The most important part of AI chatbots is being ignored
AI firms can increase the power of their chatbots without updating to new models. Columnist Jacob Aron explains how a simple bit of software can give AI a boost – and whether AI will soon be able to rewrite that code itself
Amidst the clamor surrounding new large language models (LLMs) like GPT-6, the crucial role of the "harness" that governs them often goes unnoticed. Jaap Arriens of NurPhoto via Getty Images explains that while the LLM at the heart of chatbots like ChatGPT or Claude remains the same across providers, it is the harness that differentiates their performance.
The harness is essentially an interface layer that steers the raw output of the LLM, directing it to perform tasks such as searching the web, reading documents, and engaging in "chain-of-thought" reasoning to improve responses. OpenAI's implementation of this harness has yielded remarkable results, with test scores for its GPT-6 Astra model jumping by 80 percentage points using the provider adapter harness compared to a standard one.
The harness is not merely about personalizing the experience; it fundamentally shapes the capabilities of AI chatbots. For instance, when given the same prompt "Who am I?" via API, the model produced a basic response. However, through the harness on chatgpt.com and a logged-in personal account, the chatbot delivered much more informative and context-aware replies tailored to the user's identity and interests.
Despite the critical importance of the harness, it remains largely under-discussed in comparison to the constant buzz around the release of new model families. AI firms rarely spotlight their innovative harness developments, focusing instead on unveiling the latest LLM advancements. This oversight obscures the true differentiator that holds the key to unlocking the full potential of AI chatbots.
Written by urgent.news from New Scientist's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
