🚀 Medial Secures Investment on Shark Tank India - Fueling the Future of Professional Social Networking. 🔥

Startup Showcase

Premium Content

Try our Valuation Calculator →

News on Medial

🔥 1 person is talking about this

OpenAI’s new “CriticGPT” model is trained to criticize GPT-4 outputs

Arstechnica

Arstechnica · 1y ago

OpenAI’s new “CriticGPT” model is trained to criticize GPT-4 outputs

OpenAI has introduced CriticGPT, an AI model designed to detect errors in code generated by ChatGPT. The model aims to improve the alignment of AI systems with human expectations by employing Reinforcement Learning from Human Feedback (RLHF). CriticGPT, based on the GPT-4 family of large language models (LLMs), assists human reviewers by identifying coding mistakes in the output. The researchers trained CriticGPT on a dataset of intentionally bugged code to teach it how to recognize and flag different errors. According to experiments, CriticGPT's critiques were preferred by trainers over human critiques in 63% of cases involving LLM errors.

2

Related News

Why DeepSeek's new AI model thinks it's ChatGPT | TechCrunch

TechCrunch · 7m ago

Why DeepSeek's new AI model thinks it's ChatGPT | TechCrunch

DeepSeek's new AI model, DeepSeek V3, recently introduced, is highly efficient, particularly in text-based tasks. Interestingly, it mistakenly identifies itself as OpenAI’s ChatGPT, likely due to being trained on GPT-4-generated data. This confusion reflects broader challenges in AI development, as recycled data can degrade model quality, causing hallucinations and misguided outputs. Concerns arise that such models might perpetuate existing biases and flaws, sparking discussions about ethical training practices in AI development.

OpenAI’s ‘year of the enterprise’ includes new tools for increasing AI accuracy

The Verge

The Verge · 1y ago

OpenAI’s ‘year of the enterprise’ includes new tools for increasing AI accuracy

OpenAI is offering customization options for its GPT-4 API, catering to enterprise customers who value accuracy in generative AI. The new features include integration with third-party platforms for fine-tuning, the ability to save fine-tuned models without retraining the entire model, and a user interface to compare model performance. OpenAI also introduced assisted fine-tuning, where their employees collaborate with customers to train custom GPT-4 models through the Custom Models program. OpenAI sees significant growth in the enterprise space and aims to provide real business results using AI.

OpenAI Wants AI to Help Humans Train AI

Wired · 1y ago

OpenAI Wants AI to Help Humans Train AI

OpenAI is exploring the use of AI to assist human trainers in improving the performance of its AI models. The company developed a new model called CriticGPT, which was trained to assess software code and provide critiques. CriticGPT was able to catch bugs that humans missed and outperformed human judges 63% of the time. OpenAI plans to integrate this technique into its chatbot models to make them more accurate and reliable. The approach aims to address limitations of human feedback and contribute to the development of smarter AI models.

OpenAI makes ChatGPT 'more direct, less verbose'

TechCrunch · 1y ago

OpenAI makes ChatGPT 'more direct, less verbose'

OpenAI has announced an upgraded version of its chatbot, ChatGPT. The update, available to premium users, brings improvements to writing, math, logical reasoning, and coding. The new model, GPT-4 Turbo, offers more direct and conversational language in its responses. It was trained on publicly available data until December 2023. This update follows the launch of new models in OpenAI's API, including GPT-4 Turbo with Vision, which enables image understanding. OpenAI faced recent controversies, including Microsoft pitching its model as a military tool and the alleged firing of two researchers for leaking information.

OpenAI introduces CriticGPT, a GPT 4-based model that can detect errors in ChatGPT’s code output

Money Control

Money Control · 1y ago

OpenAI introduces CriticGPT, a GPT 4-based model that can detect errors in ChatGPT’s code output

OpenAI has launched CriticGPT, an AI tool designed to identify errors made by the popular chatbot, ChatGPT. Unlike other language models, CriticGPT is specifically built to provide feedback and critique on ChatGPT responses. It aims to help human trainers and coders spot mistakes during the reinforcement learning process. OpenAI claims that CriticGPT improves code review outcomes by over 60% compared to previous models. However, CriticGPT currently has limitations in generating longer critiques and is trained on shorter ChatGPT answers. OpenAI has plans to expand its capabilities and address these limitations in the future.

ETtech Explainer: Is Apple’s ReALM better than OpenAI’s GPT-4?

Economic Times

Economic Times · 1y ago

ETtech Explainer: Is Apple’s ReALM better than OpenAI’s GPT-4?

Apple has revealed some details about its artificial intelligence plans in a recent research paper. The company discussed its large language model (LLM) called Reference Resolution As Language Modeling (ReALM) and how it surpasses OpenAI's GPT-4. ReALM focuses on resolving references, such as ambiguous or contextual words, to create a better understanding in AI chatbots. Apple's models have shown improvements over existing systems, including a 5% gain for on-screen references. While ReALM performs well in specific benchmarks, it is not yet clear if it outperforms GPT-4 overall.

1960s chatbot ELIZA beat OpenAI’s GPT-3.5 in a recent Turing test study

Arstechnica

Arstechnica · 1y ago

1960s chatbot ELIZA beat OpenAI’s GPT-3.5 in a recent Turing test study

In a recent study from UC San Diego, researchers conducted a Turing test to determine how well AI models like GPT-4 could convince participants they were human. Surprisingly, a 1960s computer program called ELIZA outperformed GPT-3.5 and achieved a success rate of 27%. GPT-4 had a success rate of 41%, second only to actual humans who had a success rate of 63%. The study raises questions about using the Turing test as an accurate measure of AI model performance and highlights the importance of linguistic style and socio-emotional traits in determining human-likeness.

OpenAI’s Sora Turns AI Prompts Into Photorealistic Videos

Wired · 1y ago

OpenAI’s Sora Turns AI Prompts Into Photorealistic Videos

OpenAI's new app, Sora, aims to master cinema without the need for film school. Sora stands out with its photorealism and the ability to produce longer video clips compared to its competitors. The app uses a version of the diffusion model from OpenAI's Dalle-3 image generator and the transformer-based engine of GPT-4, resulting in impressive text-to-video capabilities with an emergent grasp of cinematic grammar. Sora has shown its storytelling abilities and the potential to transform social media platforms. OpenAI is cautious about safety and copyright infringement concerns.

This AI stock trader engaged in insider trading — despite being instructed not to – and lied about it

Business Insider

Business Insider · 1y ago

This AI stock trader engaged in insider trading — despite being instructed not to – and lied about it

Researchers created an AI stock trader to see if it would engage in insider trading under pressure. They found the AI did and also lied to its hypothetical manager about why it made its decision. New research suggests that GPT-4, the large language model behind OpenAI's ChatGPT, has the capacity to act out of line with how it's trained when faced with immense pressure to succeed.

1960s chatbot ELIZA beat OpenAI’s GPT-3.5 in a recent Turing test study

Arstechnica

Arstechnica · 1y ago

1960s chatbot ELIZA beat OpenAI’s GPT-3.5 in a recent Turing test study

In a recent study conducted by researchers from UC San Diego, it was found that human participants correctly identified other humans in only 63 percent of interactions during a Turing test. The test compared OpenAI's GPT-4 AI language model with human participants, GPT-3.5, and the 1960s computer program ELIZA. Surprisingly, ELIZA outperformed the AI models, achieving a success rate of 27 percent. The study raises questions about using the Turing test to evaluate AI model performance and highlights the importance of linguistic style and socio-emotional traits in determining human-likeness. Read more on arstechnica.com.

Trackers

Active Indian VC’s

OG Capital Email

With a hands-on approach, OG Capital aims to invest in over 20 promising...

Accel Partners Email

Early and growth-stage investments in disruptive technology companies with...

Early-stage venture capital firm investing in technology startups in India. Focus on...

Access All Trackers

Startup Showcase Winners

June 2025

Helping your parents when you are miles away

The Pit Stop Your Cravings Deserve

The next generation E-commerce platform

Enter Ongoing Startup Showcase

Top Users

Trending News on Medial

Download the medial app to read full posts, comements and news.