AI progress incrementalism

By citing a CS paper that LLM persuasive ability plateaus with more training data, the post supports the view that recent AI advances produce bounded, incremental gains rather than sudden, qualitatively new manipulation powers.

Several links (Claude’s 1M context, world-model demos vs. TextQuests showing AIs can’t beat long games, METR’s usability gaps) portray steady usefulness gains without decisive capability leaps, aligning with the incremental-progress framing.

Author describes GPT-5 as reducing hallucinations and improving utility more than raw capability jumps.

Items like Microsoft’s CLIO boosting GPT‑4.1 biomedical QA from 8.55% to 22.37% and notes that GPT‑5 advanced without massive pretraining scale-ups reflect notable reliability/usefulness gains driven by techniques and steering rather than headline compute leaps.

The piece frames GPT-5’s advances as a 'boiling frog'—steady, reliability/usability gains masked by comparing only to SOTA—matching the idea that post‑ChatGPT progress arrives as notable but non‑dramatic improvements that still matter for policy expectations.

Rather than headline-grabbing AGI leaps, the article highlights practical productivity tools (bots drafting slides, summarizing interviews, checking argument logic) being embedded at McKinsey and in legal workflows—incremental capability gains that materially change work processes.

The guests argue that nearly three years into the hype cycle, AI remains far from AGI and is still searching for a killer app, directly reinforcing the theme that recent advances are notable but incremental rather than transformative.