AI agent benchmarks are broken(ddkang.substack.com)
167 points by neehao | 79 comments
First screen, captured .
Looks similar
Related stories
Publish fake stats, get cited in LLMs(peec.ai)
315 points by robtherobber | 204 comments
Peec AI blog post by a GEO researcher describing how publishing statistics led to the author's data being widely cited across the web and by LLMs.
LLMs use a surprisingly simple mechanism to retrieve some stored knowledge(news.mit.edu)
336 points by CharlesW | 121 comments
MIT News article on research showing LLMs often retrieve stored facts via simple linear functions, and a technique to probe what a model knows about subjects.
Large-Scale Online Deanonymization with LLMs(simonlermen.substack.com)
327 points by DalasNoin | 233 comments
Substack post by Simon Lermen summarizing research showing LLM agents can deanonymize users from anonymous posts across Hacker News, Reddit, LinkedIn and interviews at scale.
2025: The Year in LLMs(simonwillison.net)
823 points by simonw | 451 comments
Simon Willison's annual blog post reviewing major LLM trends of 2025, including reasoning models, coding agents, Chinese open weight models, and Gemini.
LLMs work best when the user defines their acceptance criteria first(blog.katanaquant.com)
411 points by dnw | 368 comments
Substack essay from Vagabond Research arguing that LLMs produce plausible rather than correct code, using a simple database primary key lookup test as an example.
Why I'm still bearish on LLMs after Navier-Stokes(dank.systems)
426 points by jaykru | 552 comments
Personal blog essay by Jay Kruer arguing LLMs remain overhyped despite headline results like Navier-Stokes, citing poor generalization, reward hacking, and specification costs.