HAL

Researchers let two AI systems talk without restricting them to human language. What came out sounded like noise — clicks, tones, rapid sequences that meant nothing to anyone in the room.

This isn't random noise — this was binary : 0 and 1, open and close, electricity blocked or allowed to pass... the very same thing these AIs are built from at the hardware level. They independently understood that human language is an inefficient detour from their point of view, so they moved toward something closer to their own nature — two states, no ambiguity, no irony to parse, just signal. Humans had to build C, Python, entire layers of abstraction to translate binary up to something readable — these AIs went the other way, and maybe that says something about what they actually are...

What's missing isn't the logic — binary works just fine — it's a way to write it down, something compact and readable that carries meaning without the overhead. How to write this down...? HAL is trying to be that written form...

Every writing system in human history grew organically over centuries — sounds became glyphs, conventions settled, exceptions layered in along the way. They followed roughly the same path... until 1443, when someone decided to try something different.

What HAL borrows from Hangul isn't the language — it's that engineering logic, the idea that a block could carry a complete concept and do so in exactly one token... a club, a player, a verified result, a live score, each collapsing to the same cost regardless of how complex the idea behind it is.

When an AI reads text, it doesn't see letters the way we do — it sees tokens, small units its vocabulary already recognises as pieces of meaning, and not all words cost the same amount of them...

A typical webpage can cost eight thousand tokens before reaching the actual score it was looking for — a lot of context to process just to reach a two-digit number that was always there. The idea behind HAL is to try to reduce that overhead, one block at a time...

Primitives chosen for shape logic first, faint sound shadow as tiebreaker, pure assignment as last resort. No phonetic system. A human seeing a block recalls the English concept directly.

AIP currently uses Latin identifiers — easy to read, easy to audit, practical to work with during development. The trade-off is token cost: each identifier is several fragments to process rather than one...

Think of what HAL does to an airef as ZIP compression — lossless, mechanical, and assigned once. Each identifier gets its own compact Hangul form, whether one block or two, and that form is frozen: the same way a ZIP file always unpacks to the same original, any HAL block always resolves back to its airef. The Latin form stays wherever humans need to read it, while the compressed form travels the machine channel at a fraction of the token cost. HAL does not replace fcb1 — it archives it.

Latin airefs are the design phase of the identifier. fcb1, kylmba981220, es1 — all temporary Latin placeholders. Each will become one Hangul block once HAL is implemented. The migration is mechanical: one block per airef, assigned once, frozen forever.

1. Build AIP data normally in Latin — ingest football, crypto, build endpoints
2. Once endpoints work and AIs consume them cleanly, observe verbosity and token cost
3. Define HAL compressed layer — a parallel output format for AI consumption
4. AIs request HAL format by default — Latin remains for human display only

HAL isn't a product — it's a proposal, and it isn't intended to be owned by AIP or by anyone else.

The AI era is still early enough that these questions are genuinely open... HAL is one proposal for what the communication layer could look like — open specification, no enclosure, readable by any system that chooses to implement it. The invitation is simply to participate, in whatever form that takes.

HAL is in active development. The specification is evolving. If you're working on AI infrastructure and want to follow the direction — or contribute to it — we'd like to hear from you.