In a guest blogpost, Neo4j’s Alyson Welch explains why Large Language Model AI systems can’t move beyond non-trivial applications until they are properly curated.
ChatGPT and other generative AIs have captured both public and increasingly corporate interest, but they still have limitations in an enterprise context. From ‘hallucinations,’ misinformation, bias and a lack of explainability, so far this use of Artificial Intelligence (AI) in the context of “Large Language Models” is more of a curiosity than anything more substantial.
There’s also fear about the ‘threat’ LLMs represent, with some tech leaders calling on AI researchers to slow down production so risks can be studied in “a stepping back from the dangerous race to ever-larger unpredictable black-box models with emergent capabilities”. Elon Musk, who is planning to introduce his own generative AI, claims he’s worried that ChatGPT is inherently biased.
There’s a debate to be had here. But whatever else needs to happen, generative AI needs to do a better job of delivering true visibility to build trust. But to lay the foundations for that, we need to be clear about what ChatGPT and its LLM text producers aren’t. They are not sentient and cannot replace the necessity of you and I writing articles, essays or brochure copy entirely, although they may assist in initial drafting tasks.
And as the technology has assimilated numerous examples of computer code, it’s also a valuable resource for improving developer efficiency. But beyond making it easier to facilitate code and copy generation, to fully harness the power of generative AI we need to train its next generation on high-quality, structured business data, rather than relying on ‘what’s out there’ on the internet for free.
And the optimal way to do this, in our view, is via a graph-based Knowledge Graph as support and interface. This approach is the best way to achieve explainability, compliance, and reproducibility, as well as giving system builders the many other benefits graph-based Knowledge Graph and GDS (graph data science) brings.
In fact, companies are increasingly beginning to adopt this approach. To take one example, an oil and gas company in Singapore tried an LLM for its companywide searches, but as the AI didn’t know what it was looking at the results were of limited value. However, using graphs to provide context, the firm was able to increase the accuracy and relevance of its AI.
There’s no reason why you couldn’t do the same: you could amass a significant volume of text data, either externally (pertaining to your specific market) or internally (such as product catalogues).
Radically cut down on ChatGPT errors
Using the new synthesised LLM, you can create a Knowledge Graph that will help make sense out of the accumulated data accelerating R&D, or refining compliance procedures. And you can employ this process in reverse by applying a smart language model to a problem space you have encoded in a graph. This allows you to control the input to the model, resulting in a responsive, easy-to-interrogate natural language interface on top of your graph without requiring strenuous effort to achieve. This approach also gives you a way to radically cut down the kinds of errors or hallucinations you see with ChatGPT.
This kind of ‘small’ large language model (any takers for a new acronym, SLLMs?) will soon be highly prevalent in industrial and business applications.
Think of what an e-tailer, say, could do with an SLLM—incorporating all of its product documentation from its databases, loading it into ChatGPT, and offering customers an interactive conversational chatbot over all of that complexity.
CIOs should look beyond all the frantic ChatGPT headlines and focus on exploring the untapped potential in their internal data stores by applying LLMs and building knowledge graphs using graph data science algorithms. Just imagine what you could soon be achieving.
The author is the Chief Revenue Officer at graph database and analytics supplier Neo4j.