GoodBarber engineering lead J. explains that agent memory is a liability and recommends tracking state rather than history, citing a Google and Purdue preprint (SKILL.state) that reports up to 94 perc…
Qwen 3.8 27B, running locally on a 16GB RAM MacBook Pro, outperformed OpenAI's GPT 5.6 Luna Max on the DABstep benchmark at over 17 times lower cost, with electricity costs under $0.50 versus over $8.…
Qwen 3.8 27B, an open-weight model, outperformed OpenAI's GPT 5.6 Luna Max on the DABstep benchmark while running locally on a laptop, costing under $0.50 in electricity versus over $8 for Luna Max, a…
A June 2026 study by LightSpeed and Tencent found that four frontier models—Claude Sonnet 4.5, DeepSeek-V4-Pro, GLM-5, and Gemini-3-Flash—misjudged their total context size with median relative error …
OpenAI's price cut for GPT 5.6 Luna by 80% has made AI-powered analytics answers cost less than half a penny each, with Luna achieving 99.8% accuracy on an agentic SQL benchmark at a 5x lower price th…
A study of OLMo-3 checkpoints by Arav Dhoot, supervised by Yixiong Hao and Zephaniah Roe, finds that chain-of-thought faithfulness to hints varies non-monotonically across training stages. The pretrai…
An arXiv paper submitted on 15 Jun 2026 introduces SearchGEO, a framework to measure how LLM-based web-search agents can be manipulated by malicious web content. Evaluating 13 LLM backends on 308 case…