How to Become a Super Individual: Taste Is the Ticket, Tokens Are the Fuel, Brainpower Is the Ceiling
How to Become a Super Individual: Taste Is the Ticket, Tokens Are the Fuel, Brainpower Is the Ceiling
In the first half of 2026, "super individual" and "one-person company" became the two most repeated phrases in Chinese tech circles.
Shanghai Lingang launched a "Super Individual 288 Action." Suzhou hosted the first AI One-Person Company Conference. State media ran the story twice in two months. At Anthropic's Code with Claude conference, Dario Amodei was asked when the first billion-dollar company with a single human employee would appear. His answer: 2026, with 70–80% confidence.
The story is sexy. But the story is the product. The structure is the engineering.
Strip away the narrative and look at what actually decides whether you can pull this off — three deeply unequal resources I call the three pillars:
Pillar 1: Taste — The Real Scarcity of the AI Era
A counterintuitive observation first.
In early 2026, Wix paid $80M cash to acquire Base44 — a company with a single developer, Maor Shlomo, who hit 250K users in 6 months and cleared $189K in profit in a single month after LLM token costs. Pieter Levels runs Nomad List, Remote OK, and PhotoAI solo, generating $3M+/year, with zero employees.
These cases keep proving the same thing: when AI can generate infinite code, design, copy, and strategy, the scarcity of "making" collapses to zero. The scarcity of "selecting" goes through the roof.
Selecting what? Selecting whether the AI's output is usable. Selecting whether the product shape is right. Selecting which feature gets cut. Selecting which line of copy sounds too "AI-ish."
That's taste.
Taste is a fast, low-cost judgment function. You don't need to explain why. You look once, you know.
In product terms, taste decides three things:
1. Product shape. "AI writes your daily standup" could be a Chrome extension, a Notion integration, a Slack bot, a standalone SaaS, or a CLI tool. Five shapes, five user segments, five retention curves. Which one fits the audience you actually want? That's taste.
2. AI output triage. Cursor / Claude Code throw 2000 lines of code at you daily. Your job isn't to write code — it's to be a chief editor for 2000 lines. What stays, what dies, what gets rewritten. Engineering taste converts directly into product quality.
3. Injecting "humanity." AI-generated copy has a distinctive uniform quality — everything is correct, nothing is memorable. Andrés Max's line has been quoted all year: "Taste is the new moat." When everyone can generate content with AI, the only thing that makes a product memorable is the personality embedded in it. AI cannot do this for you, because it has no personality.
The brutal part: there's no shortcut to taste. It accrues with time. You have to have shipped hundreds of failed products, used thousands of apps, edited tens of thousands of lines of someone else's code before you can judge "is this good?" in half a second. This happens to be where veteran engineers crush 22-year-old prodigies.
Pillar 2: Tokens — Not About Saving, About Ruthless Compression
The second pillar is most underrated because it looks like a money problem.
It isn't. Tokens are a resource-utilization art form.
A typical super individual's monthly AI spend (2026 median, based on Fortune / industry data):
- Claude Code Max / Codex Pro: $200 × 2 = $400
- Cursor / Windsurf: $40
- Research tier (Perplexity etc.): $20
- Direct API calls (user data, batch jobs): $300–$800
- Total: $760–$1260/month
Sounds fine. But that's the spend of someone who knows what they're doing. Someone who doesn't could easily burn $5000–$10000 in the API-call line — with worse output.
The gap comes from three things:
Cache hit rate. Anthropic's prompt caching gives 90% off on prefixes over 1024 tokens. A 50K-token codebase context costs $0.15/call cached vs $1.50/call uncached — 10x difference. Knowing how to write code doesn't mean knowing how to structure prompts. Putting stable content first and variable content last is basic but routinely ignored.
Model routing. A single task can be split across three models: Haiku for intent ($1/M), Sonnet for the main generation ($3/M), Opus only for critical review nodes ($15/M). Cheap models for simple tasks, expensive ones for complex tasks — table stakes. A deeper layer: high-volume tasks (batch data cleaning) should run on local Qwen / DeepSeek — zero token cost. A cleaning task that runs 100K times a year saves enough money to buy an M4 Max if you migrate it off Sonnet.
Distillation and reuse. Running Opus repeatedly for the same review is wasteful. Smart play: let Opus run once, distill its judgment into a prompt or fine-tune, then let Haiku handle 95% of requests downstream. You're solidifying the expensive model's judgment so a cheap model can execute.
Why does this matter? Because the super individual's business model is "dollars of revenue per token spent."
A SaaS at $50/month, 1M tokens per user in service cost. At $3/M, your margin is 94%. Optimize to $0.5/M and it's 99%. But pipe every request straight to Opus at $15/M — you lose $15 per customer.
I've seen too many great-engineering products with broken token economics. Using AI to write code isn't the skill. Doing AI unit economics is.
Pillar 3: Brainpower — The Real Ceiling, And No Moore's Law
The first two pillars are learnable. The third is a hardware constraint.
The human brain has roughly 4–6 hours per day of deep-focus capacity. Every context switch, per Gloria Mark's classic UC Irvine research, takes about 23 minutes to recover prior focus depth.
Now put that number in a super individual's workday:
- 9:00 — review Claude Code PR (window 1)
- 9:30 — customer ticket (window 2, 23 min loss)
- 10:30 — Xiaohongshu comments (window 3, 23 min loss)
- 11:00 — product design (window 4, 23 min loss)
- 12:00 — production bug (window 5, 23 min loss)
- 14:00 — customer call (window 6, 23 min loss)
Switching cost alone eats nearly 2 hours out of a 6-hour deep day. What's left, you still have to spend on AI output triage, taste judgments, and strategic decisions.
And here's the cruelty: AI throughput grows exponentially while human cognitive bandwidth is flat.
In 2023, GPT-4 could draft ~10 tasks per hour for you. In 2026, Claude Opus 4 + Code can run 200+ tasks per hour (parallel agents). Your brain is still the same brain — still ~8–10 meaningful judgments per hour.
This is the buried truth in the super-individual story: you're not blocked by the tech. You're blocked by yourself.
Scheduling itself burns energy. Deciding "which agent output do I look at next" is decision fatigue — a finite daily budget that, once spent, leads to a cliff in quality. This is why many super-individual products show quality drift around month 3 — the product isn't tapping out, the founder is.
Three responses, none of them comfortable:
1. Cut workflows, don't add tools. The number of projects a super individual can run simultaneously is not limited by code output — it's limited by how many independent mental models you can hold. My observed healthy ceiling is 2 core projects + 1 maintenance project. More than that, things start to break.
2. Let AI handle scheduling itself. Don't decide which agent output deserves attention — have a meta-agent score and rank. The top 3 things come to you. Everything else is handled. This is one of 2026's most underrated workflows: outsource attention allocation too.
3. Physically isolate deep time. 9am–noon: no notifications, no comments, no email, one thing. Sounds like 2015 productivity advice, but in the AI era the value is higher — because you are the only non-parallelizable resource in the system.
The Three Pillars Are Unequal
If you can only invest in one pillar, which?
My answer: build taste first, optimize tokens second, protect brainpower third.
Because the learning curves are radically uneven:
- Taste takes 5–10 years of accumulation, the earlier the better, cannot be outsourced
- Token optimization is a 1–2 week engineering problem, other people's playbooks transfer
- Brainpower management is a lifelong topic, edges differ per person, you have to find yours
But invert the question. Which pillar collapses you first? Always brainpower. The first two are slow-growing bottlenecks. Brainpower can fail tonight. That's why I put it last — it's the endgame problem.
A Counterintuitive Conclusion
"Super individual" evokes a heroic image: one person versus an entire team, AI-augmented, soloing a $10B market.
The real life of a super individual is more boring, and more sober.
The core isn't "how much can I do with AI." It's "how much can I refuse to do."
- Cut every workflow that doesn't convert directly to dollars
- Cut every project that requires more than 3 mental models
- Cut every feature whose token unit cost exceeds its revenue unit cost
- Cut every loop that requires real-time human judgment
What's left is what one person can actually do.
AI raised the ceiling on human capability by 10x. But the ceiling on human energy is the same number it always was.
Whoever sees this clearly survives year three.
If you're seriously considering this path, do three things before you touch any tool:
- Find a vertical you've been judging for 5+ years — taste must be in place
- Pull last month's token usage and find the 80% that comes from one call — token economics must be clear
- Draw a 3-month cognitive budget — how many deep hours per week, over which you are overdrawing — draw the brainpower line
Then pick a product.
Skip the order, and the better you get with AI tools, the harder you fall.