AI in 2025: gestalt

G
gleech.org
gleech.org Saved Wednesday, December 17, 2025 Readwise
Read Original Article
7
Medium
Write Soon

This article strongly resonates and has clear angles for your perspective

Quick Take

This is a comprehensive technical deep-dive into 2025 AI capabilities and safety that directly intersects Brian's interests in AI agents and future of software work. While dense and academic, it contains specific insights about reasoning models, agent capabilities, and practical deployment challenges that could inform his AI integration work and side projects.

Relevant Domains

AI/agents/future of software work Engineering craft/architecture/productivity (secondary - deployment insights) Side projects/automation/earning from skills (secondary - capability predictions)

Blog Angles

1

"Why AI 'Reasoning' Models Might Be Expensive Theater"

Thesis

Your Hook

2

"The AI Agent Reality Check: What Actually Works in 2025"

Thesis

Your Hook

3

"Building on Quicksand: Why AI Evals Don't Matter for Real Work"

Thesis

Your Hook

Key Quotes

We improved on some things they were explicitly optimised for (coding, vision, OCR, benchmarks), and did not hugely improve on everything else
Between 1 and 7% of all work hours are currently assisted by generative AI...1.4% time savings
The experienced devs with <50 hours of practice using AI showed productivity decrease
It's probably not because [pretraining] wouldn't work; it was just ~30 times more efficient to do post-training instead
Most benchmarks are weak predictors of even the rank order of models' capabilities

Tags

#ai-reasoning-models #agent-capabilities #ai-benchmarks #engineering-productivity #ai-costs #automation-reality #ai-scaling-limits