There are few things that can generate as much buzz as the artificial intelligence research lab OpenAI announcing a new number to follow the letters GPT. However, the announcement of GPT-4 represents a significant milestone in the large-language model’s development. At least for the limited number of people that will get early access to the model.
New and improved OpenAI?
As might be expected, the latest model is touted as a significant improvement over the previous versions. Sam Altman, co-founder of OpenAI, has called it their “most capable and [human-]aligned model yet”. The latest version is touted as being more creative, less biased, and more accurate than its predecessor.
The dangers of AI models simply repeating incorrect information and reinforcing human biases is well documented. Indeed, Bard, the Google rival, suffered a setback when it answered a question incorrectly at a public demonstration. OpenAI claims that the latest iteration of GPT is 40% more likely to provide a factually correct answer, and published data suggesting it would be able to pass many common exams with a high grade.
However, the model also boasts significant new features. A key change is in the ways users interact with the model. The model can act on prompts that are longer than 20,000 words, and respond at a similar length. But although the use-case for novella-length dialogue may be limited, the new version offers the most sophisticated opportunities for the ‘chat’ part of ChatGPT.
The model is now multimodal, so it can accept pictures as input. For example, it can recognise and discuss objects in a picture, or use it as a prompt to develop further work. It is also more Socratic in its approach, and for the first time can answer a question with a question. OpenAI has also said that, although not currently available, it can have different personalities.
‘It looks like you are writing an AI prompt…’
One beneficiary of the announcement was Microsoft, who saw their stock rise by 6.5% after the announcement. Microsoft has invested heavily in OpenAI and has already announced their Copilot feature, which is being rolled out to Office 365 subscribers.
Copilot marks Microsoft’s return to having an assistant for its applications. However, unlike the much-derided Clippy (which was retired in 2001, but survives in memes), Copilot actually promises to be useful. The service offers a range of behind-the-scenes and user-prompted enhancements. Copilot can summarise meetings, for example, and notify people of actions it had identified and even flag the relevant parts of discussions to people who were not present.
It will also enable people to interact directly to help work on the documents they are creating. Simple uses might include formatting spreadsheets and documents, allowing the user to focus on creating the content. But for those who would prefer to delegate that, there is even the prospect that, via the chat interface, Copilot could draft an entire document or presentation based on user requests and other documents that are available to it.
Copilot, and the similar assistants that will emerge, like Google’s Bard, could potentially transform the workplace. Perhaps the most obvious outcome will be a further reduction in the corporate need for administrative staff. But it will also mean a redefinition of the knowledge work and may well prove predictions that the next growth career will be AI prompt writer.
When OpenAI is not an open AI
However, GPT-4 has received some criticism about its opaque nature. Some builds on OpenAI’s transition in 2019 into what was termed a ‘capped for-profit’. While the transition was touted as, among other things, allowing OpenAI to be a competitive employer and to raise investment, it also moved OpenAI from its origins in research.
Critics, including Elon Musk who donated to OpenAI when it was a non-profit, have suggested that OpenAI is no longer open. Ben Schmidt, an AI expert, pointed to OpenAI’s own technical report, which stated that it would not provide some details because of the ‘competitive landscape’. While the paper contains specific examples of how the model is more accurate or less offensive, those arguing for openness say it’s impossible to know if OpenAI’s claims about it being better are actually true.
Indeed, OpenAI themselves are keen to point out that AI is still an evolving technology, and noted that it was still less capable than humans in some scenarios, with Altman stating “it still has flaws.” Despite improvements, the problems of previous AIs have been reduced, not removed, and OpenAI acknowledges that the best results are likely to come when GPT-4 is teamed with a human editor.
But despite the flaws, GPT-4 is a significant improvement over the previous models. The question it’s unlikely to be able to answer for most people, though, is when they can get to use it. Currently, there is a waitlist for access to the test interface, with no announcement on public availability for image input. For most people, it may be that their first experience of GPT-4 will be in Clippy’s reincarnation.
Aaron Goldman, Chief Marketing Officer, Mediaocean said: “Improvements in generative AI are coming along very quickly. With ChatGPT-4, users now have the ability to train the model and feed it more background information, ultimately making the end product much more tailored. When it comes to AI, more control over the inputs also means much better outputs. This additional layer of control opens up whole new ways for companies to take advantage of the technology, customise it and make it bespoke for their own use cases, their own information, and their own needs.”