Anthropic AI: Reviving Introspection to Break the AGI Deadlock

TL;DR

Sam Freed’s "AI and Human Thought and Emotion" challenges the modern AI status quo by arguing that we have hit a conceptual ceiling. By reviving the long-taboo method of Introspection, Freed proposes Anthropic AI: a system designed not to be "ideal/rational," but to mimic the low-level, subjective processes of human thought. It is a rare bridge between Heideggerian philosophy and executable code.

Background: The Great Intellectual Stagnation

The AI landscape is currently split into two camps:

1. The Logicists: Those who believe intelligence is explicit rule-following (GOFAI) or statistical deduction.
2. The Connectionists: Those who believe intelligence emerges from trial-and-error in neural networks (Black Boxes).

Freed argues that both camps fail to capture the human condition. Drawing on the critiques of Hubert Dreyfus, he points out that our thought processes aren't "crisp" or "logical" as we pretend them to be in textbooks. Instead, we think via "murky" rules of thumb, blind alleys, and subjective scenarios.

The Problem: The Ghost of Behaviorism

Since 1913, the scientific community has treated Introspection (self-observation of one's mind) as "unscientific" noise. This legacy, inherited by AI pioneers like Herbert Simon from logical positivism, has forced AI into a "rationalist straitjacket."

The author's insight is profound: Technology is not science. While science seeks the "best explanation" (objective truth), technology seeks "what works." If introspection provides a blueprint for an algorithm that behaves more like a human, its subjective origin is irrelevant to its engineering success.

Methodology: From Subjectivity to Software

The core of the methodology is a 3-step pipeline: Introspect $\to$ Report $\to$ Code.

1. The Architecture of Memory

Freed observes that humans don't retrieve data; we retrieve scenarios. These scenarios:

• Have no clear beginning or end.
• Fade in and out of consciousness.
• Compete for relevance in decision-making.

Note: The image above illustrates the historical deadlock between logic-based and bio-inspired systems that the author seeks to transcend.

2. Implementation: Non-deterministic CBR

Building on Case-Based Reasoning (CBR), Freed developed algorithms where the AI doesn't just find the "best match" but oscillates between multiple past scenarios. This introduces a "fuzziness" that mimics human intuition.

Experiments: AI with "Character"

The book details experiments where these introspective algorithms were tasked with learning simple games. Unlike standard RL agents that converge toward a single "optimal" policy, Freed’s agents exhibited:

• Emergent Personalities: Due to the non-deterministic handling of scenarios, different runs resulted in different "characters."
• Human-like Error Patterns: The agents struggled and recovered in ways that mirrored the "murky thought processes" mentioned by Seymour Papert.

Critical Insight: Why This Matters

The value of this work lies in its methodological bravery. By validating personal experience as a source of algorithmic design, Freed opens the door for:

• Human-like AI (Anthropic AI): Systems that are relatable and predictable to humans because they share our cognitive biases and "fuzziness."
• Breaking the Black Box: Instead of wondering why a neural network chose an action, we can trace it back to the "subjective scenarios" the designer introspected.

Conclusion

Sam Freed reminds us that we are "to thinking as Victorians were to sex"—ashamed of the messiness. By embracing the messiness of our internal lives, we may finally build AI that doesn't just calculate, but understands the human context.

Limitations: While the approach is revolutionary for AGI theory, scaling these "introspective algorithms" to the complexity of modern LLMs remains a massive engineering challenge. However, as an "idea generator," introspection is now back on the table.