Turns out, you don’t need to be human to master metalinguistic analysis—LLMs do it better, and now generate entire artificial languages on demand.
We always said analyzing language itself was a ‘human-only’ superpower. LLMs just called our bluff. Berkeley’s Gašper Beguš dropped this bomb at OpenAI’s forum: modern models don’t mimic speech—they dissect syntax like linguists[1].
In breakthrough case studies, Beguš showed LLMs with metalinguistic capabilities, generating complex sentences and testing language limits. Paired with Project CETI’s whale comms decoding (sperm whale ‘phonemes’ mirror human speech), it flips animal-human comms research. But the killer app? ConlangCrafter: an LLM system auto-generating coherent, diverse artificial languages for games, novels, or evo studies[1].
Devs, this is gold for procedural gen: craft lore languages for RPGs, simulate dialect evolution for chat agents, or probe LLM internals for novel insights. Beyond creativity, it tests ‘what’s uniquely human’ in your prompting workflows[1].
Vs. prior tools like fragmented conlang generators, ConlangCrafter is typologically diverse and coherent out-of-box, leveraging LLM scale. Stacks with CETI’s bio-AI bridges, pushing interpretability frontiers OpenAI’s chasing[1].
Fire up ConlangCrafter (details via Beguš’s lab), prompt a Elvish dialect for your game, and watch it evolve grammars. Does this mean AIs ‘speak’ more humanly than us—or redefine intelligence entirely?
Source: Berkeley Linguistics