The Last Year in AI

A GAME OF INCREMENTS

It was only in November of 2022 that OpenAI released the 3.5 version of its ChatGPT tool, leading to the current zeitgeist-dominating, all-encompassing rush to develop and deploy AI-based technology. 

It was the first time the general public had really started to take notice of the capabilities of tools like this and the impact was huge. ChatGPT became almost an overnight success with huge media coverage and a step change in the pace of development. With new interest and investment, more and more companies began to develop their technologies in an effort to compete.

Since then the gap has closed and we’ve seen a tit-for-tat fight developing between the key players, OpenAI, Anthropic and Meta. The margins of improvement are much smaller and more frequent, with each company also trying to carve out a niche and a USP to beat out the others. For example, Meta has been engaging with open-source technologies aimed at local deployment which gives it an edge when it comes to being able to limit security risks for business-sensitive data. That comes with development resources and budget overheads though that place it out of reach for part of the market, who turn to the ever more integrated and simple UI of something like ChatGPT.

That’s not to say we haven’t seen some real strides forward in the last year, though - let’s take a look at some of the most important ones.

CUSTOM INSTRUCTIONS

Back in July of 2023, ChatGPT announced the introduction of a new way to control the tool’s responses, called ‘Custom Instructions’. This additional set of instructions could be used to provide a second set of instructions to the chatbot alongside every prompt.

That meant it was now possible to build a ‘character’ or ‘role’ for the bot outside of the prompts it was being given. It’s been well-documented that response quality improves if generative AI tools are given a clear idea of what they are supposed to do and how they should go about it. Rather than having to provide that in every prompt, custom instructions mean they can be set once and only updated when needed.

Custom instructions also meant it was now even easier to use ChatGPT to power tools used by third-parties. If a user interacted with a tool that now had a clearer role and specific instructions on how to generate its responses, the quality of that interaction was much higher. It could also be used to give the responses more flavour and character, instead of being limited to the default approach of the base model.

The introduction of this feature was important in itself but also represents the beginning of the focus on an ‘agent’ based AI workflow in the industry. If a task could be performed not by one universal bot, but instead by lots of specialised ‘agents’ with specific roles and instructions, the level of complexity in the tasks it could complete would start to drastically increase.

In a world of increasingly specialised AI agents - the value of an individual, consistent viewpoint from a human is likely to increase where it holds expert knowledge, practical experience or other elements of what Google would refer to as E-E-A-T. 

A HISTORIC MOMENT

Shortly after the introduction of Custom Instructions, OpenAI delivered another important announcement. In September of 2023, ChatGPT gained the ability to browse the internet to perform research to answer its prompts. There had been fleeting access to links and internet content prior to this, but this was the watershed that gave the models free reign to access the vast stores of information across the web.

The real change here was how this addressed one of ChatGPTs previous biggest challenges as a practical tool. Due to the way it had been trained, the data and knowledge it had access to only ran up until September 2021. This meant it was unable to answer questions about anything that had happened since that point with any amount of factual accuracy.

Unlocking ChatGPT from being stuck using historic data meant it was far better at being able to stay up to date with current events and recent developments. While it may be able to view and process more current information, being able to filter and process that data for potential biases and veracity is still a challenge best suited to a human. 

THE NEXT FRONTIER

With tools like Midjourney and Dall-E getting better and better at generating believable images, the next frontier of generative AI visuals is most certainly video. There have been strong showings from tools like Runway previously but it was back in February that OpenAI blew us away with the capabilities of their as-yet-unreleased Video generative AI model, Sora.

A Midjourney image generated with a late 2023 model.

Previously, the issues that had plagued early image generation models were even more prominent in motion. Combine the mismatched anatomy, shifting and blending of multiple subjects and confused layouts with the uncanny valley effect of “not quite right” human faces - and you get a product that is best described as a gimmick. With no widespread practical applications, the use of video generative AI was limited to areas where it’s unique shifting and smearing qualities was used for artistic effect (such as the controversial title sequence of Marvel’s ‘Secret Wars’). 

The same prompt as above with a Midjourney model from early 2023.

What we saw in the promotional content for Sora and in subsequent posts to X from those with access was nothing short of revolutionary. Subjects were clear and generally sustained over the course of the video, accuracy and realism were impressively maintained and the improvements to motion and movement were beyond impressive. This is arguably going to be one of the most impactful model releases we’ve seen yet, as it moves into a whole new realm of applications.

Not only can companies and individuals reach for the highest form of engagement content on-demand and with minimal budget - but bad actors and those with a more nefarious agenda might be able to wield unprecedented capabilities. We already see misinformation and deepfakes raising concerns across the globe - the introduction of realistic generative AI video into this ecosystem could have far-reaching consequences. This may be one of the reasons OpenAI has been slow to release it.

LOOKING FORWARD

We’ve only scratched the surface of the developments we’ve seen in the last year, with plenty more tweaks and changes building these tools into more and more capable platforms. The pace of change looks set to continue, if not accelerate, with new techniques and ever-growing computing power delivering better and better results.

There are going to be significant shifts in the coming months, with the release of Sora being of particular importance. Keeping up to date with the shifting landscape of generative AI can be hard so it’s worth making sure that new tools and exciting possibilities don’t distract from the fundamentals. Generative AI tools should be used to enhance, optimise and support the ways we work, not replace it. Human-created content is going to become an ever more valuable commodity as it becomes less prominent - so creating engaging and authentic content should be the absolute priority. 

KEY TAKEAWAYS

• Generative AI has improved rapidly over the last year and a half, not just in quality but also capability.
• Model owners like OpenAI and Meta are trying to find niches where they can outcompete their peers, on factors such as security and ease-of-use.
• The introduction of competent video-generative AI will be a significant change not just for the marketing industry but for the world as a whole.
• Human-created content could become an increasingly valuable commodity in the coming months and years.

If you're interested in discovering how AI may be able to enhance your digital marketing strategies, get in touch with us today!