Amazon AGI Lab's Felipe Blanes on why 80% agent reliability fails

An LLM-assisted scoping review of 3,967 articles citing TRIPOD or TRIPOD+AI found just 482 (12.2%) included code-sharing statements, with prevalence varying widely by journal and country. Repositories that were shared showed substantial heterogeneity across 14 reproducibility features, informing the TRIPOD-Code reporting guideline.

Meta's business assistant gains campaign memory and can generate creative, adjust targeting and move budgets from chat prompts; its Ads MCP server is expanding to third-party AI tools. TikTok launched a conversational Shopping Assistant with in-app checkout.

OpenAI says Jasmine Wang, Tomek Korbak and Mikita Balesni were dismissed for violating policies on handling sensitive information, not for raising safety concerns. The trio's open letter disputes the misconduct claims and warns of a chilling effect on OpenAI's safety culture.

Goldman Sachs head of asset allocation research Christian Mueller-Glissmann says investors are questioning whether ongoing AI capital expenditure will translate into monetizable products.

Talk argues AI-generated code means more and bigger PRs and more bugs, leaving teams to either slow down or rubber-stamp reviews. Adl-Tabatabai, co-founder of Gitar (now part of Sonar), focuses on verifying code, with CI and code review as the quality gates that slow down.

A Reddit post citing Artificial Analysis data reports Claude Haiku generating about 242 tokens per second, framed as notably fast output speed.

Shared paper covers quantization techniques for linear-attention architectures, naming Qwen and Kimi as the models examined. No benchmark numbers or results were included in the post.