## Background Large language models (LLMs) debuted with GPT-3 back in June 2020. After roughly five to six years of development, I believe the technology is still in its infancy, with massive room for improvement. A major priority is building powerful models that run on fewer resources, especially capable models that can operate smoothly on CPUs.
More crucially, the engineering ecosystem around LLMs is also at an early stage. Many unresolved challenges remain. Among them, *hallucinations* have become a major bottleneck, greatly limiting LLMs' adoption in real-world production scenarios.
I’ve come to a thought: relying purely on natural language to interact with LLMs — the so-called prompt engineering — may have been a wrong direction from the very start.
Fundamentally, an LLM is a mathematical computation system that takes linguistic tokens as its basic computing units. Problems like hallucinations and alignment issues are not flaws inherent to the models themselves. Instead, they arise from our improper usage patterns.
We input plain text and expect accurate, high-quality responses. We feed large chunks of text and expect the model to fully understand and remember all the information. This direct text-input-text-output workflow, in my opinion, is fundamentally flawed.
## Ontology as an Intermediate Layer I propose inserting a transitional layer between humans and LLMs: a human-defined semantic space that strictly maps to the real world. All conversational semantics will be converted via this layer, ensuring every element involved in the LLM’s computation is authentic and reliable.
This layer can be built by introducing *ontology* into AI agents. While many teams are already working on this field, most of them only build ontologies for narrow, isolated domains. Every party has to develop its own domain-specific ontology, which also explains the extremely high operating costs of companies like Palantir.
## The Radical Idea We invest enormous resources, measured in trillions, to train large models. Why not spend a fraction of that cost to abstract the entire real world into a unified semantic space based on ontology and knowledge graphs?
## Final Thoughts Artificial intelligence has long had three major schools of thought: *symbolism*, *connectionism*, and *behaviorism*.
OpenAI’s ChatGPT has proven that scaling up is the key to unlocking the full potential of connectionism. Could it be that symbolism and behaviorism are also stuck in stagnation simply because they have never been scaled to a comparable level?
If we apply the brute-force scaling approach to symbolism and behaviorism on a massive scale, will we also see disruptive qualitative leaps?
Last but not least: why not integrate symbolism (ontology), connectionism (LLMs) and behaviorism (reinforcement learning) together via AI agents? I believe this combination is the true path forward for artificial intelligence.