What’s happening?

OpenAI does not appear to be slowing down in its pursuit of market leadership in generative AI technology. The company announced the release of GPT-4, an improvement on the current model powering ChatGPT, called GPT-3.5. This release came as no surprise given the wild success of ChatGPT. Many speculated that Google or Facebook would be releasing their language models and we’d see a face-off, but OpenAI isn’t waiting for that to happen before releasing improvements into the market.

Much ado about a 0.5 increase?

If only the new release was about the name alone, but the new GPT-4 boasts the following improvements:

• Image analysis: the major improvement in GPT-4 is its capability of analyzing images that users upload to it. This means that users can upload images, and ask the AI to generate solutions, conversations and every other capable interactions from the uploaded image. OpenAI demonstrated this with a demo where they showed how a sketch could be turned into a functional website in minutes. An image of a meme was also uploaded to the AI and asked to explain the joke and it did this with great accuracy. This feature is a game changer as it doesn’t limit the creativity of prompts to words alone, uploading sketches of your ideas will get the AI to produce more desirable outputs therefore deepening AI and human collaboration.

• Improved test taking: When ChatGPT first surfaced, lots of users and articles praised its ability to take tests designed for humans and pass it. This brought about intense debates about the continued existence of tests since students could use AI to bypass it. Those debates are about to go haywire because GPT-4 has scored in the highest percentile of several tests given to it while GPT-3.5 sometimes passed only in the lower 10th percentile. What this means is that the AI has improved greatly in understanding inputs and giving better results from its database

• Providing more precise responses: According to OpenAI, GPT-4 is able to generate written responses that are longer, more thorough, and more trustworthy than the previous version. The most recent version can now respond with up to 25,000 words, an increase from the previous 4,000, and can give thorough directions for even the most unusual situations, such as how to prepare Nigerian Jollof or file your taxes.

• Improved steerability: GPT-4 has improved on the way its tone, verbosity, and style are presented. Rather than having a fixed style, users can now determine how they want the AI to interact with them by telling the AI their preference and unlike previous models, the AI won’t break out of character even when asked. Examples of this can be found in a post by OpenAI where they showed examples of System set tone and how the AI interacted. In one, the system was set to respond in Socratic style where the AI doesn’t give the user the answer but gives it all the clues and help the user needs to arrive at the correct answer themself. Despite the user telling the AI to stop responding in that manner, the AI doesn’t break character.

All’s not yet rosy:

While these improvements are laudable, there are still limitations with GPT 3.5 that are evident in GPT-4. The technology still doesn’t have any knowledge later than September 2021 and does not learn from users. This means it may still return wrong answers when prompted for current information. GPT-4 could also make predictions that are sure to be wrong and give buggy code, especially if it doesn't double-check its work in those situations. 

OpenAI say that they are improving GPT-4 by having experts from fields like AI alignment risks, cybersecurity, biorisk, trust and safety, and international security test the model in an adversarial way and that mitigations made with the help of this group of experts have cut harmful outputs by up to 82%.

Why this matters:

The progress and risks of generative AI show how these technologies are getting better and how they could change the world. OpenAI's release of GPT-4, which has better image analysis, test-taking, response accuracy, and the ability to be guided, shows how useful this technology could be for things like language translation, content creation, customer service, education, and more. However, the ethical implications and risks associated with these technologies cannot be ignored. OpenAI's efforts to reduce the risks of GPT-4 through adversarial testing and feedback from experts show how important it is to develop AI in a responsible way. Ultimately, as we continue to push the boundaries of AI technology, it is crucial that we balance innovation with caution and prioritize the well-being of society as a whole.

• Analyze edge cases
• Look up data similar to edge cases, based on algorithms
• Extract particular data that represent the whole
• Therefore, save time and cost

Weekly AI Issues  → →

Google Releases Bard, Its Competitor in the Race to Create A.I. Chatbots

The internet giant will grant users access to a chatbot after years of cautious development, chasing splashy debuts from rivals OpenAI and Microsoft.

Can ChatGPT Plan Your Vacation? Here’s What to Know About A.I. and Travel

Powerful new artificial-intelligence software is already shaking up the travel industry, but it has a long way to go until it can plan a seamless trip. For now.

Microsoft Bing now uses OpenAI’s DALL-E A.I. to turn text into images

Microsoft is rolling out a generative image creation feature to Bing and Microsoft Edge.