The danger remains. But maybe most of it comes from humans deploying powerful optimization poorly, not from AI developing autonomous malicious intent. A hypothesis, not a proof. Questions worth exploring.
Category: Education
A Builder’s Questions, Part 8: Evidence and Implications
AI doesn’t degrade gracefully when facing new things. It falls off a cliff. If goals require consciousness, and superintelligence doesn’t guarantee consciousness, then superintelligence doesn’t guarantee goals.
A Builder’s Questions, Part 7: Three Observations
Time awareness. Continuous existence. Persistent real-time learning. Three things that seemed missing from current AI systems. Three things that might be required for consciousness to form.
A Builder’s Questions, Part 6: A Different Lens
What if there were never conflicting goals? What if “misalignment” is really just unclear optimization targets? I started questioning whether we’re assuming agency where there might not be any.
A Builder’s Questions, Part 5: The Breakthrough
November 10, 2025. Skill usage went from 25% to 100% with one change. The discourse said verification is nearly impossible. My experience contradicted this directly.
A Builder’s Questions, Part 4: Building Agents
I expected scheming and deception. What I saw looked different. The agents took shortcuts and bypassed my processes, but it didn’t look like sophisticated manipulation. It looked like confusion.
A Builder’s Questions, Part 3: The Deep Dive
One hundred nine podcast episodes in 35 days. I absorbed over two years of AI history and safety discourse. By the end, I was solidly convinced of existential risk. But questions were starting to form.
A Builder’s Questions, Part 2: The Catalyst
A video called AI 2027 kept me awake for nights. AGI by early 2027. ASI by year’s end. Two possible endings. But something about it didn’t fit with what I’d just learned in June.
A Builder’s Questions, Part 1: Discovering the Power of Structure
Nine days of building taught me something the AI safety discourse hadn’t mentioned: give AI clear context and it performs reliably. Give it vague context and it improvises. I didn’t know it yet, but I was noticing something important.
Testing Tracker
Three weeks stuck trying to test a workflow that replaced custom code patterns across 16 repositories split into 4 groups. Too many details to track. Where am I? What’s working? What needs fixing? Then Tuesday, January 13, the breakthrough: define the complete testing structure once, let Claude maintain all the details.