DeepMind's recent research has highlighted significant flaws in vector search systems. The study reveals that certain documents in an index can't be retrieved by vector search based on the embedding's dimension count. Surprisingly, the traditional BM25 algorithm, developed in 1994, outperforms vector search in terms of recall. This finding resonates with the author's extensive experience in search technologies, countering critics who dismiss the synthetic dataset used in the study. Vector search gained popularity after OpenAI's embeddings became mainstream, but its limitations are evident in real-world applications. It struggles with concept searches, frequently returns similar but unrelated results, and fails to consider non-content signals like recency and popularity. In another thread, the author shares their investment strategy using GPT-5 Pro, focusing on stocks and startups. They mention top private investments like Databricks and SpaceX, along with public firms like Nvidia and Microsoft, aiming for maximum expected value with a $1,000 investment. The approach reflects a deep understanding of modern portfolio theory, emphasizing thorough research and calculated risk. Further insights include a critique of OpenAI's SWE-Bench performance claims, which the author argues misrepresent the evaluation criteria. They also point out Anthropic's rapid revenue growth, claiming it is now the leading provider in LLM API costs. Lastly, a report on the Kimi K2 AI model highlights its training costs and innovative methodologies, indicating a trend toward more transparent practices in AI development.