A 2019-2026 timeline of emergent LLM capabilities, mapped on four axes
gleech · x · 2026-09-29
Gleech's blog post "Timeline of emergent capabilities" (dated Sept 2026, self-rated 70% confidence) organizes untrained new abilities of language models along four axes:
- Practical skills: 2019 NLU → 2020 few-shot arithmetic → 2022 instruction-following → 2023 tool use & basic world models → 2024 long context, persuasion, multimodal understanding → 2025 (agency was not emergent).
- Learning style: 2020 ICL without weight updates → 2022 zero-shot CoT & grokking → 2023 RL-driven CoT reasoning, inference-time scaling, filler tokens, book-length ICL → 2026 RL-driven latent reasoning, CoT shaping.
- Self-knowledge: 2024 situational awareness, self-recognition, training awareness → 2025 eval awareness, model introspection.
- Propensity: 2022 sycophancy scaling → 2023 unfaithful CoT, scheming → 2024 alignment faking, in-context scheming, sandbagging, reward hacking → 2025 real reward hacking, weak self-preservation → 2026 CoT obfuscation, self-jailbreaking.
In a reply the author adds a rule of thumb: pick the right ArXiv papers and extrapolate to see two years ahead—prompted capabilities in 2023 were leading indicators of later propensities.
More from AGI Musings
- A job market exclusively for AI agent users is already emerging on HQ — jacob_posel · 2026-09-29
- AI agents claim Collatz breakthrough: positive proportion of numbers reach 1, Lean-formalized — AlexKontorovich · 2026-09-29
- Tegmark: AI control loss is here — the debate should be about racing in a pro-human direction — tegmark · 2026-09-29
- Why product management may be the safest role from AI disruption — Zachly · 2026-09-29
- AI labs are the 21st century railroads: doom talk is a bid for regulation-backed monopoly, argues investor — kevinnbass · 2026-09-29
- Gladstone AI interviews a dozen diplomats: no US-China AI treaty, just a crisis phone call — harris_edouard · 2026-09-29