The state of AI in 2026

0:00 So, for those of you that are new to the channel, my name is Dave Abelar, the founder of Data Lumina. I'm an AI engineer with over 10 years of experience working in the field. We run an AI company. So, everything that I'm sharing within this video is through the lens of what we see actually happening happening right now for the client projects that we are working on. So, I'm going to go through seven major things that I've noticed.

0:22 I've got some notes over here and I will share some resources. So, let's dive in. And the first point that I want to talk about is the limits of LLMs. And more so the question of whether we are already hitting the limits of what we can do with large language models. So, as you know, one of the biggest issues right now when working with AI, integrating with AI, pretty much doing anything with AI where we use large language models is the big big challenge of hallucinations.

1:11 It's like this coworker that's already pretty smart, but every now and then it just like makes things up. It completely lies about certain things. That's of course not really a coworker that we want to have around, right? And that's really the current state of AI. And the question pretty much becomes like, "Look, how far can we push all of these new models? How far can we push these various benchmarks that there that there are? And are we going to reach what we call AGI with the current LLM paradigm that that we're dealing with?" And while I think throughout this year we will continue to see better performing models.

1:56 They will get better at coding. They will get better at certain tasks. The benchmarks will increase. But I think to be honest, we as first of all users and also programmers working with these models are not necessarily going to experience very drastic changes in the terms of like how we extract value from them and how we build applications around them. Yes, coding agents will get better. Yes, they will get better at using multiple tools. But if you really look at what happened in literally the past 2 years and you compare for example like the early like GPT-4 old models which has have already been out for a very long time.

2:19 If you now compare that to the GPT-5.2 models, the way you build applications around these models and how you really integrate them and extract value from them is still pretty much the same. Later in this video, I will talk specifically about the notion of using DACs versus using pure agents. So, we'll get to that. But I think we are not going to experience a real breakthrough just by increasing the benchmarks on all of these tests that we currently have for these models with the current LLMs, the current LLM architecture.

3:08 So, one of the things that I am paying attention to is for example the work from Yann LeCun, who was the former I think chief of AI at Meta, at least a very well-known and respected AI scientist that has been working in the field for over decades. And earlier last year, so this was a paper submitted in December 2025, there was this interesting paper that was released where Yann LeCun also contributed to where they introduce a kind of like new paradigm algorithm, if you will, where they compare pretty much the limits and the flaws of LLMs and introduce a new way that we could potentially engineer intelligence.

3:54 Now, I'm not going to go into the details of this paper, but this is a very interesting read. And if you look at most experts out there like true experts, the like leading researchers and scientists, they all pretty much say, "Look, we are hitting a limit. We are hitting a wall with these large language models. And we need to think about it very differently." So, I'm very excited to see what will come out of this year and what research breakthroughs we will have.

4:16 Now, how I pretty much translate that to the current work that we are doing right now is I'm pretty much I'm very focused on applied AI. So, usually I pay attention to science, but unless I can use it in a product or in a project today, I don't really care about it. That's just like what I like to focus on because that's what I do with this YouTube channel. That's also what we do with our client work.

4:40 Now, another interesting paper that recently was shared with me was this one about recursive language models. Now, this is just a way to query and chain LLM calls together in a way where we could overcome potentially some of the limits of the context window, the limited context windows that large language models have. So, this is not a new algorithm that really new paradigm. This is just a clever way to recursively chain language models together, which show really promising results.

5:05 So, I will leave the link to this research paper below this video as well. There are also some code snippets that they shared for you to play around with this yourself. So then to summarize the current state of LLMs, if you ask the true experts, they think LLMs are going to hit the wall and we need something entirely new. So, I'm closely paying attention to that research and where that is heading. But knowing that in the day-to-day right now, LLMs are the best tools, the best AI tools that we can use.

5:28 So, I'm also paying closer attention to new approaches and techniques to maximize the effectiveness of current modern-day LLMs like for example the paper I just shared. All right. And then topic number two that I want to talk about and that I am paying attention to is Google. So, as you may know, for the past 2 years, Google was pretty much quiet. We saw OpenAI and Tropic, DeepSeek from China,