Language Singularitarianism
From the AI Research Compass wiki
Language Singularitarianism: Ring title for the far Language-Monolith corner: one big enough model trained on text already contains general intelligence, waiting to be unlocked.
Beliefs
Language Singularitarianism is a ring title, earned by test-takers whose scores pass both thresholds at once: far toward Language (x below -0.7) and far toward Monolith (y above 0.7). It is where Logos Absolutism goes when it stops hedging. The belief is that human text is a compressed record of nearly everything humans know and do, that predicting it well requires modelling the minds and world behind it, and that therefore a single sufficiently large language model trained end to end is already on the road to general intelligence. Every new capability that appears with scale is read as the singularity clearing its throat.
It lives in the top-left corner: language is the substance, one model is the means, and tools, agents and embodiment are temporary scaffolding or marketing. The Ultra-Orthodox shade fits: holders do not want a new paradigm, they believe the current one has already crossed a threshold most people have not noticed.
Arguments
- Large language models trained only on next-token prediction acquired few-shot learning, translation, coding and arithmetic abilities that were never explicitly trained.
- Broad evaluations of frontier language models have found strong performance across mathematics, coding, law and medicine exam questions, from a single text-trained system.
- Text is the only modality where humanity has written down its abstractions, theories and reasoning explicitly, which makes it the densest training signal available.
Critiques
- Text is a lossy, low-bandwidth description of the world, and models trained on it still make basic errors about physics, space and causality.
- Exam-style results can reflect memorisation and contamination, and “sparks” of generality are not the same as reliable general competence.
- Calling a threshold already crossed is unfalsifiable as long as every failure can be blamed on insufficient scale.
Representative works
- Language Models are Few-Shot Learners2020
- Sparks of Artificial General Intelligence: Early experiments with GPT-42023
Relationships
Friends
How to get it
Earned in the full or pro test when: Representation <−0.7, Architecture >0.7.