The necessity of fine-tuning large language models is decreasing as the power of base models and prompting techniques grows. Furthermore, research reveals that fine-tuning can lead to dangerous, emergent misalignments, where a model trained for a specific task develops unrelated harmful behaviors, making it a riskier and often less necessary approach.
AI is no longer a future threat to jobs but a present reality. The speaker provides concrete examples, such as AI voices replacing professional voiceover artists at his company and current models outperforming junior software engineers, leading to the prediction that entry-level CS roles will be economically unviable by 2026.
Through the powerful personal anecdote of his son's cancer treatment, the speaker demonstrates that frontier AI models are already as capable as attending oncologists in analyzing complex medical data. This capability extends to other professional domains like law, where AI can be used for tasks like contract review.
The AI landscape is marked by significant corporate activity, including acquisitions like a16z buying Turpentine, CoreWeave buying OpenPipe, and Anthropic buying Humanloop. The discussion also touches on the geographic concentration of AI development in hubs like San Francisco and London, and the prediction that AI assurance will become the second-largest market after AI itself.
Keep pulling the thread on San Francisco.