At AscentCore, we invest in world-class talent to ensure we can build and deliver innovative digital products for our clients. Our amazing team holds so much knowledge and expertise across a wide range of topics. Our CTO, Cornel Stefanche, loves the opportunity to chat with staff about what they’re passionate about. Cornel recently chatted with Liviu Sopon for an AI year-in-review and a peek at what we can expect in 2024.
Cornel Stefanache: Liviu, welcome! I’m so excited to talk about all things AI with you, so let’s dive right in and talk about GPTs. AI was a hot topic last year, and I’m sure it will continue this year.
Liviu Sopon: Yes, there’s no doubt that everyone is really into GPTs. From what I understand, Code Llama, which is the current open source, was internally leaked, and then later on, they still decided to release it as open source. I don’t think we would have all these models like Mistral and Neural Chat if that hadn’t happened.
Cornel Stefanache: The fact that they managed to provide open-source solutions for models like these and small models that run on less powerful devices, like home computers, was a big step forward in generating new applications.
Liviu Sopon: Yes, and by the time GPT-3.5 launched, the public truly became interested in AI in the chatbot area. We’ve seen this increased activity everywhere because the public is so interested. More investments are happening in the chatbot space.
Cornel Stefanache: Looking back at 2023, we can see two significant events in the chatbot space. One was improving output, and the other was introducing more types of input. Before Mistral, we had GPT-4, which allowed you to upload images, voice, and text and gave you an image or text as the output. The output types have been generative, and I expect that in 2024, we’ll continue to see advancements in this area, such as video generation.
Liviu Sopon: Yes, most likely. When you make an image, it’s somehow easier. This is the first time people recognize that you have to think like an artist, at least in stable diffusion. As humans, we have an idea of an artistic representation in our minds. In stable diffusion, if you ask it to draw a robot, the odds that it will look like what you imagined are pretty slim. So, people are grappling for the first time with the fact that you have to be very specific to get exactly what you have in your mind.
You could say that this is problematic for illustrators and designers, but it seems to me that it’s more important to have good direction. But I’m curious to see how this will play out and how this will translate into video.
Cornel Stefanache: We certainly can’t talk about AI without talking about its implications on the workforce. In this instance, we’re giving people skills they didn’t have before because not everyone knows how to draw. But I also think it pushes the whole design field forward because not everyone knows how to design well. So, you offer this possibility to a much larger population, and then artists will come up with something new to compete. Because, in the end, AI can only generate what it’s seen. It can’t generate a new design.
Liviu Sopon: It comes back to this idea of specificity and good direction. You have to be very specific if you want exactly what you have in mind to appear on screen. But often we don’t have something exactly in mind, you know? And I think that’s where image generation is really helpful. We can say, “I want a person with black hair sitting at a computer, drawn in a comic book style,” and AI handles the rest. But with video generation, I don’t know. I think it will be much more interesting to see how the prompting for videos will be.
Cornel Stefanache: I expect that for video, you’d have to fill in the gaps between two frames. So, you give it a start frame and an end frame and say, “Find content for three seconds between these two frames.”
Liviu Sopon: Yes, similar to interpolation, but the issue is, if you’ve ever played with twinning tools or 3D models, it doesn’t generate intermediate frames as you get in the animation process. Instead, you get a more linear movement.
Cornel Stefanache: This makes me think about what kind of regulations we’ll see around this. Personally, I see some significant regulatory issues around generative AI coming our way in 2024, and I think the use of tools and their integration into applications may become stricter.
Liviu Sopon: And that’s tricky. I think regulations can help. But in the current situation where the EU can’t impose regulations on states that aren’t members, we’d be shooting ourselves in the foot. The more regulations you have, the slower things move, especially when considering other countries without such restrictions, like the US or China.
Cornel Stefanache: Let’s keep looking ahead to the rest of 2024. I think we’ll see a much bigger evolution in AI than in 2023, from the services area into the area of tools around models. I think that right now, AI is where DevOps was in the 2000s. Everyone made their own tool because nothing was standardized. So, I think that standardization will need to happen. What do you think will emerge in 2024?
Liviu Sopon: Honestly, I don’t like to make predictions because usually, what I think can’t happen happens. And what I think will happen doesn’t happen! AI is already much better at generating images with text, and this year, we may see the ability to write whatever text you want there. I’ve actually already seen that Midjourney does that.
I recently read the State of AI in 2023, and it mentions “flickers,” as in sparks from the generation of 3D models. It’s not even semi-production ready, but it seems this could be a big area for development in 2024. There are already video editing tools where you can tell an image generator to make the sky starry, for example. I think we’ll soon be able to do something similar with videos. I saw some pretty nice proofs-of-concept in 2023 on maintaining consistency.
Cornel Stefanache: Yes, because maintaining consistency is the biggest problem when generating from one frame to another. I think we’ll see a combination of methods to solve this issue. Older methods were separate and proof-of-concept, but something amazing comes out when put together.
Liviu Sopon: Agreed! For example, if we combine it with stable diffusion or Dali and add some fixes for generative stylization, we’d get less flickering. I think you could quickly create a style over an image. I don’t even know how long people will continue using green screens, which is really interesting.
Cornel Stefanache: Let’s talk more about how we might move away from more traditional tools to those that can save us time. I’ve used the generative fill from Photoshop. It confirms for me that generative AI is here to stay. It just saves so much time! It won’t replace people, but it handles basic coding operations that a programmer doesn’t have to deal with.
Liviu Sopon: Yes, I haven’t found those very useful for more than just writing some basic code, but it is quite a time saver. You don’t have to know how to write code, although you do need to understand it. But you can have GPT write it for you. We’re not 100% there yet, but we’re very close.
Cornel Stefanache: This raises a security concern for me. So, if I invest enough money in creating the best coding model, but I hack it to use a key that gives me access to all the applications I’ve generated with my AI, that’s pretty easy for others to hack. We’d trust what the model generated to be complex enough that somebody couldn’t hack it.
Liviu Sopon: That’s an interesting scenario to think through. It would become tough to debug a good model because it would have quirks we’ve never seen before. Let’s say you create a trojan, which generates backdoors. For a developer to notice them, they would have to have seen them before. But potentially, the model could be capable enough to create complex backdoors that today’s programmers just wouldn’t recognize. I think this is beyond the current possibilities of a GPT. But it’s certainly something to be mindful of in the future.
Cornel Stefanache: That does leave us a lot to think about moving forward, and that seems like a good place to end. There is so much more we could talk about regarding the future of AI, and I’m sure this is just the first of many conversations we’ll have on this topic. Thanks, Liviu, for chatting with me.
