LessWrong (Curated & Popular) “AI in 2025: gestalt” by technicalities
Dec 8, 2025
This discussion dives into the current landscape of AI, projecting its capabilities for 2025. It highlights improvements in specific tasks, yet notes a lack of generalization in broader applications. The conversation contrasts arguments for and against the anticipated growth, including concerns about evaluation reliability and safety trends. A look at emerging alignment strategies and governance challenges adds depth, while pondering the future of LLMs amidst evolving models and metrics. Intriguing questions linger about the real implications for AI safety.
AI Snips
Chapters
Transcript
Episode notes
Impressive But Not More Useful
- Compared to last year, models are much more impressive but not correspondingly more useful.
- Progress mainly reflects bringing more tasks into distribution rather than far generalization.
Post‑Training Over Pretraining
- Labs prioritized post-training (RL and fine-tuning) over huge pretraining runs for cost-effectiveness.
- This made apparent capabilities jagged and obscured true frontier ceilings.
Better Metrics, Clearer Signals
- New composite metrics (ECI, ADEL, CAST) give more AGI‑like signals than single benchmarks.
- These measures show 2025 keeping up a fast rate of capability growth despite noisy benchmarks.
