A recent study conducted by Purdue University in the United States has shed light on a concerning issue related to the accuracy of ChatGPT’s responses to programming questions. The study’s findings indicated that more than half of ChatGPT’s answers were incorrect, and its advanced language abilities managed to mislead a significant portion of the participants.

Why Relying on ChatGPT May Be Counterproductive

The research team examined 517 programming questions sourced from Stack Overflow and assessed various aspects of ChatGPT’s responses, including correctness, consistency, comprehensiveness, and conciseness. The outcomes of the evaluation were disappointing, as it was revealed that 52% of the provided answers were inaccurate, and a substantial 77% were unnecessarily lengthy. What raised even more concern was the observation that the AI’s eloquent and methodical language style often led the participants astray. Only in cases where the errors were glaringly obvious were the participants able to identify the inaccuracies.

In spite of the incorrect responses, nearly 40% of the participants preferred ChatGPT’s answers. However, a significant 77% of those favored responses turned out to be incorrect. The researchers, including individuals like Samia Kabir, David Udo-Imeh, Bonan Kou, and Assistant Professor Tianyi Zhang, clarified that many errors stemmed from ChatGPT’s inability to grasp the contextual nuances of the questions.

These findings present a compelling argument that current generative AI, in its existing state, may not be a suitable tool for assisting with code generation and might even have counterproductive effects. Acknowledging this reality, various tech giants such as Google, Apple, Amazon, and Samsung have issued warnings or imposed bans on the utilization of generative AI for code suggestions.

According to reports, OpenAI is working on its next GPT iteration GPT-5 which is expected to solve these errors. Expectations include reduced hallucinations, improved multi-modality with text, images, videos, and audio, enhanced computational efficiency, memory, and contextual understanding. GPT-5 could enable more detailed interactions, expand into new domains, and offer a higher number of parameters for more powerful AI content generation.

RELATED:

• OpenAI rolling out new ChatGPT features to enhance user experience
• Here is how AutoGPT is enabling Business Automation and Encouraging Side Hustles
• OpenAI Makes Conversational AI more Accessible with the Launch of the ChatGPT Android App
• OPPO Reno10 vs OPPO K11: Specs Comparison
• OPPO Reno10 Pro vs Nothing Phone (2): Specs Comparison

(Via)

The post ChatGPT’s programming answers wrong over 50% of the time, study finds appeared first on Gizmochina.

OpenAI CEO, Sam Altman, has confirmed that the company is not currently training a GPT-5 language model. Altman addressed this issue during a virtual event at MIT, where he emphasized that OpenAI is currently focused on upgrading and updating their existing GPT-4 model. He added that the company is working on addressing safety issues related to the model and that they had spent over six months training GPT-4 before its public release.

Altman also discussed some of the concerns raised in an open letter, allegedly signed by Elon Musk, calling for a pause in the development of large-scale language models. Altman noted that the letter lacked technical nuance and did not provide specific guidance on where the pause should occur. He emphasized the importance of increasing safety rigor as capabilities become more serious, and he stated that moving with caution is critical.

While OpenAI plans to continue being transparent, the company has not released any information on GPT-4’s training data or its architecture, construction, or other inner workings. This decision was based on concerns related to the competitive landscape and safety implications of large-scale language models like GPT-4.

Overall, Altman’s comments suggest that OpenAI is taking a measured approach to the development of language models and is focused on addressing safety concerns as it moves forward. While there is no GPT-5 in training, the company is committed to upgrading and updating its existing models to improve their performance and safety.

RELATED:

• Amazon Unveils Bedrock to Compete with OpenAI’s ChatGPT and DALL-E 2
• ChatGPT can resume service in Italy if OpenAI agrees to meet watchdog’s terms
• ChatGPT is Offering Rewards of Upto $20,000 for Finding Bugs
• Microsoft Edge Brings DALL-E-Powered AI Image Creator to Your Desktop
• Google’s Race Against Time: AI Chatbot Confirmed for Search, But When Will It Launch?
• ChatGPT Reportedly Blocks Large Number of Asian Accounts, Causes Panic Among Users
• OpenAI may face world’s first defamation lawsuit over ChatGPT by Australian Mayor
• Motorola Edge 40 Pro vs OnePlus 11: Specs Comparison

The post OpenAI CEO clarifies company is not currently training ChatGPT-5 appeared first on Gizmochina.

OpenAI, the artificial intelligence research lab, has been making headlines recently with the release of ChatGPT-4, the latest iteration of its Large Language Model (LLM). However, reports have already surfaced about the development of ChatGPT-5, with rumors suggesting that OpenAI expects to complete its training by the end of 2023.

The details surrounding GPT-5 remain murky, as OpenAI has not publicly discussed the project. However, a tweet by developer Siqi Chen has suggested that some individuals within the company believe that GPT-5 could achieve Artificial General Intelligence (AGI), a significant milestone in the development of AI that would mark the point at which machines can display human-like understanding and intelligence.

It is unclear at this time whether GPT-5 will actually achieve AGI, and Chen later clarified that his original tweet was not representative of a consensus within OpenAI. Nevertheless, the potential for GPT-5 to surpass the capabilities of GPT-4 is significant, given that the latter is already capable of analyzing both text and images, performing complex tasks, and displaying creativity.

Microsoft is already using GPT-4 for its new Bing search engine, which was launched in preview earlier this year. GPT-4’s advanced capabilities have already been put to use in a variety of applications, and it seems likely that GPT-5 will be even more capable.

Despite the potential for GPT-5 to advance the state of the art in AI, there remain significant concerns about the idea of machines achieving human-level intelligence and understanding. While some reports have suggested that ChatGPT has already passed the Turing test, which evaluates a machine’s ability to exhibit intelligent behavior that is indistinguishable from that of a human, other experts have disputed this claim.

We might get an intermediate ChatGPT version before GPT-5, with OpenAI expecting GPT-4.5 to launch in September or October. Nevertheless, the continued evolution of OpenAI’s LLMs represents an exciting development in the field of AI and natural language processing.

RELATED:

• Google Accused of Stealing ChatGPT: Is There Truth Behind the Claims?
• Man claims ChatGPT saved his dog’s life after vets couldn’t make proper diagnosis
• Cerebras Systems Unveils Open Source ChatGPT-like AI Models
• The AI Crime Wave: Police Warn of ChatGPT’s Dark Side
• Microsoft threatens to restrict Bing search data access to AI chatbot competitors
• China’s AI ambitions hindered by funding and chip import issues

(Via)

The post ChatGPT-5 rumored to launch at the end of 2023, will it achieve AGI? appeared first on Gizmochina.