Ep 27: AI & Human Alignment With AI Solopreneur Yanqing Cheng

<h1>Shownotes</h1><p><b>Are AIs Learning Humanity's Better Nature? with Yanqing Cheng</b></p><p>Anne Currie talks with software engineering leader and <a rel="noopener noreferrer nofollow" href="http://Tollens.ai" target="_blank">Tollens.ai</a> founder Yanqing Cheng about how large language models acquired the helpful assistant persona, and what their training tells us about human learning, morality, and alignment. Yanqing traces the assistant from early role-played prompts through reinforcement learning and synthetic post-training, while Anne asks whether books and long-form writing carry more of humanity's better nature than the short-form internet.</p><p>The conversation moves from AI training to parenting, cautionary stories, Chinese AI research, recursive improvement, regulation, and the risks created by both malice and ordinary negligence. They finish with a cautiously hopeful question: if today's systems are a form of baby AGI, can we deliberately raise them to help solve humanity's problems?</p><p><b>Key Topics</b></p><ul><li>How early language models role-played the helpful assistant persona into existence</li><li>The difference between a base model, RLHF, reasoning training, and later synthetic post-training</li><li>Whether books contain a more reflective version of humanity than short-form social media</li><li>What raising a child suggests about learning, alignment, judgment, and the desire to be good</li><li>Why fairy tales and science fiction let people update their mental models using fictional evidence</li><li>The limits of text-only training when reality contains far denser information</li><li>How open research and shared techniques are accelerating Chinese AI development</li><li>Recursive improvement, international regulation, open weights, accidents, and deliberate misuse</li><li>Why Anne and Yanqing remain hopeful even while acknowledging that optimism may be wrong</li></ul><p><b>Action Items</b></p><ul><li>Distinguish the behaviour of a base model from the behaviour created by post-training and the surrounding environment.</li><li>Use fiction and cautionary stories to explore failure modes before equivalent systems exist in the real world.</li><li>Pair optimistic experimentation with monitoring, feedback, and a willingness to learn when an idea fails.</li><li>Treat international cooperation as part of AI safety rather than assuming that one country can regulate the problem alone.</li></ul><p></p><p><a rel="noopener noreferrer nofollow"