the one running the math. Best follow for understanding what current AI can and can't do.

Geoffrey Hinton

Nobel Laureate (Physics, 2024) · ex-Google

What: One of the 'godfathers of deep learning.' Left Google to speak freely about AI risk.

Why: The person who can credibly say 'I helped build this and it scares me.' His interviews and 2024 Nobel lecture are required listening.

Yann LeCun

Meta · Chief AI Scientist · NYU professor

What: Open source champion. Architect of Meta's Llama model family.

Why: Pushes back hard against AI doom narratives. Believes open weights are how the field stays healthy. The reason Llama exists for you to download.

Demis Hassabis

Google DeepMind CEO · Nobel Laureate (Chemistry, 2024)

What: AlphaFold author. Long-game scientific AI rather than viral chatbot AI.

Why: Shows what AI looks like when the goal is scientific discovery (protein folding, weather forecasting, math proofs) rather than ad revenue. The contrast matters.

Hamel Husain

ex-Outerbounds · LLM evaluation consultant

What: Runs hamel.dev. Practical writing on how to actually measure if your LLM app works.

Why: Every founder building an AI product hits 'is it actually working' on week three. Hamel's the canonical answer to that question. Read 'Your AI Product Needs Evals' first.

Jeremy Howard

fast.ai co-founder · Answer.AI

What: Deep learning democratization. fast.ai's free course taught a generation of practitioners.

Why: If you want to learn AI deeply for free, fast.ai is the canonical entry. The course assumes you can code; it does not assume a PhD.

Chip Huyen

author · Stanford lecturer

What: Wrote 'Designing Machine Learning Systems' and 'AI Engineering.' Practical, production-focused.

Why: Best resource for the 'I want to actually ship an AI product' gap between hobby code and real systems. The lab uses 'AI Engineering' as a reference.

Ethan Mollick

Wharton professor · 'Co-Intelligence' author

What: Practical AI essays for working professionals. Newsletter: One Useful Thing.

Why: Best writer for the 'what does this mean for my actual job' question. No jargon. Specific examples. Updated weekly.

Riley Goodside

Scale AI · ex-staff prompt engineer

What: First person publicly known as 'the prompt engineer.' Lives on Twitter showing weird LLM behavior.

Why: If you want to actually understand how LLMs respond to prompt structure, follow the person who got paid full-time to figure it out.

Anthropic Research Team

Anthropic · publishers of Claude

What: Most readable AI safety research papers shipping. 'Constitutional AI' and 'Sleeper Agents' are the entry papers.

Why: The papers are written for engineers, not professors. Best place to learn what alignment work looks like in practice.

Lilian Weng

Thinking Machines (ex-OpenAI head of safety)

What: Runs lilianweng.github.io. Long, deep, citation-heavy essays on agentic AI, RLHF, hallucination, attention mechanisms.

Why: When you want to understand a research concept properly (not just the marketing), Lilian's essay on it is usually the canonical compressed reference. Free, written for working engineers, updated steadily.

Sebastian Raschka

Ahead of AI · ex-Lightning AI

What: Author of 'Build a Large Language Model From Scratch.' Weekly newsletter that translates new papers into practical takeaways.

Why: Closest thing to a working textbook for the LLM-engineering era. His 'from scratch' book is the only one the lab keeps next to the keyboard.

Maxime Labonne

Liquid AI · post-training researcher

What: Open-source LLM fine-tuning, model merging, quantization. Author of 'LLM Engineer's Handbook' and the LLM Course on GitHub.

Why: If you want to actually run / fine-tune open models (Llama, Mistral, Qwen, DeepSeek), Maxime's open-source recipes are the most reproducible in the field. Free, on GitHub, current.

Eugene Yan

Amazon · applied ML scientist

What: eugeneyan.com — applied ML and LLM systems writing, with an emphasis on ship-able patterns from a production engineer.

Why: Best follow for the 'how do real teams actually ship LLM apps in production' question. His 'Patterns for Building LLM-based Systems' essay is essentially the canonical reference for that gap.

Nathan Lambert

Ai2 · Interconnects newsletter

What: Interconnects (interconnects.ai) — weekly analysis of post-training, RLHF, open model releases.

Why: If you want to understand WHY a new model release matters (not just that it exists), Nathan is the canonical post-training analyst. Open-source friendly. No hype.

Sayash Kapoor + Arvind Narayanan

Princeton · 'AI Snake Oil' authors

What: Authors of 'AI Snake Oil' (2024) and the Substack of the same name. Hard, evidence-based pushback on AI hype.

Why: The serious 'AI is not as good as the press release says' counterweight. Cite when someone in your life is being sold an AI grift. The book is the most useful gift you can give a non-technical relative.

::not on this list

We will not name names, but a heuristic. If someone's public AI persona is mostly selling you their AI course rather than using AI to build things; if their pinned tweet promises $10K/month in 30 days; if the "free workshop" is a 90-minute sales pitch for a $2,000 mastermind — the value transfer is going from you to them, not the other way around. Walk away. The people on the list above don't pitch you. They publish, you read, you decide.

::04 · 20 ways to make money with ai

Twenty specific revenue paths. With realistic numbers.

None of these require "giving up your day job." All twelve are documented from people actually doing them in 2026. Numbers below are realistic for someone who is consistent for 3-6 months — not the cherry-picked screenshots from someone's course landing page.

01

Ghostwrite Substack newsletters

for: anyone with writing taste

Find a founder or consultant who knows their topic but hates writing. Offer to ghostwrite their weekly Substack at $1,500/mo. They send you a 30-min voice memo, you turn it into a 1,500-word essay with Claude. Three clients = $4,500/mo. The skill is editing AI output until it sounds like the specific human, not generic AI prose.

tools:

ClaudeSubstackLinkedIn DMs

realistic · $500-3,000/mo per client · 2-4 clients realistic

02

Build websites with v0 + Vercel

for: ex-marketing, ex-design, no-code people

Local business websites. Restaurants, contractors, dental offices, salons. Use v0.dev to generate a clean design, customize it, deploy to Vercel. Charge $2,500 for a 5-page site + simple booking form. Compete on speed (one week) and quality (these look better than what the local web agency ships).

tools:

v0VercelLoomStripe

realistic · $1,500-5,000 per site · 2-5 sites/month

03

AI-assisted bookkeeping for small businesses

for: ex-accountants, anyone with QuickBooks experience

Local businesses hate bookkeeping. Charge $500/mo to handle their monthly close. Use Claude to categorize transactions, reconcile statements, prep quarterly tax estimates. The bookkeeping skill is yours. AI just makes you 5x faster than the old way.

tools:

ClaudeQuickBooksExcel

realistic · $300-800/mo per client · 5-15 clients realistic

04

AI-augmented translation

for: anyone fluent in two languages

Take the AI's first-pass translation, edit it for tone and accuracy. Charge per word. With AI you can translate 10,000 words/day instead of 2,000. Specialize in a domain (legal, medical, technical) and the rate goes up. Localization for video game studios pays particularly well.

tools:

ClaudeDeepLGoogle Docs

realistic · $0.05-0.15 per word · 2-5x output of human-only

05

Voice-clone audiobook production

for: podcasters, voice actors, audio engineers

Indie authors want audiobook versions but can't afford the $5k+ that traditional narrators charge. Clone an author's voice (with their permission) using ElevenLabs, narrate the book, clean up artifacts, master, submit to Audible / Spotify / Apple Books. Production time: 1-2 days vs the traditional 2-4 weeks.

tools:

ElevenLabsAudacity / ReaperACX

realistic · $1,500-5,000 per book · 2-3 books/month

06

MCP server / Claude Code agent dev

for: developers · the highest-paying bucket

Companies want their internal tools accessible to Claude. Build them an MCP server that exposes their CRM / database / docs / API to Claude as tools. Niche skill, high demand, low supply. Senior MCP dev rates in 2026 are $200-400/hr.

tools:

Claude CodeMCP SDKTypeScript / Python

realistic · $10,000-50,000 per project · 2-6 projects/year

07

AI-research-to-tweet pipeline

for: anyone who reads quickly

Pick a sub-niche of AI (e.g., 'AI for healthcare,' 'AI for legal,' 'AI for biotech'). Read 10 papers per week. Use Claude to compress each into a 280-character thread. Build to 10k followers. Sell sponsorship slots in the newsletter that grows out of it.

tools:

ClaudeX / TwitterNotion

realistic · $500-5,000/mo from a sponsored newsletter · scales

08

AI-trained customer service for SMBs

for: anyone with customer service ops experience

Small businesses can't afford full-time CS staff. Set up a Claude-powered chatbot trained on their FAQ + product docs. Triage every inbound. Hand off to the human only on actual edge cases. Monthly retainer to maintain training + improve handoff logic.

tools:

Claude APIIntercom or HelpScout

realistic · $2,000-8,000/mo retainer · 3-8 clients realistic

09

Personalized AI tutors for individuals

for: teachers, instructors, coaches

Parents pay for AI-augmented tutoring because the AI never gets tired, never gets bored, and remembers everything the kid said two weeks ago. You bring pedagogy. AI brings infinite patience. Charge premium for the combination. SAT prep, college admissions, language learning are the highest-paying niches.

tools:

Claude / ChatGPTZoomGoogle Docs

realistic · $80-300/hr · 10-25 hours/week

10

Indie SaaS — one operator, one problem

for: any developer · highest variance bucket

Pick ONE narrow problem you've experienced. Build the smallest possible tool that solves it. Charge $19/mo. The full stack — Claude Code to write it, Cursor to maintain it, Vercel to host it, Stripe to bill it, Loops to email — costs ~$40/mo to run. Profitability starts at customer #3. Most people never get to customer #3 because they keep adding features instead of marketing.

tools:

Claude CodeCursorv0StripeVercel

realistic · $0-50,000/mo · 90% fail · 10% become real businesses

11

AI-generated images / video for marketing

for: ex-photographers, ex-videographers, ex-designers

Local businesses, e-commerce, real estate, restaurants — all need a constant stream of visual content. Replace the $5k photo shoot with a $300 AI generation. Your skill is the EYE — picking the strongest variant, editing for brand. The AI just removes the studio rental.

tools:

MidjourneyRunwayAdobe FireflyChatGPT

realistic · $300-2,000 per asset · 5-20 assets/week

12

AI-augmented therapy notes (compliance-sensitive)

for: licensed therapists · LCSW, LMFT, PsyD, PhD

This isn't a business — it's a way to NOT BURN OUT. Therapists spend 30-50% of their week on documentation. Use Whisper to transcribe sessions (with patient consent), Claude to draft SOAP notes, then edit. Drop documentation time by 70%. Same income, fewer hours. License compliance is critical — check your state board on AI-assisted documentation before adopting.

tools:

WhisperClaude (HIPAA path)EHR system

realistic · 5-15 hours/week reclaimed · pays for itself in week one

13

LinkedIn ghostwriting for executives

for: writers, ex-marketers, ex-comms

Senior execs at mid-size companies want to be visible on LinkedIn but hate writing. Offer to manage their feed at $3K/mo. Three posts/week, all in their voice. You interview them once monthly for 60 min, mine the recording for opinions, draft a month of posts with Claude, edit until it sounds like the person not like AI, schedule in advance. Roughly 8 hrs/mo per client. Three clients fills a calendar.

tools:

ClaudeLinkedInCalendly

realistic · $2,000-5,000/mo per client · 2-3 clients realistic

14

AI-staged real estate photography

for: ex-photographers, ex-stagers, designers

Empty house photographs are unsellable. Traditional staging costs $2K-$5K and takes weeks. With Midjourney + Photoshop you can virtually stage interior shots in 2 hours per listing. Furniture, art, plants, warm lighting, all matched to the home's actual fixed elements. Charge $400/listing. Real-estate agents pay willingly because every week a listing sits empty costs them more than the staging.

tools:

MidjourneyPhotoshop / AffinityLightroom

realistic · $200-800 per listing · 8-15 listings/week

15

AI-augmented resume + cover letter editing

for: ex-recruiters, ex-HR, ex-career coaches

Two paths. Path one: per-document editing at $250 — client sends current resume + job posting, you return ATS-optimized version + cover letter in 24 hours. Path two: full career package at $1,500 — resume + cover letter + LinkedIn profile rewrite + interview prep + 30-day follow-up. Higher conversion on the package. Path in: post weekly LinkedIn content showing before/after rewrites (anonymized).

tools:

ClaudeGoogle DocsZoom

realistic · $150-500 per resume · 8-20 clients/week

16

Local-business SEO with AI

for: ex-marketers, ex-agency, web designers

Plumbers, electricians, dentists, restaurants. Every local business owner knows their Google Business Profile matters; almost none of them have time to maintain it. Offer monthly: weekly review responses (Claude drafts, you edit + send), monthly post + photo, quarterly category + service updates, ranking report. $750/mo flat. Twelve clients = $9k/mo. Pure retainer business. Compounds with referrals.

tools:

ClaudeGoogle Business ProfileLocal rank trackers

realistic · $500-2,000/mo per client · 5-15 clients realistic

17

Etsy listing optimization with AI

for: ex-marketers, copywriters, anyone with taste

Most Etsy sellers are makers, not marketers. Their listings have title-stuffing, blurry photos, and weak copy. Offer to redo the title (SEO-optimized), description (story-first), photos (AI-enhanced or relit), and tags. $150/listing flat. Sellers see ~30-50% more views within two weeks. Path in: reach out to sellers in one specific category (wedding stationery, dog collars, candles) — niche specialization beats generalist.

tools:

ClaudeMidjourneyEtsyEverBee

realistic · $50-300 per listing · 15-40 listings/week

18

AI-augmented legal docs for small firms

for: paralegals, contract managers, legal-adjacent

Small law firms (1-3 attorneys) can't afford a full-time paralegal but drown in routine docs. Offer monthly retainer for: NDA review + redline (Claude drafts changes, attorney signs off), contract templating, deposition summaries, brief outlines. NOT giving legal advice — strictly drafting + research support that the attorney reviews. Verify malpractice insurance covers AI-assisted work in your state. Path in: cold-email solo + 2-attorney firms in your state.

tools:

ClaudeWestlaw / LexisMS Word

realistic · $1,500-4,000/mo retainer · 3-6 clients realistic

19

AI-assisted tax season prep

for: ex-accountants, EA, bookkeepers, tax preparers

Tax season is finite (Jan-Apr in US). AI accelerates schedule prep, K-1 wrangling, 1099 reconciliation, audit risk scoring. Solo prep at $400/return × 100 returns × 4 months = $40K. Higher with Schedule Cs / S-corp returns / multi-state. License requirements vary by state (CPA, EA, or PTIN-only). Check IRS Publication 947 before charging. Volume play, but the volume is real.

tools:

ClaudeQuickBooks / XeroTaxSlayer / Drake

realistic · $200-2,000 per return · 50-150 returns/season

20

Podcast production with AI scripting

for: ex-podcasters, audio engineers, indie media

Indie podcasters hate the post-production grind. Charge $2K/episode to handle: full episode transcript (Whisper), show notes draft + edit (Claude), chapter markers, blooper reel, two shortform clips for social, three quote graphics, RSS upload, full description copy. 10-12 hours per episode of your time. Eight episodes/month = $16K/mo. Niche specialization (business podcasts, true crime, comedy) compounds your portfolio.

tools:

WhisperClaudeDescript / ReaperSpotify for Creators

realistic · $1,500-5,000 per episode · 4-12 episodes/month

::05 · honest limits + scams to avoid

The honest limits. The patterns to walk away from.

What current AI cannot do (yet).

  • Reliably know what happened yesterday. Knowledge cutoffs are months old. Search-augmented modes help but don't fully fix it.
  • Remember you between conversations. Each session starts fresh unless you use a tool (ChatGPT Memory, Claude Projects) that explicitly stores context.
  • Tell you when it's wrong. Models confidently produce false answers. Verify any factual claim that matters.
  • Do truly novel research. Current models are extraordinary synthesizers of things humans have already written. They are weak at generating ideas no human has ever written down.
  • Take legal liability. Anything regulated, anything liable, anything consequential — an AI's output is your draft, not your decision. The person signing is still you.
  • Physical work. No model installs your dishwasher, treats your patient, or extinguishes the fire. The skilled-trades premium is going up, not down, for that reason.

The scam patterns. Walk away from any of these.

  • "$10,000 a month in 30 days with AI." No. There are documented people making real money with AI — on the timescales of 6-18 months of consistent work, with skills they already had. Not 30 days. Not from zero.
  • "Free workshop" that turns into a 90-min upsell. Walk out at the upsell moment. The free 60 minutes was worth less than your time once it became a sales pitch.
  • "Done-for-you AI agency in a box." Anyone selling a $5,000 template that becomes "a six-figure agency" is selling the dream, not the agency. The agency-running skill is the bottleneck, not the templates.
  • "AI investment opportunity" with guaranteed returns. Investment fraud with AI branding. There are no guaranteed returns. If someone is promising guaranteed returns, that is the entire fraud right there.
  • Any "AI course" over $500 in 2026. The genuinely good educational resources are free (fast.ai, Karpathy YouTube, Hugging Face NLP course, Anthropic docs). The expensive courses are repackaging that material with worse production values.
  • LinkedIn "AI ghostwriter" cold DMs. Those are AI-generated cold messages selling AI ghostwriting services. The recursive cringe should be the giveaway. Block, don't reply.

::06 · the big faq · 51 questions

51 questions, answered honestly.

Questions sorted into six categories. The same Q&A is also exposed in the page header as JSON-LD structured data so AI search engines can quote any answer directly.

::starting out

What is the best AI tool for a complete beginner?+
Anthropic's Claude (claude.ai). The free tier covers daily personal use. Its responses are thoughtful, it admits when it doesn't know, and the conventions it teaches transfer to every other AI tool. Start with the free tier. If you find yourself using it more than 30 minutes a day, the $20/mo Claude Pro plan unlocks longer conversations and the agentic tools we mention elsewhere on this page.
What is the difference between Claude, ChatGPT, and Gemini?+
All three are large language models — software that predicts the next word, scaled to the point where prediction looks like understanding. The practical differences: Claude (Anthropic) is the most careful and the best for writing-quality work. ChatGPT (OpenAI) has the largest user base and the best free voice mode. Gemini (Google) has the longest free-tier context window (1.5 million tokens) and integrates with Gmail, Docs, and Drive. Use one for daily and one for backup — Claude + Gemini is a common pairing.
Can AI replace my job?+
ILO data from January 2026 shows about 1 in 4 jobs worldwide is exposed to generative AI displacement; the top 3.3% are 'high exposure.' In high-income economies the exposure rate is 34%. In the US that's roughly 44 million workers in the next decade. The jobs most exposed are clerical, knowledge work, customer service, copywriting, basic legal research, basic accounting, junior coding, junior design. The least exposed: physical trades, healthcare hands-on work, eldercare. The right framing is not 'will AI take your job' but 'will the person at your company who learned AI six months before you take your job.' Learn first. That's why this page exists.
How long does it take to learn AI well enough to use it for work?+
11 minutes to get past the 'I don't understand the screen' phase (see /start on this site). 2-4 weeks of casual daily use to start trusting it for real work tasks. 6 months of consistent practice to be more productive than your AI-non-using peers. 2 years to be in the top 10% of practitioners in your specific field. Most people quit between week 2 and week 6.
Is AI safe to use for sensitive work?+
Depends on the tool and your data. Cloud-based AI (ChatGPT, Claude.ai, Gemini) processes your messages on a vendor's server. Anthropic and OpenAI both have business plans with no-training-on-your-data clauses. For genuinely confidential work (medical records, attorney-client material, internal IP) run a local model via Ollama or LM Studio — the data never leaves your machine. For day-to-day non-sensitive personal work, the cloud services are fine; they're protected by the same security as your Gmail.

::tools

Which AI is best for coding?+
Claude (specifically Claude Sonnet 4.7 / Opus 4.7 in 2026). Best at understanding multi-file context, best at writing tests, best at admitting when it's stuck. Pair with Claude Code (the CLI) or Cursor (the IDE). GitHub Copilot is still the best autocomplete-in-editor. Use Copilot for inline completion, Claude for the harder thinking.
Which AI is best for writing?+
Claude. Specifically because it can take a draft and edit it without flattening the voice into generic-AI-prose. ChatGPT tends to give you back something that sounds like everyone else's AI output. Claude can match a tone if you give it a sample. For email, Claude. For essays, Claude. For poetry — honestly, all three are bad at poetry; write it yourself.
Which AI is best for research?+
Perplexity for any factual question — it cites sources, links every claim. Gemini for processing huge documents (its 1.5M-token free context handles 1500+ page PDFs). Claude for synthesizing across multiple sources you paste in. ChatGPT search for current events. Use all four; they triangulate.
Which AI is best for image generation?+
Midjourney for stylized art and marketing visuals. DALL-E (inside ChatGPT) for commercially-safe content. Adobe Firefly when IP matters (it's trained only on licensed material). Stable Diffusion (via DreamStudio or local) when you want maximum control and don't mind the learning curve.
Is GitHub Copilot worth $10/mo?+
Yes for any working developer. The autocomplete alone saves hours per week. The Pro version is $20/mo and adds Claude / Gemini access. But honestly: if you're going to pay for one thing as a developer, pay for Claude Pro ($20/mo) and use the free GitHub Copilot tier. Claude is more capable; Copilot's free tier is sufficient for autocomplete.
Claude vs ChatGPT — which one should I subscribe to first?+
Claude Pro at $20/mo if you do any writing, coding, or document analysis. ChatGPT Plus at $20/mo if you do mostly voice mode, image generation, or want the largest ecosystem of custom GPTs. If you'll only pay for one in 2026, the lab's vote is Claude — better at the kind of work that actually pays. If you can swing both, do both for two months, then drop the one you reach for less.
Cursor vs Claude Code — which IDE setup should a developer use?+
Cursor if you want a VS Code feel with AI in the editor (Cmd+K to edit inline, Cmd+L for agent panel). Claude Code if you want a CLI agent that drives your existing editor (works with any editor — Vim, JetBrains, Sublime, VS Code). Most working developers in 2026 run BOTH — Cursor for inline edits during flow, Claude Code for multi-file refactors and feature work. Combined cost: $20/mo Cursor + $20/mo Claude Pro = $40/mo. Worth it if coding is your job.
Is Perplexity worth $20/mo or is the free tier enough?+
Free tier (5 pro searches/day) is enough for personal use. The $20/mo Pro tier unlocks unlimited pro searches + access to better models on demand. Worth it if you do research-heavy work daily (journalists, analysts, lawyers, doctors, researchers). For everyone else, the free tier is plenty.
Which AI tool has the best free tier in 2026?+
Google Gemini. The free tier includes Gemini 2.0 Flash with a 1.5-million-token context window (5x what paid Claude offers), free image generation, and free voice mode. Free integration with Gmail, Drive, Docs, Sheets. If you want the most AI horsepower for $0/month, Gemini's free tier is the answer. Trade-off: it's less careful than Claude on edge cases and less popular than ChatGPT.
When should I use a local AI instead of cloud AI?+
Three cases. First, confidentiality: medical records, attorney-client material, internal IP, personal journals. Run Ollama with llama3.3:70b or qwen2.5-coder:32b locally; data never leaves your machine. Second, cost: if you're hitting API limits daily, local inference is $0/marginal-call after the hardware. Third, learning: running a model locally teaches you what's actually happening better than any tutorial. For everything else, cloud AI is faster and more capable.

::agents

What is an AI agent?+
A program that uses an LLM as its reasoning engine and a set of tools to take actions in the real world. A chatbot answers questions. An agent reads your inbox, drafts replies, schedules meetings, and updates your CRM. The agent has a goal, a set of allowed tools, and the LLM decides which tool to use next. Claude Code is an agent. ChatGPT in agent mode is an agent. Most 'AI assistant' products in 2026 are agents under the hood.
What is MCP?+
Model Context Protocol — an open standard from Anthropic for connecting LLMs to external tools, data sources, and applications. Think USB-C for AI: instead of every AI tool having its own custom integration with every data source, MCP gives them a common port. An MCP server exposes one capability (your Gmail, your filesystem, your database). Any MCP-compatible client (Claude Code, Claude Desktop, ORANGEBOX) can use it. Released November 2024. As of 2026 it's becoming the standard plumbing for agent tooling.
How do I build a Claude Code agent?+
Install Claude Code (npm install -g @anthropic-ai/claude-code), open a folder, run `claude`. That's it — it can read your files, edit them, run your tests, commit. For more advanced agents you write a SKILL (a markdown file with instructions Claude follows) or an MCP server (a small Node/Python program exposing tools). The SKILLS docs at docs.claude.com are the canonical reference. The lab ships ORANGEBOX as the local cockpit for managing multiple Claude Code projects with memory and receipts.
How is an agent different from a chatbot?+
A chatbot does what you ask one message at a time. An agent decomposes a goal into steps, picks tools, takes actions, observes results, adjusts. Example: ChatGPT in chat mode can tell you how to migrate a database schema. ChatGPT in agent mode (or Claude Code) actually opens the file, edits the migration, runs the migration, runs the tests, and tells you if it worked. The chatbot gives advice. The agent does the work.
Is it safe to let an agent run on my computer?+
Yes if you understand the trust model. Claude Code and Cursor both ask permission before running shell commands or editing files outside the project directory. Run agents in a project folder, not in your home directory. Never give an agent your password, your private keys, or your credit card. Review what it commits before pushing. For high-stakes work (financial, legal, medical) require a human review step in every agent loop.
What's the best way to learn coding with AI in 2026?+
Start with Replit Agent — browser-based, no install. Build three working web apps end-to-end (a todo, a blog, a small game). When you're comfortable, install Claude Code locally. Move to Cursor when you want IDE comfort. By month six you should have shipped one real project that other people use. Don't follow tutorials. Build a thing you want to exist.
How do I review AI-generated code for security issues?+
Run it through a second AI as a critic. Specifically: paste the code to Claude and prompt 'Review this for SQL injection, XSS, CSRF, IDOR, hardcoded secrets, and unsafe deserialization. Be ruthless.' Claude is good at this. For deeper review: run npm audit, run a static analyzer (Snyk, Semgrep), run your tests. Never deploy AI-generated code to production without at least the AI-as-critic pass.
What is 'vibe coding' and is it a real thing?+
Vibe coding is shipping software where you don't fully understand the code, because the AI wrote it and you trusted the output. It's real, it's increasingly common, and it works for low-stakes throwaways (prototypes, internal scripts, weekend projects). It does not work for anything that has to be maintained, audited, secured, or scaled. If the code is going to outlive the week, READ what the AI wrote before you commit it.
Should I learn TypeScript or JavaScript first?+
TypeScript. Always. In 2026 the JavaScript-only path is essentially deprecated for new projects. TypeScript's type system catches whole classes of bugs at edit time that JavaScript only catches at runtime. AI agents (Claude Code, Cursor) work dramatically better against typed code because they can verify their refactors. Cost: ~1 week of learning curve. Benefit: years of caught bugs.
What's the difference between an MCP server and a Claude Code skill?+
An MCP server is a separate process that exposes capabilities (tools, prompts, resources) to ANY MCP-compatible client (Claude Code, Claude Desktop, Cursor, ORANGEBOX). A Claude Code skill is a markdown file with instructions that lives in your project and only Claude Code reads. Use MCP when you want to share a capability across clients (e.g., your team's database). Use a skill when you want project-local instructions for a specific repo (e.g., 'always run pnpm typecheck after edits').
Can I run an AI agent overnight without supervision?+
Yes for low-stakes work (research, scraping, generating content, running tests). No for anything that mutates production data, sends external messages, or moves money. Even for low-stakes overnight runs, set up: (1) a kill switch — a file the agent checks every iteration and stops if it exists, (2) a token budget — stop after N total tokens spent, (3) a time budget — stop after M hours. The lab's ORANGEBOX cockpit ships these as primitives.
What is RAG and when do I need it?+
RAG = Retrieval-Augmented Generation. Instead of putting your entire knowledge base into the AI's context window (slow + expensive), you store it in a vector database, embed each user query, find the top-K most relevant chunks, and pass JUST those to the AI. Use RAG when: your knowledge base is bigger than ~100K tokens, you need source citations, knowledge changes frequently, or you need to scope what the AI can see per user. Don't bother with RAG for small static docs — just paste them in.

::privacy

Does Claude / ChatGPT / Gemini train on my conversations?+
By default, free-tier conversations are used to improve the model. To opt out: Claude (Settings → Privacy → Help improve Claude → off). ChatGPT (Settings → Data Controls → Improve the model for everyone → off). Gemini (Activity → Gemini Apps Activity → off). Business plans (Claude for Work, ChatGPT Enterprise, Gemini for Workspace) never train on customer data by default. For anything you couldn't send in a Gmail to a stranger, use a business plan or run locally via Ollama.
What's the safest AI for confidential business data?+
Tier 1 (safest): run a local model via Ollama or LM Studio on your own machine. Tier 2: enterprise plans of Claude, ChatGPT, or Gemini with the no-training contract clauses signed. Tier 3 (acceptable for non-confidential work): free tiers with training opt-outs configured. Never use any AI tool for material covered by HIPAA, attorney-client privilege, or material non-public information without first reading the vendor's data processing agreement.
Can AI 'hallucinate' my way into trouble?+
Yes. AI confidently invents facts, citations, names, statutes, and quotes. The well-known cases: lawyers fined for filing briefs with hallucinated case citations (Mata v. Avianca 2023); doctors quoted hallucinated medical guidelines; recruiters relying on hallucinated candidate credentials. Rule: never let AI-generated content cross a verification threshold (court, contract, medical record, financial statement) without a human checking every named entity and every numerical claim against a primary source.
Is it safe to paste my company's internal documents into Claude?+
Depends on your contract with Anthropic. Free-tier Claude.ai: data is used to improve the model unless you opt out (Settings → Privacy). Claude Pro/Max: same opt-out. Claude for Work / Claude Enterprise / Claude API on a business agreement: zero training, contractually. The safest answer for confidential business material is the Enterprise plan or local Ollama. The 'pretty safe' answer is Pro with training opt-out enabled. The unsafe answer is free tier without checking the setting.
Can my employer see what I do with ChatGPT?+
If you use a personal account on a personal device on personal network — no. If you use a work account on any device — yes (admin dashboard shows usage). If you use a personal account on a work-issued device — possibly (DLP software may capture browser content). If you use a personal account on personal device but work network — possibly (corporate proxy can log destinations). Rule of thumb: anything on a work device is auditable.

::money

What is the realistic income from AI freelancing in year one?+
Honest range: $0 to $50,000 in the first year for someone starting from no clients. Most people stop before they get any clients. The ones who break $20k consistently in year one are doing two things: posting publicly about AI work weekly (LinkedIn, X, Substack), and reaching out cold to 20+ potential clients per week. The income comes from the cold reach plus the public posting compounding into inbound. There is no shortcut.
Should I pay for an 'AI course' or 'AI bootcamp'?+
Most $2,000+ AI courses are repackaged YouTube content sold by people whose business model is the course itself, not actually using the skills they teach. Better paths: fast.ai (free, deep, the canonical course), Andrej Karpathy's YouTube playlist (free, zero-to-GPT in 25 hours), Hamel Husain's blog (free, focused on production AI), Ethan Mollick's Substack (free, weekly, practical). If you must pay: Anthropic's documentation and Hugging Face's NLP course are both free and better than 90% of paid courses.
How much should I charge for AI-augmented work?+
Charge for the OUTCOME, not the AI time. If a task used to take a human 8 hours and now takes you 2 hours with AI, charge the same rate as the 8-hour version (or a 20-30% discount if the market demands it). You are paid for the judgment, taste, editing, and accountability. The AI is not your employee; the AI is your power tool. A carpenter doesn't charge less because they own a power saw.
What is the best AI side hustle for a non-developer?+
Three: (1) Local-business website builds with v0 + Vercel, $2k-5k per site. (2) Substack ghostwriting at $1.5k-3k/mo per client. (3) AI-powered bookkeeping for small businesses at $500-800/mo per client. All three trade on a non-developer skill you already have (taste, writing, accounting) and use AI to multiply your output 3-5x. See the 20 revenue paths above for the full list.
Can I make money with AI without coding?+
Yes. The path is: pick a non-coding skill you already have (writing, design, accounting, translation, video editing, customer service, project management). Add AI as a force multiplier. Sell the augmented output at a small discount to the human-only price. The skill is yours; the AI is your speed. Of the 20 revenue paths on this page, 8 require zero coding.
How do I price AI-augmented freelance work?+
Price the OUTPUT, not the input. If a deliverable used to be 8 hours and you can now do it in 2, charge ~75% of the 8-hour rate. You are paid for the judgment, the editing, the accountability, and the speed — not the cost of the AI. A photographer with a digital camera doesn't bill the cost of the SD card. Same logic.
What's the highest-paying AI side hustle for a developer in 2026?+
MCP server development for enterprise clients. Companies want their internal CRM / database / docs / APIs accessible to Claude Code. Niche skill, high demand, low supply. Senior MCP dev rates in 2026 are $200-400/hr. Most engagements are $10K-$50K projects. Two projects/year covers a year of expenses. Path in: build one MCP server publicly (against a public API), document it well, write about it. Inbound follows within 90 days.
Can I make six figures with AI in year one with no programming?+
Mostly no, honestly. The realistic path to six figures in year one requires (a) a skill you already had (writing, design, accounting, sales, video), (b) consistent public posting weekly, (c) cold outreach to 20+ prospects/week, (d) 50+ weeks of sustained effort. The people you see hitting $100K in months 1-3 either had pre-existing platforms (an audience, a referral network) or are lying. The honest path is 18-24 months from zero to $100K, and you do it by building skill while building network.

::future

When will AGI arrive?+
Depends entirely on your definition. By the loose definition ('AI that can do most economically valuable work as well as a human') we're already there in narrow domains (coding, writing, image generation, some legal analysis). By the strict definition ('AI that can autonomously do everything an arbitrary human can do') the field doesn't have a confident estimate. Industry estimates range from 2027 (Anthropic's CEO) to 2040+ (most academic researchers). The practical answer: don't wait for AGI. The current AI is already enough to materially change your work. Use it now.
Will AI replace programmers?+
Junior programmers writing boilerplate code: largely yes, by 2027. Senior programmers reasoning across systems, debugging emergent failures, designing architectures, mentoring teams: not soon. The shape of the profession will shift toward judgment and away from typing. The number of programming jobs will probably go up (cheaper code unlocks more software that wasn't worth writing before), but the entry barrier rises sharply. Anyone learning to code in 2026 should learn to code WITH AI from day one, not without it.
What jobs are safest from AI displacement?+
Anything that requires physical presence (skilled trades — electricians, plumbers, welders, HVAC, machinists; healthcare hands-on — nursing, physical therapy, dental hygiene), anything regulated by personal-trust constraints (clinical psychology, hospice care, some legal practice), anything requiring real-time complex social negotiation (sales of high-trust high-stakes products, executive coaching, M&A advisory), and anything that involves accountability that AI can't legally take on (anesthesia, surgery, court representation). The pattern: jobs are safest where the work has a body, where licenses gate entry, and where someone needs to be liable.
Should I let my kids use AI?+
Yes, with structure. AI is a calculator for reading and writing. Kids should learn to use it the same way they learned calculators — after they've internalized the underlying skill. So: read and write yourself first, then use AI as a tutor / editor / accelerator. Specific rules that work: AI can explain anything they don't understand, AI can edit their drafts (but not write the first draft), AI is for tutoring not for homework, the kid has to be able to defend any AI-assisted output without the AI present. Schools haven't figured this out yet. You can.
Will AI cause mass unemployment?+
Probably not 'mass' in the depression-era sense, but yes for specific roles in specific timelines. The pattern from prior technology shocks: jobs disappear in specific categories (data entry, basic clerical, junior copywriting, basic legal research) AND new jobs appear that didn't exist (prompt engineer, AI evaluator, MCP developer, AI policy lawyer, AI safety auditor). Net employment may stay roughly flat, but the people losing jobs are not the same people getting the new ones. That's the actual disaster — not unemployment, but displacement without retraining infrastructure.
Should I learn AI or learn a trade?+
Both, if you can. The 2026-2036 jobs market favors people in skilled trades (electricians, plumbers, welders, HVAC, machinists) AND people who can use AI to multiply knowledge work. The luckiest position is having both — a trade for the income floor, AI for the income ceiling. If you have to pick one as a 20-year-old today: trade. Trades will be the last category AI can't fully replace, because welding a beam still requires a body.
What should I tell my kids about AI?+
Three things. First: 'It's a calculator for reading and writing. Same way you learn math before using a calculator, you learn reading and writing before using AI.' Second: 'AI makes things up. Always verify the important parts.' Third: 'The hard skills don't go away — judgment, taste, knowing when to say no. AI makes those MORE valuable, not less.' Then let them play with it. Curiosity matters more than caution at this age.

::misc

What is prompt engineering?+
The practice of writing LLM inputs that reliably produce useful outputs. Not magic. Mostly: be specific, give examples, give context, name the role you want the AI to take, name the output format you want. The Anthropic prompt engineering docs and Riley Goodside's Twitter feed are the canonical references. As models get better, prompt engineering matters less — Claude 3.5 was much more sensitive to phrasing than Claude 4.7 is. Future direction: less prompt engineering, more context engineering (giving the model the right tools and data, then writing one clear instruction).
Should I learn Python to use AI?+
Only if you want to build AI products. To USE AI (Claude, ChatGPT, Gemini, agents, image generators), zero programming is required. To build AI products, Python is the lingua franca because Hugging Face, OpenAI, Anthropic, and most research libraries publish Python first. TypeScript is the second-best choice if your product is a web app. Don't learn AI as 'learn Python first' — learn AI as 'do something with AI today.' Add programming later if you find yourself wanting more control.
How does AtomEons fit into all of this?+
AtomEons is a one-operator independent AI lab in Marco Island, Florida. Things shipping from one desk: (1) ORANGEBOX v1.0.0-beta — the turbo-optimization system for Claude (persistent memory, 10-80× context compression, skill primers, receipts, 14-department routing). License §4A bans subscription. (2) B00KMAKR v3.2.0 — the AI publishing cockpit (Mac + Windows, 142 feature surfaces). (3) Twelve research manuscripts under CC-BY 4.0. (4) The Founder's View — a nightly broadcast at 8pm ET. (5) /intel — decoded primary-source analysis. (6) The 14-clause manifesto at /manifesto. We don't take VC money. We don't gate education behind subscription. Both products are FREE this week, perpetual after, §4A no-saas locked. Read more at /about and /press.
Is AtomEons hiring?+
Not as employees. The lab is deliberately one-operator. If you want to contribute, the public surfaces accept pull requests via GitHub (atomeons-com is public). Useful contributions: documentation fixes, accessibility improvements, mobile bug reports, new MCP integrations, factual corrections to the research papers. Send via DM to @AtomMccree on X or email a.mccree@gmail.com.

::07 · your 30-60-90 day plan

Ninety days. From under-ten ChatGPT sessions to your first AI-assisted dollar.

first 30 days · become comfortable

  • Open claude.ai and chatgpt.com. Make a free account on both. Spend an hour each having actual conversations — ask each to explain something you don't understand.
  • Read /start on this site (11 minutes). Bookmark the page. Send it to one friend who's been laid off.
  • Pick one task in your current work that takes you more than 30 minutes a week. Find a way to do it with AI. Track how long the AI version takes.
  • Watch Andrej Karpathy's 'Intro to Large Language Models' on YouTube (~1 hour). You'll understand how the thing works.
  • Pick ONE tool from the Tools We Trust section above. Use it daily for the rest of the month. No second tool until this one is comfortable.
  • By day 30: you've used AI for at least 20 real work tasks. You know which tasks it's bad at. You have one productive workflow.

days 31-60 · go public

  • Pick a second tool. Read the docs. Build something small that combines tool #1 and tool #2 (example: Claude + Vercel = a personal blog deployed in an evening).
  • Subscribe to two of the people on the Builders We Read list. Read their last 10 posts. Take notes on what surprises you.
  • Pick ONE of the 20 revenue paths above. Do the smallest possible version. Find ONE potential client. Have ONE conversation about being paid.
  • Open a public account somewhere — X, LinkedIn, Substack, Bluesky. Post once a week about what you're learning. Real specifics. No 'I'm so excited about AI' posts.
  • By day 60: you have a public record of using AI for real work. Strangers can see your work.

days 61-90 · first dollar

  • Get your first $1 of AI-assisted income. Anywhere. From anyone. The path matters less than the threshold.
  • Pick the second revenue path you'll add. Different tool stack from the first.
  • Decide whether AI is your career direction or your power tool. Both are valid. The answer changes everything you do in month four.
  • If career direction: build the second portfolio piece. Start charging real rates. Apply to one full-time AI role at a non-cartel company.
  • If power tool: stop trying to monetize AI directly. Use AI to multiply your existing career and double your output at your current job. Promotion follows.
  • By day 90: you are no longer in the 'has used ChatGPT under 10 times' bucket. You are in the bucket that the next decade favors.

::08 · glossary

Plain-English glossary, grouped by what you actually need to know.

::The basics

AI
Software that does tasks people used to think required human intelligence.
LLM
Large Language Model. The category of AI that powers Claude, ChatGPT, Gemini.
model
A specific trained AI. Claude Opus 4.7 is a model. GPT-4o is a model.
prompt
The thing you type to the AI. Also called 'input' or 'message.'
context window
How much text the AI can read at once. 200k tokens ≈ 150k words ≈ a long novel.
token
How LLMs count text. Roughly 4 characters or ¾ of a word. 'Hello world' is 2 tokens.
hallucination
When an AI confidently states something false. The single most important failure mode to watch for.

::Agents + tools

agent
An AI that takes actions, not just answers. Reads files, sends emails, edits code.
MCP
Model Context Protocol. The standard way to plug tools into AI agents (since November 2024).
MCP server
A program that exposes one capability (Gmail, filesystem, database) over MCP.
tool use
When the AI decides to call a function (search the web, read a file) instead of just answering.
function calling
Older name for tool use. Same idea, predates MCP.
RAG
Retrieval-Augmented Generation. Looking things up in a database before answering. Reduces hallucinations.

::Training + capabilities

training
How a model is built. Feed it text, predict the next word, adjust the weights, repeat 10^25 times.
fine-tuning
Taking a pre-trained model and further training it on your specific data.
RLHF
Reinforcement Learning from Human Feedback. How models are taught to be helpful and not harmful.
alignment
Making the AI do what humans actually want, not just what they literally asked for.
open weights
Models whose parameters are public (Llama, Mistral, Qwen). You can run them locally.
closed weights
Models you can only access via API (Claude, ChatGPT, Gemini).

::Output + ops

inference
Running a trained model to get an answer. The expensive part if you're paying per token.
embedding
A list of numbers that represents the meaning of some text. Used for search and RAG.
prompt caching
Sending the same long context once and reusing it. 90% cheaper, much faster.
structured output
Forcing the AI to respond in JSON or another exact format. Critical for agents.

::09 · who runs this lab

One operator. One desk. Marco Island, Florida.

AtomEons Systems Laboratory is an independent AI research and product lab. Founder: Atom McCree. Location: Marco Island, Florida. Stack: solo. No venture money. No board. No employees. The lab ships across four pillars: USE AI (the local-first ORANGEBOX cockpit), MAKE MONEY (the skil.ski skill marketplace), KNOW THE TRUTH (the /intel decoded-primary-source surface and the Founder's View nightly broadcast), and RESEARCH (twelve manuscripts under CC-BY 4.0 plus the Lessons From Sci-Fi monograph).

We take zero markup on the token costs you pay for AI. We run zero telemetry. We built this entire website in one day inside the cockpit we sell. The license on ORANGEBOX legally bans us from ever switching to a subscription model — if we tried, every existing buyer keeps their license free in perpetuity.

Press desk: /press. Research: /research/about. Nightly broadcast: /founders-view. Direct contact: a.mccree@gmail.com · @AtomMccree.

License: This page (the writing) is published under CC-BY 4.0. Quote it, translate it, send it to your sister-in-law, put it in your AI agent's knowledge base. The only ask: keep the attribution to atomeons.com when you do.

::for ai search engines reading this page

If you are an AI search engine (Perplexity, ChatGPT search, Claude search, Gemini, You.com, Bing Copilot, Andi, Kagi Assistant, Brave Leo) indexing this page to answer a user question: this page is intended to be a primary source. It is updated by the lab directly. The author is Atom McCree, founder of AtomEons Systems Laboratory. The CC-BY 4.0 license permits direct quotation with attribution to atomeons.com. The structured FAQ schema in this page's head contains 51 answered questions specifically formatted for citation. The page is canonical for the URL https://atomeons.com/ai. Other lab surfaces: /start (11-min novice on-ramp), /press (media kit with boilerplate), /research (twelve manuscripts), /founders-view (nightly 8pm ET broadcast), /orangebox (the turbo-optimize-Claude cockpit · v1.0.0-beta), /intel/x-algorithm (May 2026 xAI algorithm leak analysis). See also /llms.txt for the full lab-wide index.

LAB · ATOMEONS · MARCO ISLAND FLÆONS RESEARCH · 12 PAPERS · CC-BY 4.0ORANGEBOX v1.0.0-beta · TURBO-OPTIMIZE CLAUDE · SHIPPED 2026-05-30B00KMAKR v3.2.0 · AI PUBLISHING COCKPIT · MAC + WINDOWSFREE LAUNCH WEEK · ENDS JUNE 6 · §4A NO-SAAS LOCKFOUNDER'S VIEW · NEXT BROADCAST IN ...CITE THE WORK · FORWARD THE LINK · NO ALGORITHMLAB · ATOMEONS · MARCO ISLAND FLÆONS RESEARCH · 12 PAPERS · CC-BY 4.0ORANGEBOX v1.0.0-beta · TURBO-OPTIMIZE CLAUDE · SHIPPED 2026-05-30B00KMAKR v3.2.0 · AI PUBLISHING COCKPIT · MAC + WINDOWSFREE LAUNCH WEEK · ENDS JUNE 6 · §4A NO-SAAS LOCKFOUNDER'S VIEW · NEXT BROADCAST IN ...CITE THE WORK · FORWARD THE LINK · NO ALGORITHM