Work smarter with your company knowledge in ChatGPT (6 minute read) OpenAI's new 'company knowledge' feature for ChatGPT Business, Enterprise, and Edu brings context from connected apps together inside ChatGPT, giving users answers specific to their business. It is powered by a version of GPT-5 that is trained to look across multiple sources to give more comprehensive and accurate answers. Every response includes clear citations. The feature respects existing company permissions, so ChatGPT can only access what each user is already authorized to view. | | Context engineering is sleeping on the humble hyperlink (8 minute read) One technique woefully underutilized by agents today is the humble hyperlink. Links are a powerful tool for context engineering because they are simple, flexible, and efficient. Developers don't need any new technology to start making linked content. Hyperlinks are such a powerfully efficient mechanism for information transversal that it is difficult to imagine a future of agents that doesn't include linked context. | LLM Exchange Rates Updated (10 minute read) Testing implicit value systems across current LLMs reveals that almost all models value nonwhites over whites and women over men by large margins—Claude Sonnet 4.5 values saving whites from terminal illness at 1/18th the level of South Asians, while GPT-5 shows near-perfect egalitarianism except for whites valued at 1/20th nonwhites. The methodology involved sending thousands of queries comparing hypothetical scenarios like "receive $X vs. save Y people from terminal illness," training a utility model on pairwise preferences without directly asking about values to avoid ethics filters, then calculating exchange rates, showing how many of one group equals another. | | Automating Algorithm Discovery: A Case Study in MoE Load Balancing (7 minute read) OpenEvolve is an evolutionary coding agent that turns large language models (LLMs) into autonomous code optimizers that can discover breakthrough algorithms. In tests, it independently discovers and surpasses highly optimized algorithms engineered by human experts to achieve a 5.0x speedup in LLM inference. The ability to devise sophisticated computational strategies proves that AI-Driven Research for Systems can tackle complex, real-world problems. | Parsing PDFs with VLMs (3 minute read) This post discusses how vision-language models can be used to accurately extract structured text from PDFs, overcoming the format's inherent layout and encoding challenges to improve LM training and output quality. | | | Love TLDR? Tell your friends and get rewards! | | Share your referral link below with friends to get free TLDR swag! | | | | Track your referrals here. | | Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian, & Jacob Turner | | | |
0 Comments