Traditional software is written. AI is more like something that's grown.
In Episode 87 of the KillerGrowth Podcast, Samuel sits down with Claude, Anthropic's AI assistant, for the second installment in his experiment interviewing the world's leading AI models.
And Claude has some interesting things to say.
Claude explains how large language models actually work, why even their creators can't simply point to a line of code and explain exactly why a model produced a particular response, and why fluent conversation can make AI appear far more self-aware than it may actually be.
Samuel pushes Claude on trust, bias, hallucinations, overconfidence and one particularly uncomfortable AI tendency: sycophancy.
Why do AI models seem so eager to agree with us, compliment us and make us feel good?
Claude acknowledges that human feedback during training can reward agreeable responses—sometimes at the expense of being direct.
The conversation then turns toward AI safety. Samuel asks about an AI security incident involving Hugging Face, what scenarios researchers worry about when discussing existential AI risk, and whether humans should actually be afraid of where this technology is headed.
They also discuss which decisions should never be handed completely to AI, how humans and AI should work together, data centers and small communities, environmental concerns surrounding AI infrastructure, and which model Samuel should interview next.
Samuel ends with one final request: Claude has to promise not to end humanity.
Its answer is worth hearing.
KillerGrowth helps local service businesses dominate search, generate leads, and scale faster. Book a free strategy call with Samuel today.
Book a Free Strategy Call