Remember that recent story about the senior student from Beijing University of Posts and Telecommunications, born in the 2000s, who got $30 million in funding from Chen Tianqiao within 24 hours? Their team @evermind just dropped another bombshell project: MSA (Memory Sparse Attention). In simple terms: Compared to large models = super smart but with goldfish memory, MSA = actually remembers things, evolving from open-book exams to photographic memory. Here’s the project link: https://(github.com)/EverMind-AI/MSA Although the project isn’t open-sourced yet, I’ve already thought of a potential use case. Feel free to discuss and share your thoughts. Currently, most quant systems rely on backtesting historical candlestick charts and indicators. Recently, I’ve noticed @CryptoPainter and @Meta8Mate working on “pattern-based trading systems” using historical candlestick patterns. The biggest issue with these backtests is that they rely solely on technical indicators without incorporating the fundamental context or news events of the historical cycle, which makes them somewhat incomplete. But now, with MSA’s powerful memory capabilities, I think achieving true “learning from the past” is entirely possible! 100M token context = Reading all S&P 500 earnings reports from the past 15 years + every Federal Reserve meeting transcript + the complete history of the crypto market in one go. Candlestick patterns + technical indicator strategies + news events and fundamentals. I think it’s entirely possible to truly emulate human-like decision-making: recognizing a familiar pattern and recalling the fundamental events or breaking news that occurred at that time, leading to the most accurate judgment! https://(x.com)/elliotchen100/status/2034479369855590660?s=46&t=akNrvwH7J8CfQVwvmQ4AfQ