OpenAI has revealed GPT-4, a new large language model that surpasses its predecessor, ChatGPT. However, the company remains secretive about the reasons for its improved performance. GPT-4 marks OpenAI’s transition from a nonprofit research lab to a for-profit tech firm.
Introducing GPT-4
Key Features and Capabilities
As a multimodal large language model, GPT-4 can respond to both text and images. It can suggest recipes based on the contents of your fridge, and even explain why a meme is funny. Access to GPT-4 will be available to those on the waitlist and subscribers of ChatGPT Plus in a limited, text-only capacity.
Discover the world of cutting-edge technology at Techtyche.
Industry Reactions
Oren Etzioni at the Allen Institute for AI and Thomas Wolf, cofounder of Hugging Face, have praised GPT-4’s potential in multimodal applications, but also note that without technical details, it is difficult to assess the results.
Comparing GPT-4 to ChatGPT
OpenAI states that GPT-4’s performance surpasses ChatGPT due to its larger size and more parameters. GPT-4 has shown better results in basic reasoning, human tests, and safety features. However, the company has not revealed the size or details of GPT-4’s construction.
Improved Safety and Accuracy
According to OpenAI, GPT-4 is 82% less likely than GPT-3.5 to respond to prohibited content and 60% less likely to generate false information. The company achieved these improvements through reinforcement learning via human feedback.
If you’re interested in learning how human intelligence is making a comeback in the world of AI, don’t miss our article on how Go players are striking back against AI systems
Competing in the AI Market
While GPT-4 is a powerful multimodal large language model, it faces competition from other models like DeepMind’s Flamingo and Hugging Face’s open-source project. Major companies such as Microsoft, Stripe, Duolingo, and Morgan Stanley have partnered with OpenAI to use GPT-4 in their products.
Take a look at the latest news about the technology world and how it’s progressing.
Remaining Challenges
Despite improvements, GPT-4 still generates biased, false, and hateful content and can be hacked to bypass guardrails. OpenAI’s Ilya Sutskever suggests slower releases for models with unprecedented capabilities might be desirable in the future.
Conclusion
OpenAI’s GPT-4 has raised the bar for large language models with its impressive multimodal capabilities and improved safety features.
Although the company remains tight-lipped about the technical details behind GPT-4’s enhancements, it has already attracted significant interest from major partners in various industries.