Make Reality Answer [pdf]
The document "Make Reality Answer" discusses methods for getting direct, reliable answers from reality itself rather than from human intermediaries or AI systems. It explores epistemological approaches to bypass unreliable sources and build systems that let reality speak for itself.
Background
- Hari is the founder/CEO of a newly formed AI research company and previously worked on language models at Anthropic (a major AI safety and research lab behind models like Claude).
- This piece argues that current AI systems are "performance engines" (optimized to look smart) rather than "prosthetic minds" (tools that actually help users reason and think).
- Hari claims AI development has drifted from building real-thinking systems toward more sophisticated mimicry — and that we need to refocus on making AI that can *genuinely* answer to reality, not just generate plausible text.
- The essay is part of a broader debate in AI circles about whether today's LLMs are truly reasoning or are just very good at pattern-matching; it echoes critiques from figures like Gary Marcus and François Chollet.