OpenAI’s language machine – ChatGPT – leverages its algorithm to provide users with everything from essays, computer code, to philosophical conversations. Its latest version, GPT-4, is an improvement over its predecessor GPT-3.5 in many ways, chiefly its ability to accept images along with text inputs. 

Additionally, GPT-4 now shows human-level performance on various professional and academic benchmarks such as the bar exam, which it passed in the 90th percentile, and a biology Olympiad which finished in the 99th percentile. This is compared to GPT-3.5’s score which ranked in the 10th percentile for the bar exam, and 31st percentile for the Olympiad. 

Image understanding capabilities will not be available to customers immediately. OpenAI is currently testing the tool with Be My Eyes for which it will be used as a digital assistant for people who are seeing impaired.

The new image input feature will see GPT-4 using images, graphs and text in unison as input to influence its output. As of now, users can simply input text in a variety of languages, for ChatGPT to use as a prompt to respond to. 

The development of GPT-4 is in part thanks to Microsoft, who worked with OpenAI to develop a supercomputer in Microsoft’s Azure cloud which was used to train GPT-4. Microsoft announced that their Bing chatbot has already been running on an early model of GPT-4 for the last five weeks. 

Using feedback from the experts, GPT-4 also offers advanced selection and filtering, evaluations and expert engagement, and monitoring and enforcement according to the release. Thanks to the mitigations in the new version of GPT, they were able to decrease the model’s tendency to respond to requests for disallowed content by 82%, responding to sensitive requests in accordance with OpenAI’s policies more often. 

There is currently a waitlist to access the GPT-4 API, and pricing for the service is £0.025 for 1,000 prompt tokens and £0.049 for 1,000 completion tokens. Default rate limits for the service are 40k tokens per minute and 200 requests per minute. Prompt tokens are the raw text fed into GPT-4 while completion tokens are content generated by the tool. 

Recommended

• What does Digital Transformation Mean to Incremental CEO, Mark Skelton?
• New MI5 Organisation to Combat State-backed Security Threats
• UK Broadband: Is 5G Making a Difference?

With enhanced capabilities, questions remain about the risks involved with GPT-4. The UK’s National Cyber Security Centre identified potential risks in Large Language Models (LLMs) and AI chatbots. 

They found that organisations should take great care with the data they choose to submit in prompts, in order not to place organisational data at risk. 

“A common concern is that an LLM might ‘learn’ from your prompts, and offer that information to others who query for related things”, said the report. “The query will be visible to the organisation providing the LLM (so in the case of ChatGPT, to OpenAI). Those queries are stored and will almost certainly be used for developing the LLM service or model at some point.”

In the process of training GPT-4, OpenAI placed importance on making the tool safer and more aligned from the beginning. To understand the extent of the risks involved with GPT-4, OpenAI worked with over 50 experts in AI alignment risks, cybersecurity, biorisk, trust and safety, and international security to test the model.