OpenAI、ChatGPT Instant 5.3は「恥ずかしさ」が減り精度が向上したと発表

2 Min ReadOpenAI released its latest update for ChatGPT, which the generative AI vendor claims delivers more fluid conversations and greater accuracy.ChatGPT Instant 5.3, released on March 3, has been designed to address some users’ complaints about stilted, unnatural chats -- criticisms that don’t appear in benchmarking but affect how people feel about the model.Explaining the revisions in a blog post, OpenAI said the update is focused on “tone, relevance and conversational flow," while delivering better contextualized results, following feedback that ChatGPT Instant 5.2 would fail to respond to questions it should be able to answer safely and occasionally reply in an “overly cautious or preachy” fashion.These traits, OpenAI acknowledged, meant that the previous model could feel “cringe” or “overbearing". The new approach dispenses with unnecessary preambles and lengthy caveats to respond to queries more directly, although, as has previously been the case, users can still determine elements of the tone -- such as warmth and enthusiasm -- using the settings.Related:Neura Robotics, AWS Collaborate to Bring Physical AI to the Real WorldThe model also improves the quality of its answers by reducing its tendency to over-index on web results and lean more heavily on its own reasoning and understanding. As an example, OpenAI says the latest ChatGPT Instant version is now more likely “to contextualize recent news rather than simply summarizing search results."By doing so, accuracy has been improved, and to back this up, OpenAI published the results of two internal evaluations that demonstrated this.The first, which focused on high-stakes domains such as medicine, law and finance, saw hallucination rates fall by 26.8% when using the web and 19.7% when relying on its own internal knowledge, according to OpenAI The second, which studied user-flagged feedback, highlighted a 22.5%drop when using web search and a 9.5% reduction without web access.However, one area in which ChatGPT Instant 5.3 does not offer a significant improvement is in filtering banned content, with OpenAI’s safety benchmarking recording slight “regressions” compared with versions 5.2 and 5.1 for disallowed sexual content and 5.2 for self-harm. OpenAI said it was monitoring these areas and noted that benchmarking values can change after launch.Other areas of immediate focus, the vendor said, are enhancing the model’s natural tone in foreign languages, with OpenAI conceding that it can sound particularly literal in Japanese and Korean, and further expanding customization options.ChatGPT Instant 5.3 is available now to all users in ChatGPT, as well as developers in the API as "gpt-5.3-chat-latest." Updates to Thinking and Pro are promised soon.Related:Anthropic Seals $100B Infrastructure Deal With AmazonIt's likely there won’t be long to wait for another update, with OpenAI posting on X: “5.4 sooner than you Think.”About the AuthorContributing WriterGraham Hope has worked in automotive journalism in the U.K. for 26 years, including spells as editor of leading consumer news website and weekly Auto Express and respected buying guide CarBuyer.