Does the machine understand, or just mimic better?

In 1966, Joseph Weizenbaum's secretary at MIT asked to be left alone in the room so she could keep talking to ELIZA — a program that did nothing but reassemble canned patterns into replies. Sixty years later you can ask Kimi K3 (released July 2026) about the philosophy of mind and get back something that reads like understanding. The question has not moved: does the machine understand, or has it simply become much better at imitating understanding?

This article walks through 70 years of AI — ELIZA, the Perceptron, backpropagation, word2vec, the Transformer, RLHF, Grok 4.5 and Kimi K3 — not to celebrate progress, but to make the part that did not change visible. If you run a small business in Vietnam and you are weighing up an AI agent, that boundary is not theory. It decides how far you trust the system and where you keep a person in the loop.

ELIZA (1966): the illusion of understanding starts here

ELIZA was never designed as a chatbot. It was a research platform for studying how people interpret language (Shrager, 2024). Users still assigned it feelings, intent, even friendship. Weizenbaum spent the rest of his life warning about it: when a machine says "I understand you", we believe it because we feel understood, not because the machine understands anything.

That response — the ELIZA effect — has not worn off. A CHI 2024 study (N = 2,165) found that an AI using first-person pronouns raised perceived accuracy and lowered perceived risk (Cohn et al., DOI: 10.1145/3613905.3650818). The effect was stronger among users with lower financial literacy — exactly the group most exposed when AI is deployed without any evaluation frame.

What this means for a Vietnamese SMB: do not judge an AI by the feeling that it understands you. Judge it by its error rate on a specific task, plus a mandatory verification step.

The Perceptron and the AI winter: not failure, a correction

Minsky and Papert (1969) did not kill neural networks. They showed that a single-layer perceptron cannot compute XOR or connectedness. But the book Perceptrons created a chilling effect that lasted 20 years at DARPA (Bottou, 2017). The AI winter was really a correction: it forced the field toward backpropagation (Rumelhart, Hinton, Williams, Nature 1986).

You can read today as a Perceptron moment for large language models. The unsolved limits — grounding, genuine understanding — may force a similar correction. Not a collapse, but a demand for more transparency.

From word2vec to the Transformer: the scale changed, the question did not

word2vec (Mikolov et al., 2013) learned word vectors in a day over 1.6 billion words. The Transformer (Vaswani et al., 2017) replaced recurrence with pure self-attention. RLHF (Ouyang et al., NeurIPS 2022) made a 1.3B InstructGPT model preferred over GPT-3 at 175B in 85% of comparisons — progress in alignment, not in understanding.

Grok 4.5 (xAI, July 2026) scores 64.7% on SWE Bench Pro with a 500,000-token context window. Kimi K3 (Moonshot, July 2026) carries 2.8T total parameters with roughly 104B active, 896 experts, and a 1M-token context. Two architectural details matter and are still unverified: KDA (Kimi Delta Attention) and Stable LatentMoE have no independent technical report yet, and the claimed scaling gain of 2.5x over K2 is internal and has not been reproduced.

The practical conclusion: scale and speed changed. The underlying pattern matching over symbols did not.

"Understanding" is still open — which is why SMBs need a trust frame

The latest philosophical work (Trends in Cognitive Sciences, 2026) sorts understanding into four families: model-based, capacity-based, causal, and phenomenological. None of them shows LLMs reaching understanding as grounding — linking symbols to physical reality (Harnad, 1990; reassessed in 2026 in Philosophy & Technology, DOI: 10.1007/s13347-026-01079-4).

For a business running agents, the useful question is not "does the AI understand?" but "on task X, with error rate Y, how far can I trust it?" That is the five-question frame before any rollout: define the task, set the acceptable error threshold, keep a human in the loop, add a verification mechanism, and close the feedback loop.

G-Company OS at 5ac is built on exactly that principle: AI agents for Vietnamese SMBs — control instead of blind trust.

Takeaway

Seventy years from ELIZA to Kimi K3, the progress is real. What is still unresolved — understanding, grounding, architectural transparency — has not moved. For a small business owner in Vietnam, the lesson is not to refuse AI. It is to refuse trusting it without a control frame.

References

• Weizenbaum (1966), "ELIZA", Communications of the ACM — UMBC
• Shrager (2024), "ELIZA Reinterpreted", arXiv:2406.17650
• Cohn et al. (2024), CHI 2024, DOI: 10.1145/3613905.3650818
• Minsky & Papert (1969), Perceptrons, MIT Press
• Rumelhart, Hinton, Williams (1986), Nature 323 — DOI: 10.1038/323533a0
• Mikolov et al. (2013), arXiv:1301.3781
• Vaswani et al. (2017), NIPS 2017 — arXiv:1706.03762
• Ouyang et al. (2022), NeurIPS 2022 — arXiv:2203.02155
• Grok 4.5 — x.ai/news/grok-4-5; docs.x.ai
• Kimi K3 — github.com/moonshotai/Kimi-K3; kie.ai/blog/kimi-k3
• Shumailov et al. (2024), Nature, DOI: 10.1038/s41586-024-07566-y
• Trends in Cognitive Sciences (2026), DOI: 10.1016/j.tics.2026.04.003