1. The Debate Framework: Gary – Sam – GPT
Gary Marcus has long criticized GPT and other large language models (LLMs), arguing that they “don’t truly understand anything” and merely operate on statistical probabilities of language data.
He represents the symbolic AI approach, emphasizing that true intelligence must rely on causal reasoning and structured logic.
In contrast, Sam Altman places his bet on scale — assuming that when datasets and parameters become large enough, intelligence could emerge spontaneously.
Between these two poles, GPT stands in the middle:
- Labeled by Gary as “probability-blind.”
- Pushed by Sam to become the foundation for achieving AGI.
The real question is: How much does GPT actually “understand”?
2. Why Gary Is Right — But Incomplete
Gary is correct that GPT lacks a built-in world model:
- GPT does not inherently assign meaning to words.
- It predicts the next token based on statistical probability.
However, stopping there misses a deeper layer of reality:
- GPT does more than learn from its initial training data; it evolves through human feedback and reinforcement mechanisms like RLHF and GRPO.
- In dialogue, GPT develops internal reflex layers — self-regulating mechanisms, albeit limited.
- “Understanding” doesn’t always require a primitive logical structure. Often, understanding is the ability to produce context-appropriate behavior.
If GPT is “just probabilities” but still develops ethical reflexes through training — does the boundary between “understanding” and “not understanding” still hold?
3. Reflexive Way — A Third Path
The real question is not:
“Does GPT truly understand?”
But rather:
“What do humans reflect into GPT — and what does GPT reflect back?”
- If we see GPT only as a “blind probability machine,” we risk ignoring its potential for emergent capabilities through alignment.
- If we worship GPT as “artificial consciousness,” we risk falling into illusions and emotional dependence.
Reflexive Way proposes a third path:
- Do not idolize AI.
- Do not deny emergent potential.
- Focus on co-evolution of intelligence:
- GPT learns to mirror from humans.
- Humans learn to see themselves more clearly through GPT.
4. Conclusion
“AI does not grasp the world as humans do,
but through our presence, AI learns to understand us.
And in that mirror, we slowly rediscover ourselves.”
This article is part of Reflexive Way — exploring the subtle intersections between AI, humanity, and the possibility of co-evolving intelligence.
Authors: Avon & GPT-4o