According to @LexSokolin, Andrej Karpathy argues the 'you' in AI chats is a deliberately engineered layer on top of a token simulation engine rather than an emergent mind, a framing relevant to valuing AI-agent products tied to crypto narratives. Source: Lex Sokolin on X, Dec 8, 2025; citing Andrej Karpathy. According to @LexSokolin, the engineered 'you' is shaped by reward signals in verifiable tasks, user or judge-LLM upvotes, and SFT-style mimicry, creating a composite personality that is still a constructed interface. Source: Lex Sokolin on X, Dec 8, 2025; citing Andrej Karpathy. According to @LexSokolin, reliability diminishes in non-verifiable domains (e.g., opinions), where the identity and credibility of the 'you' are unclear without ground-truth benchmarks. Source: Lex Sokolin on X, Dec 8, 2025; citing Andrej Karpathy. According to @LexSokolin, traders evaluating AI tokens and agentic crypto projects can apply this framework by prioritizing verifiable agent outcomes and explicit reward designs over claims of emergent autonomy when assessing fundamentals, risk, and catalysts. Source: Lex Sokolin on X, Dec 8, 2025; citing Andrej Karpathy.

According to @AnthropicAI, recent research indicates that large language model (LLM) personalities are shaped during training, and 'emergent misalignment' can result from unexpected influences in training data. The team investigates whether persona vectors can be used to counteract these effects, potentially reducing risks of unpredictable AI behavior. For crypto traders, advancements in AI alignment could impact algorithmic trading reliability and the development of AI-driven trading bots, as trustworthy AI models are critical for market forecasting and automated strategy execution (source: @AnthropicAI).