ChatGPT 2.0: Creator of AI bot that took the world by storm launches an even more powerful version dubbed ‘GPT4’ – admitting it’s so advanced it could ‘harm society’
- The new version of ChatGPT can understand images and draw from more online data
- It can pass law exams with results in the top 90% – a huge leap from its earlier model
- But fears it will still have inherent biases like its predecessor
OpenAI, the creator of the wildly popular ChatGPT, has launched a new and even more powerful AI bot, GPT-4 – and admits it’s so advanced it can be ‘harmful’.
The bot can now accept both image and text input, but still output its responses in text, meaning it can provide detailed image descriptions.
When asked ‘What’s funny about this image?’ GPT-4 will reply. Among other things, it can instantly calculate people’s tax liability and pass the Uniform Bar Exam.
OpenAI said in a blog post, “We’ve created GPT-4, the latest milestone in OpenAI’s efforts to scale deep learning.
“GPT-4 is a large multimodal model (accepts image and text input, sends text output) that, while less capable than humans in many real-world scenarios, exhibits human-level performance across a variety of professional and academic benchmarks.
Open AI, the maker of the wildly popular ChatGPT, has launched a new and even more powerful AI bot, GPT-4 — and admits it’s so advanced it can be ‘harmful’
The company warned that the model is still prone to “hallucinating” imprecise facts – and can be persuaded to produce false or harmful content.
It can now pass the legal exams with results in the top 10 percent (GPT-3, which previously powered ChatGPT, was able to pass the exams, but with results in the bottom 10 percent).
GPT-4 is now available to subscribers via ChatGPT: free users will still experience the older GPT-3.5.
It also powers Microsoft’s Bing chatbot (and the company has revealed it’s been doing this for six weeks).
Yusuf Mehdi, Microsoft’s head of consumer marketing, said: “We are pleased to confirm that the new Bing runs on GPT-4 modified for search.
“If you’ve been using the new Bing in preview for the past six weeks, you’ve already had a first look at the power of OpenAI’s latest model.”
The company says GPT-4 will be integrated into other products from companies like Duolingo and Stripe.
OpenAI said: ‘GPT-4 and successor models have the potential to significantly impact society, both in beneficial and detrimental ways.
ChatGPT can achieve impressive exam results (OpenAI)
“We’re working with outside researchers to improve the way we understand and assess potential impacts, and to build assessments for dangerous possibilities that may arise in future systems.”
The ability to accept images means that users can now query ChatGPT with screenshots and other media.
Speaking this week at SXSW, the OpenAI co-founder said the fear of AI tools taking over people’s jobs was overblown — and that AI would free people up to focus on important work.
Brockman said, “Most important will be these higher level skills: judgment and knowing when to dig into the details.
“I think the real story here, in my mind, is an amplification of what people can do.”
ChatGPT averaged 13 million daily users in January, making it the fastest-growing Internet app of all time, according to analytics firm CompareWeb.
It took TikTok about nine months after its global launch to reach 100 million users, and Instagram more than two years.
OpenAI, a private company backed by Microsoft Corp., made ChatGPT available to the public for free at the end of November.
The new features of GPT-4
The big new feature here is that GPT-4 can respond to visual prompts – so it can answer, for example, “What’s funny about this picture?”
Longer prompts and answers
GPT-4 can now accept prompts of up to 25,000 words and output text of a similar length.
It can learn your writing style
GPT-4 is able to teach a certain writing style, making it an ideal partner for creative projects.
Better exam results
GPT-4 is now in the top 10% for exams like the Uniform Bar Exam.
Better to get the facts right
GPT-4 is 40% less likely to have factual errors – although that’s still the case – and 82% less likely to have banned content.
It can interact with web links
You can now ask GPT-4 to interact with web page content.