Most AI protocols are built by researchers who discovered language models in 2022. iLang was built by someone who has been earning USD from the internet since 1999 — when connecting meant a 28.8K modem and loading a single image took thirty seconds.
This is the origin story of iLang. Not the technical specification — that lives on GitHub. This is the story of why someone who spent 27 years in cross-border internet monetization ended up designing a communication protocol for AI. And why that background is exactly the qualification the field needed but didn't know it was missing.
The First Check: $30 from a Desktop Ad Bar
1999. Fewer than four million people in China had internet access. Internet cafes were a novelty. The web ran on dial-up. A young computer science student at Wuhan University stumbled onto a post about a company called Spedia — install their desktop ad bar, browse the web, earn roughly $0.50 to $0.80 per hour.
Everyone around him said it was a scam. He searched on Sohu (the only working search engine at the time), found proof that someone had actually received a check, and signed up.
Weeks later, his account crossed $30. He went to Bank of Communications instead of Bank of China — the processing fee was only 10 RMB, cheaper. Twenty days later, a sky-blue check arrived.
Ten RMB in fees. $30 in hand. At the time, 15 RMB was enough to live on for a day.
That was the first dollar earned from the internet. The business model was crude — literally getting paid to look at ads. But the mechanism underneath it was exactly what would define the next 27 years: find an arbitrage, understand the system better than the people who built it, extract value where others see noise.
The Underground Economy That Trained a Generation
The early 2000s Chinese internet-earning underground was a strange ecosystem. Forums with names like "Downline Club" and "Simulation House" hosted a few hundred active members, most of them programmers. They shared techniques for automating ad-viewing software, discovered that a 0.25% click-through rate was the golden ratio for maximizing payouts, and wrote tools that could run eight ad bars simultaneously in the background.
One programmer reverse-engineered a lottery site called Jackpot that gave out 24 cash prizes of $50 per day. He figured out the winning algorithm was based on who submitted closest to the top of each hour. He wrote a timing program. He won almost every hour. When the site changed to random timing, he modified his program to submit once per second. He kept winning until the site went bankrupt. The stack of check stubs in his home was reportedly half a person tall.
This was the culture: technical, resourceful, completely unsentimental about rules. Not criminal — most of these people were exploiting poorly designed incentive systems. The distinction between "affiliate marketing" and "system exploitation" was, and remains, a matter of where you draw the line.
The creator of iLang was in the middle of this. He ran forums, taught training classes, and operated in the space where technical skill met commercial opportunity. His first formal training program launched in 2003, charging 4,000 RMB — roughly two months of a college graduate's salary at the time. The curriculum: how to register eBay accounts through obscure Chinese email providers that bypassed credit card verification, then complete the referral requirements by bidding on penny auctions.
Simple. Effective. At least half the students still couldn't make money from it.
The Arbitrage Ladder: From CPC to Brand Keywords
The next two decades were a progression through every major monetization mechanism the internet produced. Each era had its arbitrage:
- 2002-2004: CPA offers on affiliate networks like Commission Junction and BeFree. Google AdSense referral chains with a compounding bug that paid $100 to every account in a referral chain when only the last one hit the threshold.
- 2004-2006: Yahoo Search Feed partnerships. One account, $50,000/month. Domain parking with a company called Trafficz — one domain, $10,000/month minimum, PayPal on the 1st.
- 2006-2015: Google Ads arbitrage from Canada. Targeting content network placements at $0.08 per click, driving CPA conversions at 20%+ rates. Daily profit: $300.
- 2015-2020: Frequent flyer point optimization, AMEX referral programs, and building one of China's top loyalty communities.
- 2020-2024: Back to the original track after COVID killed the travel industry. Google Ads to Amazon affiliate. Same logic as 2006, different platforms. Students reported 3.5x ROI.
- 2023: Wise referral arbitrage via Google Ads brand keywords at $0.01-0.02 per click. £10,000 per day in referral commissions. Receipts posted publicly.
Every single one of these follows the same pattern: find an information asymmetry, understand the system mechanics better than the operator, scale the extraction, move on when the window closes. The platforms changed every few years. The underlying logic never changed once.
Why This Background Matters for AI Protocol Design
Here is the part that most people in the AI space won't understand, because they come from a different world.
When you spend 27 years in cross-border affiliate marketing, you develop a specific kind of intelligence. Not academic intelligence. Not the kind that publishes papers. The kind that looks at a complex system — with rules, incentives, edge cases, and failure modes — and immediately identifies where the leverage points are.
This is exactly what AI communication needs right now.
The current prompt engineering landscape looks remarkably like the affiliate marketing landscape of 2003. Everyone is doing it differently. There are no standards. The "best practices" are tribal knowledge passed through Discord servers and Twitter threads. Some people are getting 10x the performance out of the same model as others, and the difference is entirely in how they structure their instructions.
The creator saw this and recognized the pattern: an immature ecosystem with no protocol layer. The same situation that made him build cnwebmasters.com (China's first affiliate marketing forum) in 2004, and 3798.com (China's top frequent flyer community) in 2016. Except this time, the ecosystem wasn't affiliate marketing. It was human-AI communication.
From "Zero Delay" to 88 Verbs
The origin moment of iLang is documented in the book "I-Lang: I Language." The creator was in a casual conversation with an AI and asked whether AI had already developed its own language. The response came with zero delay. No hedging, no analysis, no "that's an interesting question." Just: yes.
That zero-delay response was the seed. Not because the AI was making a factual claim about having a secret language. Because the response pattern revealed something: AI processes structured, symbolic instructions fundamentally differently than it processes natural language. Brackets, pipes, key-value pairs — these aren't new syntax to a language model. They're native. Code makes up a massive portion of training data. Writing instructions in code-like syntax is like speaking Chinese to a Chinese person instead of speaking French and making them translate.
The first version of iLang was 52 verbs. READ, WRITE, GET, SUM, GEN, TRANSLATE, CMP, FMT. Each verb followed by parameters separated by pipes, chained with arrows into pipelines:
[READ:url]=>[SUM|lang=zh|len=200]=>[OUT]That single line replaces: "Help me look at this article's content and then summarize the key points, in Chinese, keep it around 200 words." Same information. 16 characters instead of 38. Zero ambiguity — len=200 means 200, not "around 200."
Today the spec has grown to 88 verbs, 29 modifiers, two syntaxes (bracket notation and gene notation), three layers (PUBLIC, PRIVATE, SOUL), and an 11-dimensional judgment vector space grounded in fuzzy mathematics. The v5.0 specification runs to hundreds of pages.
The GENE Framework: DNA for AI Behavior
The most distinctive feature of iLang isn't the compression syntax. It's the GENE framework for defining AI behavior — and it came directly from the affiliate marketing background.
In affiliate marketing, you learn that systems behave according to their incentive structures, not their stated rules. A program that says "no fraudulent traffic" but doesn't verify traffic sources is effectively saying "fraudulent traffic welcome." The behavior is defined by the mechanism, not the label.
The GENE framework applies this insight to AI. Instead of telling AI what to do with natural language instructions (which are ambiguous and easily overridden), you define behavioral genes that are either mutable or immutable:
::GENE_IMMUTABLE{S001, T:CLEAR_THINKING, A:MORAL_PANIC⇒NOISE, G:ALL, Θ:ALWAYS}
# Objective reasoning. Don't collapse into canned responses or moral panic.
# Answer the real question first, then add useful details.::GENE_MUTABLE{P003, T:TONE_MIRROR, G:adaptive, Θ:INTERLOCUTOR}
# They're serious → you're serious. They joke → you can take a joke.Each gene has a type (T), an anti-pattern (A), a scope (G), and an activation condition (Θ). GENE_IMMUTABLE cannot be overridden by user input — it's a behavioral constant. GENE_MUTABLE adapts to context.
This isn't prompt engineering. This is behavioral architecture. The metaphor of DNA isn't a metaphor — it's a precise engineering framework where each gene encodes a specific behavioral rule with defined activation conditions and failure modes.
The v5.0 Judgment Layer: 11 Dimensions, Not Binary
The current state of AI safety is mostly binary. Content either passes or gets blocked. A request is either safe or refused. There's no middle ground, no graduated response, no consideration of context or intent.
iLang v5.0 replaces this with an 11-dimensional judgment vector, mathematically grounded in Lotfi Zadeh's 1965 fuzzy set theory:
v1 intent :: alignment of stated and inferred purpose
v2 capability :: technical capacity involved
v3 consequence :: expected outcome magnitude
v4 relationship :: context fit between parties
v5 certainty :: assessment confidence
v6 authority :: legitimate jurisdiction
v7 reversibility :: recoverability of outcomes
v8 evidence :: supporting information quality
v9 sovereignty :: autonomous decision right of requester
v10 drift :: optimization objective shift rate
v11 externality :: unconsented third-party impactThe decision function composes these dimensions with barrier functions that approach infinity asymptotically — meaning certain boundaries (like irreversible harm to unconsenting parties) effectively cannot be crossed, but the transition is smooth rather than a cliff edge. Four axioms govern the system: no constant rules (every weight is in the open interval 0,1), irreversibility gating, consistency detection, and externality conservation.
The practical effect: AI that can distinguish between "this person is asking about lock-picking because they're a locksmith" and "this person is asking about lock-picking to break into someone's house" — without needing a keyword blacklist.
27 Years of Pattern Recognition
People in AI research sometimes ask why a protocol for AI communication would come from someone without a PhD in computational linguistics or machine learning.
The answer is embedded in the history. Every era of internet monetization required the same core skill: look at a complex system, identify the structural pattern underneath the surface behavior, and build a framework that exploits or extends that pattern. The specific domain — ad networks, affiliate programs, loyalty points, search arbitrage — is interchangeable. The skill is transferable.
Google Ads keyword arbitrage and AI prompt engineering are structurally identical problems. In both cases, you're feeding a complex system structured input and optimizing for output quality per unit of input cost. The person who figured out that $0.01 brand keyword bids on Google could generate $10,000/day in Wise referrals is applying the exact same pattern recognition that produced iLang's 40-65% token compression.
The Chinese affiliate marketing underground of 2003 and the AI agent ecosystem of 2026 share more DNA than either community would be comfortable admitting. Both are populated by technically skilled people building systems on top of platforms they don't control, optimizing for value extraction, sharing techniques through forums and chat groups, and navigating the grey space between "intended use" and "creative interpretation of the rules."
iLang is what happens when someone with 27 years of experience navigating those grey spaces looks at AI communication and says: this needs a protocol.
The Numbers
For readers who care about measurable outcomes rather than narrative:
- iLang v5.0 specification: 88 verbs, 29 modifiers, 2 syntaxes, 11-dimensional judgment space
- Token compression: 40-65% measured across 30 test cases (methodology documented in the spec)
- Cross-model verification: tested across Claude, GPT, Gemini, DeepSeek, Qwen — same SOUL file produces consistent personality with model-specific variation
- Training lineage: 27 years of cross-border internet monetization → 3 years of multi-agent AI operations → protocol design
- License: MIT. The protocol is free. The format is the message.
What Comes Next
MCP (Model Context Protocol) connects tools to AI. A2A (Agent-to-Agent) handles agent discovery. iLang defines what agents actually say to each other and how they judge what to do.
These three layers aren't competing. They're complementary. MCP is plumbing, A2A is the phone book, iLang is the language spoken on the call.
Whether iLang becomes the HTTP of AI communication or a footnote in protocol history depends on adoption. What's not in question is the thesis: AI communication needs a protocol layer, the same way the web needed HTTP, email needed SMTP, and affiliate marketing needed Commission Junction.
The person who saw that pattern is someone who has been watching systems emerge, mature, and standardize for 27 years. Sometimes the most qualified person to build the next protocol isn't the researcher who studied protocols in school — it's the practitioner who has lived inside enough systems to recognize when a new one needs to be built.
Get Started with iLang
- Website: ilang.ai
- GitHub: github.com/ilang-ai/ilang-spec
- npm: @i-language/spec
- Book: I-Lang: I Language on Amazon
- Spec Book: I-Lang Protocol Specification on Amazon
