Social Media Briefing — February 23, 2026

Emollick highlights a paper showing that weaker LLM judges cannot properly evaluate smarter models, arguing benchmarks should be viewed as triplets of dataset+model+judge, and judges are becoming the saturated bottleneck

Emollick argues people systematically overestimate the speed of corporate AI adoption and underestimate the limiting effect of AI's jagged abilities, noting companies have significant inertia

Emollick argues that AI companies' failure to articulate non-ominous visions of the future (even Dario Amodei's 'Machines of Loving Grace' fails to describe what life would actually be like) will be a problem for the next phase of AI adoption

Emollick argues that historical backlash to industrial revolutions wasn't Luddite destruction but regulation, redistribution, unions, and safety nets - and similar responses are likely for AI

levelsio argues that $100/mo Claude Code replaces low-to-mid-level developers who are slower and more expensive, leaving only top-tier devs and those who can effectively lead AI as valuable

Ethan Mollick highlights a research paper showing that weaker LLM judges cannot properly evaluate stronger models, arguing that benchmarks should be understood as triplets of dataset+model+judge, with judges becoming the saturated bottleneck.

Emollick explains that a large diverse pool of humans has skills that cancel out individual weaknesses, but 1000 identical AI agents share the same blindspots and may be more vulnerable to groupthink-like problems

Emollick argues that AI's ability to understand video/images is largely underexplored commercially, with many economically valuable real-time monitoring applications remaining untapped despite current limitations

Burkov argues RL-trained coding LLMs are incentivized to write spaghetti code with security holes because it's theoretically impossible to generate a reward signal for code being vulnerability-free, so models optimize solely for producing expected outputs