One of the most impressive features of GPT-4 is its ability to process text and image inputs. However, its responses are limited to text only. OpenAI acknowledges that GPT-4 still faces some of the same issues as earlier language models, such as the tendency to fabricate information (known as “hallucinating”) and generate violent or harmful text.
But the good news is that OpenAI has already partnered with several companies, including Duolingo, Stripe, and Khan Academy, to integrate GPT-4 into their products. As a subscriber to ChatGPT Plus, OpenAI’s monthly subscription service costing $20, you can access the new model right away. And Microsoft’s Bing chatbot is also being powered by GPT-4. Furthermore, the model will be accessible as an API for developers to build on. You can join the waitlist today, and OpenAI is expecting to start admitting users tomorrow.
Microsoft revealed that its Bing chatbot was already using GPT-4, but some users could bypass Bing’s safety measures and make the bot provide dangerous advice, make threats, and spread false information. Additionally, GPT-4 still lacks knowledge about events after September 2021, as this is where its data cut off.
OpenAI recently shared in a research blog post that the difference between GPT-4 and its predecessor, GPT-3.5 (which powers ChatGPT), may be subtle in casual conversation. OpenAI CEO, Sam Altman, tweeted that although GPT-4 “is still flawed, still limited,” it also appears more impressive upon first use, though its limitations become more apparent with extended use.
It’s reported that GPT-4 has demonstrated significant improvements in various tests and benchmarks, including the Uniform Bar Exam, LSAT, SAT Math, and SAT Evidence-Based Reading & Writing exams. In these exams, GPT-4 achieved scores in the 88th percentile and higher, indicating its exceptional performance.
There has been a lot of speculation surrounding the capabilities of GPT-4 over the past year, with many expecting a significant leap forward from its predecessors. However, OpenAI’s announcement suggests that the improvements are more gradual, as the company had previously cautioned.
In an interview back in January, Sam Altman, OpenAI CEO, expressed concern that people’s expectations were too high and that they would ultimately be disappointed. He explained that while GPT-4 represents a significant step forward, it’s still not the all-encompassing artificial general intelligence (AGI) that many are hoping for.
Last week, rumors about GPT-4 were further fueled when a Microsoft executive revealed in an interview with the German press that the system would be launched this week. The executive also hinted that the system would be multi-modal, which could generate more than just text. Many experts believe that multi-modal systems that integrate text, audio, and video are the key to developing more capable AI systems.
GPT-4 is now confirmed to be multi-modal, but it only supports text and image inputs and generates text output, contrary to some speculations. OpenAI explains that the model’s capability to process text and images simultaneously allows it to handle more complex inputs.
It has been a gradual journey for OpenAI and AI language models in general to achieve GPT-4’s level of sophistication. The original research paper on GPT was released in 2018, followed by GPT-2 in 2019 and GPT-3 in 2020. These models were trained on vast amounts of text data, most of which were scraped from the internet and used to detect statistical patterns. These patterns were then used to predict what word should come next. Although the mechanism is relatively straightforward, the outcome is highly flexible systems that can generate, summarize, and rephrase text, translate languages, or generate code.
Initially, OpenAI hesitated to release the GPT models, fearing they could be used maliciously to generate spam and misinformation. However, in late 2022, the company launched ChatGPT, a conversational chatbot based on GPT-3.5, which was available to anyone. ChatGPT’s release sparked a frenzy in the tech industry, with Microsoft soon following with its own AI chatbot, Bing. Google also scrambled to catch up.
As expected, the wider availability of AI language models has brought some problems and challenges. The education system is still adapting to software that can produce college-level essays. Online platforms such as Stack Overflow and sci-fi magazine Clarkesworld had to close submissions due to an overwhelming influx of AI-generated content. Early use of AI writing tools in journalism has been somewhat difficult. Nonetheless, some experts contend that the negative impact of AI language models has been less than anticipated.
According to the company, GPT-4 is “82 percent less likely to respond to requests for disallowed content and 40 percent more likely to produce factual responses than GPT-3.5” in internal tests. However, this does not mean the system is error-free and does not generate harmful content.
Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: https://t.co/TwLFssyALF pic.twitter.com/lYWwPjZbSg
— OpenAI (@OpenAI) March 14, 2023
