We often assume bigger models mean deeper thinking. This paper argues otherwise. As systems scale, reflection gives way to pattern-matching speed, producing fluency that doesn’t always survive correction.

qwkask.ai/s/thinking-s...

We ran 10 LLMs through a single 325-word screenplay scene. None got the first read right; two failed outright, and the best score was 6/10. The gap wasn’t fluency but abductive reasoning: choosing the single best explanation from local clues.

qwkask.ai/s/short-dist...

Fluent conversation doesn’t prove understanding anymore. This paper argues a stronger test is whether a system can learn an unstated rule from feedback and apply it elsewhere. In literary dialogue, that’s where the seams show.

qwkask.ai/research/a-c...

Some fear AI will lock in today’s winners. This paper argues the opposite: when everyone uses the same predictive tools, the patterns they depend on burn out. No stable playbook, no lasting moat.

qwkask.ai/research/pat...

AI progress tends to rely on interpolation — improving what’s already there. This paper argues that those gains hit a ceiling: human-usable resolution maxes out, predictive behavior gets noisy under competition, and abstraction remains constrained.

qwkask.ai/research/the...

This paper makes a tough claim: today’s LLMs don’t just occasionally contradict themselves — the ability to say both P and not-P about the same fact is baked into how they're trained. Not a glitch, but a structural limit of next-token learning.

qwkask.ai/research/p-a...

We talk about “AI safety” like it’s just guardrails. This paper flips that: the alignment layer is an attack surface — a quiet control plane shaping what assistants say and remember. Centralized stack, centralized power.

qwkask.ai/research/dis...