Title: Assistive AI and Large Language Models: Addressing the Hype and Challenges
Generative Artificial Intelligence (GIA) and large language models (LLMs), spearheaded by OpenAI’s ChatGPT, have generated a lot of excitement due to their potential. However, one major issue with LLMs is their ability to hallucinate or make things up. To address this, companies like Wolfram Research are working on intermediaries that inject objectivity into the process.
Wolfram’s approach involves human curators who give data meaning and structure, lay computation upon it to synthesize new knowledge, and provide a general extension to ChatGPT. By teaching LLMs to recognize what Wolfram|Alpha might know – their knowledge engine – they can help generate more accurate and plausible information.
While OpenAI’s plugin architecture and Wolfram’s computational technology seem opposingly different, both share a common goal: using computation to automate knowledge. As the automation has gone higher up the intellectual spectrum, the approaches have become more general but ultimately execute rules.
Looking forward, there will be incremental improvements in LLMs, better training practices for better results, and potentially greater speed with hardware acceleration. However, the reliability problem for LLMs will remain forefront, as computation is still weak in following rules beyond basic things.
The combination of ChatGPT’s mastery of unstructured language and Wolfram’s mastery of computational mathematics can be advantageous in various use cases, such as performing data science on unstructured GP medical records.
In the coming years, incremental improvements and better training practices are expected for LLMs. While a sea-change similar to the last 12 months may be ruled out due to compute costs and potential setbacks, the combination of ChatGPT and Wolfram’s technology can still yield strong results when giving clear instructions.
