As we have showcased in our blog throughout 2024, the year has been a whirlwind for artificial intelligence, not just for the technology behind the latest AI models themselves and their multiple applications, in our case focusing on those to biological sciences, but also for AI markets and businesses. With rapid advancements, shifting markets, and adapting business strategies, it is crystal-clear that the AI landscape is growing and maturing in many ways, some unexpected.

As we step into 2025, we at Nexco have distilled some key insights about the AI landscape — what’s changed, what’s working, and where things might be headed next. We bring you this summary here, focusing our presentation on mainstream AI models such as those for language, code, or image generation, and on those AI models from the very niche applications to chemistry and biology which are having such a huge impact. Read on to find key moments and insights, what didn’t turn out exactly as expected, and a look into how AI technology and AI-related economy for scientific applications will evolve next.

The Competitive Edge in AI Is Shrinking

We open our discussion with a probably unexpected fact: The competitive margins in AI are rapidly narrowing. Certain products that were once revolutionary products and positioned their developers companies as market leaders, are swiftly imitated and often surpassed by competitors.

Consider for example OpenAI’s GPT-3, which debuted as a groundbreaking large language model (LLM) after previous attempts by OpenAI and others that were mere curiosities among computer scientists. At its launch, GPT-3 made even industry giants like Google appear to lag far behind. OpenAI quickly followed with more advanced models and systems, especially ChatGPT, but just two years later, dozens of companies now compete directly with that company, many offering comparable — or even superior — products, even totally free and open versions. Just visit the LLM arena leaderboard to be amazed at how many LLMs are out there that match or even outperform OpenAI’s top models as of today: https://lmarena.ai/ And learn from our previous post about the applications of LLMs beyond summarization, copyediting, information retrieval, consultation and programming.

A similar pattern emerged in image generation. While it’s harder to pinpoint who led the charge, it is a fact that the market is today flooded with alternatives.

In niche applications, DeepMind’s AlphaFold 2, which achieved a major breakthrough in protein structure prediction during CASP14, eventually found itself in a similar situation. Within two years, companies and academic researchers developed competing models. And Deepmind’s multimodal AlphaFold 3 system, released less than one year ago, already faces effective competitors that presumably perform on par and have less restrictive licenses.

For users and businesses, the increased competition has made more models available and with less restrictions, many open-source, and drastically reduced costs — by over tenfold in some cases for LLMs and image generation tools. However, the lower running costs and less protection forced by this race among AI model developers make it harder for them to stand out and to recover their substantial investments. It is then possible that without a new groundbreaking leap, such as the development of general-purpose AI (AGI) or multimodal systems that fully integrate physics, chemistry, and biology, businesses may face prolonged waits for significant returns.

Innovation Keeps Moving to New Frontiers

Obviously, AI doesn’t stand still. When it seems like one area of the technology is slowing down, another emerges to take the spotlight. For example, traditional LLMs like GPT-4 haven’t seen groundbreaking updates recently — there’s no GPT-5 yet for a reason. But new models focused on smarter reasoning and decision-making, like OpenAI’s o1 which is supposed to have superior problem-solving capabilities, are pushing boundaries in exciting ways. In molecular modeling, the innovation lead is as of early January 2025 on multimodal systems like AlphaFold 3, Chai-1 and Boltz-1, along with RoseTTAFold-AllAtoms by Nobel D. Baker’s lab, all of which can model not just protein structures but also their complexes with nucleic acids, ligands, lipids, and ions. Such modeling capabilities are crucial for deep understanding of biological systems and for more rapid developments in pharma and biotech.

If we were to place AI in a so-called “S curve” to show how the technology emerges slowly to then accelerate and then plateau as it matures, no doubt AI for science is well at the fast growing point, still not reaching the plateau but certainly with already a quite good track already walked.

Sometimes AI is Overkill, And Simpler Tools Are The Go-To

One surprising trend in 2024 has been the return to simplicity. While it is unclear to what extent companies are reverting to older tools for some of their procedures, we have seen many ideas fail; for example, automatic image generation works very well for many things but can’t still beat human designers for some applications, especially those that require factual accuracy. The reason? Familiar tools are reliable, and businesses prefer solutions they know will work, even if they’re not cutting-edge. Startups in the AI tools market are learning that to succeed, their products must deliver real, unique value — not just the novelty of automation.

AI for science isn’t an exception. For example, as recent editions of CASP showed, modeling of nucleic acids is still poor by AI but works much better by using homology modeling if a template is available. Likewise, ligand docking and virtual screening are expected to benefit tremendously from AI-powered systems, but detailed and extensive tests are still missing so one may be well-off with a poor-performing but at least well-known, traditional approach.

As an additional example of this, we reported previously on a Nature Methods study that explored various ways to preprocess single-cell RNA-sequencing (scRNA-seq), finding that the simplest method is in practice the best.

Services Are Back in Style

For years, the tech world avoided services like consulting, installation, or ongoing support, favoring products that customers could set up themselves. That’s changing in the AI era. Just watch ourselves at Nexco, whose main activity consists exactly in providing AI and computational services for bioinformatics, pharma, the clinic, and life sciences.

Most software, and especially so AI-based software, requires quite a lot of customization to fit into different business environments. Whether it’s training a model for a specific industry or integrating existing AI tools or pieces of software into a company’s workflows, businesses need hands-on help. That is how we fit in the market with our services along the front of AI and computers for biology.

Some companies are even building entirely new service-based models, which combine cutting-edge tools with expertise to deliver results. This is again our case with the services and consulting associated to our software ONex for standardized NGS analysis, which we can extend with customized services and consulting. Similarly, our neighbors from Adaptyv Bio at the Biopole in Switzerland sell experimental protein testing as a service, and companies like Tamarind.Bio provide software for molecular structure prediction, analysis and design as services.

Truly Smart AI is as Difficult as Expected, But Context Can Help For Complex Practical Applications

In November 2022, Meta launched Galactica, an LLM supposed to be a game-changer for science thanks to its purported superior thinking capabilities and its being trained on a huge scientific corpus. In just two days, Meta shut it down as users all around the world found just how bad it was in terms of knowledge, problem-solving capabilities, strong biases, and authoritative tone even when wrong. We nowadays know that these are among the biggest problems that affect all LLMs, and companies are very careful in suppressing these problems — to the extent that sometimes, the safety protocols preclude some LLMs from displaying outputs that are actually innocent.

Since the Galactica times, many are working on improving LLMs to make them better at solving problems, while others pursue AGI and other means of artificial intelligence — some even incorporating them into robots. Along the way, means to get the most out of existing systems are explored, for example when Deepmind optimized prompts to largely increase the problem-solving capabilities of publicly available LLMs.

But even the “smartest” AI models are limited by what they know, struggling with tasks that require detailed knowledge of your unique situation, like say the specifics of a project or a business’ goals. Providing the AI model with information together with the request upon prompting can help tremendously; however, many applications may require excessively large contexts. While all LLM creators are working on expanding context windows in parallel to making their systems “truly smarter”, there will always be limits. And, what’s more of a problem, some studies have suggested that information gets “diluted” in very long prompts, making it effectively difficult for the LLM to actually find the important bits of information or the logical patterns that could allow it to produce the right answers.

AI is Transforming How Software is Built And Used, And How Data is Analyzed

AI isn’t just helping science, our daily work and businesses, but it’s also changing the way software itself is created and used, on which all three rely.

With advanced coding tools, developers can work faster and more efficiently, while non-programmers are gaining the ability to create software with minimal technical knowledge. For example, OpenAI’s GitHub Copilot serves as an intelligent pair programmer, suggesting entire functions and debugging code in real time. Similarly, tools like Tabnine and Codeium help automate repetitive coding tasks, enabling developers to focus on creative problem-solving.

And as we covered extensively, AI is making an impact in how we can interact with data and software by alleviating or even removing the need to write any code or instructions in order to achieve results. An example of this is R-Tutor, an AI-driven tool designed to help users learn and apply R programming for data analysis and visualization by casting a user’s request asked in natural language into R code. We also covered in our blog how requests in natural language can be processed by a LLM to assist users of virtual reality systems for molecular graphics.

All of this is leading to two major trends: A boom in small, specialized software tools tailored to specific problems; and a resurgence in custom software development as companies realize how much they can accomplish with AI-powered workflows. As an optimistic result, more businesses will have access to the tools they need, driving innovation across industries. And data analysis and navigation will be more powerful and seamless as ever.

Nexco is Always Looking Ahead

All work and research in science has been severely impacted for positive since two core AI technologies stroke between 2020 and 2022: AlphaFold 2 and GPT-3. And everything that came afterwards, of course.

Two years on, 2024 showed us that AI isn’t just about the technological breakthroughs themselves but also about how businesses and people adapt to those breakthroughs. From falling costs and new competition to smarter tools and services, the AI world is more dynamic than ever.

As we step into 2025, then, one thing is clear: we might be at the highest slope of the S-curve for AI. How will you make AI work for you in the year ahead?