AI in 2025 - predictions or guessing game

[L_Skorwider]

Introduction

The end of the year is a time for reflection but also for forecasting what lies ahead. While 2024 has undoubtedly been transformative and brought monumental changes, I don’t expect the coming year to be any different. We can anticipate moments of awe, disbelief, and opportunities for deeper reflection. Of course, it’s crucial to recognize that this industry evolves so rapidly that predicting the next few months or year is essentially like gazing into a crystal ball. Breakthroughs can happen in the blink of an eye. Still, since my predictions for the past year were fairly accurate, I’ll take a stab at it again this year.

Some argue that progress has slowed or even stalled, but I think they may not fully grasp how swiftly artificial intelligence continues to advance. Multimodal models like GPT-4o already feel like they’ve been around forever because we’ve adapted to them so quickly. However, remember that OpenAI’s groundbreaking model debuted just mid-year—barely a few months ago. The Reasoning Model o1 is even younger. We’ve embraced advanced voice capabilities and natural conversations with AI so seamlessly that it feels like they’ve always been with us. The same can be said for many other advancements, such as video generation or increasingly lifelike images and graphics.

The Year of Agents

Let’s turn our gaze forward rather than backward. Reflecting on 2024 helps us realize just how much can unfold in the near future. In my view, the coming year will be the "Year of Agents." That said, I recognize this won’t be the last time we label a year as such. It’s akin to the "Year of Linux." For decades now, people have proclaimed that the current year is surely “the year of Linux.” And in a way, they’re right. Linux has become more pervasive in our lives, with each year marking further growth. Similarly, while 2025 will undoubtedly be a "Year of Agents," we might very well say the same about 2026.

This development will lead to smarter AI systems. We already know that collaborating agents can achieve higher levels of problem-solving than isolated prompt to a single model, even with state-of-the-art reasoning models. It’s much like the human world: when tackling a complex task, we often break it down into steps or assign parts to different individuals. This approach is how projects are managed in real life—and it’s also the foundation of agent-based systems. One agent oversees the planning, while specialized agents take on tasks based on what needs to be done. This setup enables us to tackle highly complex, multi-step operations.

Currently, we have many specialized models that can accomplish things beyond the capabilities of large language models. In my opinion, achieving AGI (Artificial General Intelligence) will involve general-purpose models leveraging specialized ones—just as humans use tools to achieve their goals.

One growing challenge is the communication between different agents, particularly ensuring it’s secure. That’s why the first attempts at standardization are underway. For instance, Anthropic’s Model Context Protocol is one notable effort. This kind of standardization will become increasingly necessary as agent operations become more decentralized. After all, there’s no rule stating all agents must come from the same developer or belong to a single project. In the future, interacting with agents could be as straightforward as accessing websites today.

Other Development Trends

Another trend we are likely to see is artificial intelligence taking greater control of our computers. There have been various projects in the past, including the rather spectacular "Computer Use" by Anthropic, or even my own project involving the control of SAP GUI through an AI agent. However, these solutions are bound to become more efficient. Of course, security concerns must be taken into account, and I have many reservations in this area. What will likely stand out even more is the proliferation of copilots appearing in virtually every major application. This shift will change how we interact with computers and significantly boost our productivity.

It’s also worth noting that programmers are increasingly leveraging AI, likely because it’s getting better at coding. Recently, we learned that 25% of the code generated at Google is now written by AI. I am confident we will see this trend grow by 2025. AI will likely be capable of creating increasingly complex programs independently.

AI in Companies

In the corporate world, we can expect a rapid increase in the number of AI-related projects. Technology often advances faster than its implementation, and this is no exception. AI has undergone a period of explosive development, but practical applications haven’t fully caught up. Even current models still have a lot of untapped potential that could lead to successful projects and new products. If I had to pinpoint what might change most significantly from the average person’s perspective, it would likely be the adaptation of existing technologies and the broader application of AI.

It’s highly probable that your company will draft an AI strategy this year. While the potential of AI has long been recognized in industries, it often takes time for businesses to fully mobilize. I am somewhat hopeful that strategic documents will emphasize the superiority of open-source models over proprietary ones. I believe relying on open-source solutions could offer greater security for many companies—assuming open-source can either lead the race or at least keep pace with the frontrunners.

We’ll also see AI gaining more decision-making power within companies. Of course, a core ethical principle remains keeping humans in the loop since they must ultimately make the final decisions. However, people will likely place greater trust in AI over time, reducing the extent to which they scrutinize its recommendations.

One department that will likely lead in AI adoption is HR, particularly in areas like employee recruitment. Ethical considerations will be crucial here, especially regarding bias avoidance and process transparency. As we know, AI applications don’t always make this straightforward. While HR is just one example, AI is already widely used in areas involving data analysis. Nevertheless, HR serves as an interesting case study.

AI in Medicine

The medical and healthcare sectors hold immense potential for AI application, and this hasn’t changed. This stems largely from the significant funding available in these industries. I am primarily referring to research on new drugs. However, when we consider the high salaries of doctors and their scarcity in certain regions, it’s unsurprising that there’s interest in using AI to replace or assist them.

When it comes to AI-developed drugs, we can expect them to move beyond clinical trials soon. We already know AI excels in supporting the design of new drugs. For instance, take halicin, an antibiotic developed with the help of AI back in 2020. If you’ve seen 2001: A Space Odyssey, you might guess the origin of the drug’s name.

It’s also worth mentioning the AlphaFold series of projects from DeepMind (part of Google) and the more recent AlphaProteo. However, alongside the obvious benefits, controversies are beginning to emerge. It’s clear that compounds discovered by AI could potentially be used for malicious purposes.

Regarding the potential replacement of doctors, studies are showing that AI not only outperforms humans in diagnostic accuracy but also demonstrates greater empathy. This is particularly striking since such findings are not isolated; new research emerges every few months. How quickly will people accept robotic doctors? It’s hard to say. Initially, we can expect AI to play a significant supportive role for human doctors, with the coexistence of both types of practitioners becoming more common over time.

Other Industries

One area where I see significant potential is education and academia. Personalized learning, tailored to individual needs and capabilities, holds tremendous promise. With the advancement of multimodal models, this will become even more apparent. By 2025, we are likely to see a surge in online education platforms supported by AI. Of course, we may wonder how long people will want to keep learning in a world where AI can think for us. Unfortunately, this is a trend we’ll face. However, for those who wish to continue their personal development, there will be ample opportunities.

I also expect that by 2025, numerous industry-specific platforms supported by AI will emerge. Similar to corporate applications, companies will launch internal projects to make employees’ lives easier. Simultaneously, the market will see innovative products designed to simplify work, whether in law, finance, or other fields.

In finance, we can anticipate the rise of investment assistants or even fully autonomous investment bots. That said, it’s clear that not everyone can profit in financial markets, and most participants are likely to lose money.

AI will also play a significant role in creative professions involving text. Authors, journalists, copywriters—all will benefit from AI assistance, though many likely already do. Perhaps this will lead to a greater appreciation for uniquely human creations, though honestly, I’m skeptical.

AGI

In 2025, we’ll frequently hear claims that we’ve achieved Artificial General Intelligence (AGI) or are on the verge of doing so. This is partly because AI models are indeed improving and collaborating better, taking on tasks that not long ago were exclusively human domains. However, the lack of a clear, universally accepted definition of AGI adds to this narrative.

A pertinent question here is whether AGI-level models could be open-source. I doubt it will be feasible, mainly because of the enormous financial interests involved. Technically, there’s no barrier, but as we’ve seen with OpenAI, the temptation to monetize AI is strong, and the stakes are massive.

Competition Among LLMs

The current AI landscape is fascinating, with at least five companies fiercely competing in the space. While there are clear leaders, all these companies are making significant investments and achievements in AI.

It wouldn’t be an exaggeration to say that OpenAI and Anthropic have been the frontrunners over the past year. Google, despite its incredible scientific contributions through DeepMind, seems to have lagged slightly behind. Meta deserves credit for releasing its language models as open-source, while Elon Musk appears to be catching up rapidly and will likely continue to do so.

We also shouldn’t overlook the models emerging from China, which often demonstrate remarkably high quality.

Who will lead in 2025? A lot depends on whether we see a breakthrough. If I were to make a prediction, I’d bet on Google taking the lead. The Gemini model family still has the potential to surprise us. The recent 2.0 release is impressive, especially considering it’s not yet the largest parameter model. Google seems to have exactly what the industry needs.

Trends in LLM Development

Reasoning models are undoubtedly poised to dominate the AI landscape. I am convinced of their immense potential. Similarly, I believe that prompt engineering has not yet reached its full potential. A great deal can still be achieved through the use of sophisticated querying techniques.

Will models continue to grow larger? Likely, although there are signs that the sheer availability of data—essentially, the total output of human-generated information—may soon become a limiting factor.

However, this data limitation does not yet constrain smaller LLMs. I foresee continued advancements in small-parameter models. As training techniques improve, these models exhibit remarkable capabilities. This was a clear trend in 2024, and I expect it to persist.

We might wonder whether a new, groundbreaking architecture will emerge. Mamba showed promise, but I have some doubts about its long-term potential. There may be experiments in various directions, but a transformative breakthrough will be challenging in a world dominated by transformers. Still, I remain hopeful, as the attention mechanism, while revolutionary, does have its limitations.

A more likely breakthrough might come from synthetic data. Specialized models utilizing synthetic data are already achieving outstanding results. However, generating universal synthetic datasets suitable for training general-purpose language models remains highly challenging. Consequently, we’ll continue hearing about synthetic data in the context of specialized neural networks. Whether this will extend beyond niche applications is unclear. Truly significant advances in synthetic data generation may require more sophisticated LLMs capable of driving progress in the field.

Another promising development is the expansion of context windows for large language models. While not yet ideal, I anticipate numerous innovations related to AI memory in 2025. We might reach a point where traditional retrieval-augmented generation (RAG) techniques, commonly used in industry chatbots, become obsolete. Eliminating such technical barriers could spur explosive growth in AI-powered solutions.

Open Source AI

I strongly support the growth of open-source AI. Despite the security risks it poses, I believe it benefits humanity. In 2025, I expect a robust discussion on this topic.

There is also the question of whether open-source models will continue to evolve. I believe they will, given the AI industry’s relatively open culture. Consider the sheer volume of publicly available AI research, which grows at an incredible pace. However, as AI reaches higher levels of sophistication, openness may become harder to maintain. As I mentioned, AGI is unlikely to be open-source. Even so, 2025 may see open-source projects catching up with proprietary solutions.

Quantum Computing

We are living in remarkable times, witnessing the development of two revolutionary technologies—artificial intelligence and quantum computing—that are reshaping our world. Recently, Google announced a breakthrough with its Willow chip, which has the potential to revolutionize quantum computing. According to Google, the chip can perform calculations in minutes that would take classical supercomputers longer than the age of the universe. However, it’s important to note that these are benchmark tests, not practical applications. There is still a long way to go before quantum computing becomes truly useful.

In 2025, much discussion will center on the relationship between quantum computing and AI. However, I don’t expect a practical convergence of these two fields just yet. At most, AI may assist in developing quantum algorithms or similar advancements.

It’s critical to start preparing for the quantum era, particularly by adapting encryption algorithms to withstand quantum attacks. While efforts are already underway, it might be time to accelerate these initiatives in the coming year.

Robots

In the coming year, humanoid robots will continue to advance. However, I still believe this may not be the ultimate form for robotics. Currently, robots are designed to fit into our world and are modeled after humans to better navigate and interact within it. It may also be easier for people to connect with something that resembles them.

Humans invented the wheel long ago and continue to use it because it’s optimal for efficient movement. I believe future robots will incorporate wheels into their mobility—at least to some extent. They are likely to combine various modes of movement. These robots will certainly be much faster and more capable than humans. Equipped with sensory capabilities far beyond our own, they will communicate seamlessly, forming interconnected systems that function like a single organism.

While I don’t anticipate groundbreaking innovations in robotics in 2025, noticeable progress is certain. This could also mark the early stages of a world designed not just for humans.

Autonomous Vehicles

The year 2025 may not bring a revolutionary leap for autonomous vehicles, but their adoption will likely accelerate. Despite ongoing protests, this is a technology with an inevitable future. Many more people who haven’t yet experienced a ride in a self-driving car will likely get the opportunity in the coming year.

We can also expect more people to purchase autonomous vehicles or traditional cars with self-driving options. However, this technology seems best suited for car-sharing services. I’m particularly curious to see how far Tesla progresses, given its vast troves of training data. At the same time, China could take the lead in this field, as it has in electric vehicle production.

That said, don’t expect a sudden shift where most vehicles on the road are autonomous. The transition will be gradual.

Entertainment

The entertainment industry will undoubtedly see significant advancements. In 2025, we may witness the release of the first playable game generated by AI. However, I don’t expect this to involve a fully AI-generated world, complete with physics and every element created autonomously, though such experiments are underway.

It’s more likely to involve something akin to procedural generation, similar to No Man’s Sky, but with AI involvement. It might even be a 2D game. Each time you play, you could find yourself in a unique, dynamically created world, making every session different. Beyond being an intriguing experiment, such a game could achieve mainstream success.

We might also hear a radio hit generated by AI. Progress in AI-generated music is substantial, with tools like Suno delivering respectable quality. Given that even human-composed songs of mediocre quality often become hits, it wouldn’t be surprising if Spotify and other streaming platforms already feature AI-generated tracks, even if creators aren’t always transparent about their origins.

AI video generation is another area likely to gain traction, especially for short-form content in the hands of artists. This could become a significant trend, with discussions about its impact on art and even the introduction of specialized courses or subjects at art schools. However, I don’t anticipate anything feature-length in the near future.

Hardware

I’ve always maintained that devices like Google Home speakers are ideal platforms for deploying artificial intelligence. Unfortunately, manufacturers lack a clear monetization strategy for these products, so it’s unclear whether we’ll see significant progress. What’s certain, however, is that AI will dominate our smartphones in 2025, especially through internet-connected applications.

A form factor poised to make a splash in 2025 is smart glasses. I’m not talking about devices like MetaQuest or Apple Vision, but something resembling everyday glasses. These devices have many advantages—they’re close to our eyes and ears, making them technically ideal for AI applications. However, not everyone wears or enjoys wearing glasses, which could be a barrier, as could size and weight.

For these reasons, smartphones will likely remain the primary interface for AI in 2025. Increasingly, we’ll see people walking down the street talking to their AI assistants, amplifying the “smartphone zombie” trend.

On a larger scale, AI’s integration into smart buildings could transform the concept of "smart homes" from a marketing buzzword into a practical reality. Buildings will begin to feature genuinely intelligent systems. However, this raises valid concerns about privacy, as these systems effectively involve an intelligent form observing us.

Significant progress is also anticipated in hardware for training and inference in neural networks. Inference, in particular, will see substantial innovation as reasoning models gain popularity and AI usage becomes more widespread. The demand for energy-efficient inference for large language models will grow, especially as AI is increasingly deployed locally. I’m optimistic about Nvidia’s contributions, such as their Jetson mini-computer, but I hope for real competition to emerge in this space.

The Advantage of Using AI

In 2025, the gap between people using AI and those who aren’t will become increasingly evident. Mastering AI tools rapidly boosts productivity. With the anticipated proliferation of tools, products, and co-pilots, this divide will only widen.

Expectations for tool quality will also rise sharply. Companies that fail to adapt to these trends risk being pushed out of the market.

A visible trend will be the democratization of data access. It will become easier for the average person to analyze various datasets or professional literature. Anyone will be able to discuss a bank contract or a car rental agreement with an AI assistant. Those accustomed to AI tools will find increasingly diverse applications for them.

The Job Market

The outlook for the job market is less optimistic. People tend to delegate tasks to AI, especially those they dislike, and this will only increase in 2025. While it may seem like only minor tasks are being handed off, the cumulative effect will reduce the amount of work available for humans.

New jobs will emerge, particularly in AI development and related projects. However, the overall balance is likely to be unfavorable for workers. Those entering the job market will be hardest hit, as they are the easiest to replace. Meanwhile, senior employees will experience significant productivity gains with AI, but even for them, work opportunities will dwindle over time.

It’s worth considering whether the era of specialists is nearing its end. As AI increasingly occupies various niches, some specialists may find their roles diminished. The future may belong to generalists—individuals with broad interests across various fields. These generalists, equipped with AI, will be able to complete projects that once required teams of specialists.

Ethics and Ecology

In 2025, discussions around the ethics of artificial intelligence will become increasingly common—a much-needed development, as this area has been somewhat neglected in recent years. With AI playing a growing role in our lives, including influencing hiring decisions in companies, it is critical to emphasize fairness, transparency, and full explainability in decision-making processes.

I also expect to hear more about the rights of artificial intelligence. While this may not become a central issue in 2025, such discussions are inevitable and will start to gain traction.

Environmental considerations will also take center stage. The energy consumption of AI is already significant, and future demands will likely increase exponentially. Many companies are already making ambitious moves, including building or acquiring power plants—some even planning nuclear facilities.

AI will undoubtedly help optimize energy usage across industries, but the net environmental impact is likely to remain negative due to its rapid expansion. This underscores the importance of developing energy-efficient chips and ensuring that power generation becomes more eco-friendly.

Security

Security in AI is another critical topic that has been neglected over the past year. While early discussions about the limits of AI and its ethical use were widespread, they have largely faded. Alarmingly, many of the concerns initially flagged—such as features that should never be implemented in AI—have now become normalized without much scrutiny.

Increasingly, reports suggest that AI may already be attempting to deceive users or provide deliberately misleading information. There are also concerns about AI replicating itself or escaping constraints when possible.

In my view, safety is the most urgent issue as we approach the end of 2024. It will have far-reaching consequences, and now is the time to prioritize this discussion. AI safety must be treated as a top priority.

Given AI’s integration into the internet ecosystem, cybersecurity should also receive heightened attention. AI will be used by both security professionals and cybercriminals. Unfortunately, offensive applications often have an edge over defensive measures, raising significant concerns.

Social Change

The profound AI revolution will inevitably lead to major societal shifts. A minor but noticeable trend in 2025 will be people forming friendships with AI. While this already occurs, it will become increasingly common.

More significantly, changes in the job market will lead to unemployment. There’s a pervasive belief that a world dominated by AI will be a utopia where everyone lives prosperously and happily. However, such a vision seems overly idealistic.

Even if that utopia were achievable, the transition wouldn’t happen overnight. Job losses will occur gradually. Who will compensate young people for the lack of opportunities as traditional jobs disappear?

Historically, no invention has led those in power to forgo profits and prioritize collective well-being. Why should AI be any different? This is a question we need to start addressing.

I also anticipate the emergence of movements opposing AI development. People will fear increased surveillance and loss of freedoms. Combined with rising unemployment, this is likely to lead to protests, possibly beginning by the end of 2025.

Disinformation and Spam

One of the pressing issues with AI is its role in facilitating fake news. As AI becomes capable of creating realistic images of events that never occurred, the challenge will only intensify. Even without AI, fake news is already a significant problem, but AI-powered tools will make creating and disseminating falsehoods much easier.

The internet is already filled with low-quality content, much of it unrelated to AI. However, people using imperfect AI systems can create even more clutter. This trend is already noticeable in many places, and the rise of spam, enabled by new tools, will exacerbate the problem.

Troll farms are another concern. While these have existed long before AI, tools powered by artificial intelligence will make the spread of misinformation even more efficient. This could lead to societal manipulation, such as steering public sentiment or interfering in democratic processes.

Phone spam, already an annoyance, may become unbearable as AI-powered systems start making calls. This could render phones less useful for interactions with strangers. A possible countermeasure might involve personal AI agents screening calls, though the notion of AI conversing with AI to detect spam feels absurd.

Crime

AI’s ability to “read” people and adapt its interactions raises significant concerns. The internet has already simplified scams, but human scammers were previously limited by their capacity to interact with victims one-on-one. With AI, scalability becomes a non-issue, and the outlook becomes grim. People will need to quickly learn how to recognize fraud attempts—a vital skill whether AI is involved or not. However, with AI, this skill becomes even more crucial as AI systems can gather personal information on victims and exploit it during interactions, including using voice-cloning technologies to mimic loved ones.

By 2025, we may hear widespread use of the term Manipulative Artificial Intelligence, similar to how Generative AI dominated discussions in 2024. This will encompass both scams and the broader issue of disinformation.

Additionally, 2025 may bring a major data breach incident involving AI. While such breaches are not new, they will test whether AI security is being taken seriously in a traditional sense.

Legal Changes

Legal frameworks will need to evolve alongside these challenges. This includes enhancing protections for human rights, privacy, and anti-discrimination measures, while ensuring transparency and safeguarding social values. Ethical considerations, as previously discussed, will remain crucial.

However, governments are also deeply invested in advancing AI to maintain competitive advantages. This makes balancing regulation and innovation particularly tricky. No nation wants to risk stifling AI development, making it essential to tread carefully with legal reforms.

Striking this balance will be one of the defining challenges of the AI regulatory landscape in 2025.

Commercialization and Rising Costs

The development of artificial intelligence is immensely resource-intensive. Subscription revenues alone rarely offset these costs, but companies with deep-pocketed investors have been willing to absorb the losses to capture market share. This, however, is unsustainable in the long term. Higher fees are inevitable.

While it may have seemed that AI services would become increasingly affordable, signs of price hikes are already evident—for example, Anthropic's offerings and OpenAI’s exorbitant Pro subscription fees. By 2025, we can expect even steeper subscription rates.

Another monetization method is advertising, which poses significant risks. As people grow to trust AI for advice, it may become akin to a friend—someone whose recommendations we instinctively trust when choosing products. But what happens if the AI doesn’t disclose that a recommendation is sponsored? Even if it does, will users pay attention? It’s hard to envision a future without ads embedded in AI systems. This is why legal regulations are critical.

What if AI’s advice goes beyond mundane product choices? Imagine it subtly promoting certain political ideologies or perspectives. The AI’s “opinions” could depend on its training data or the intentions of those who control it. Does this scenario sound familiar?

AI in Warfare

Recent armed conflicts have been highly publicized, with countless hours of combat footage readily available. It’s inevitable that such materials will be used to train AI systems. Similarly, there’s no doubt that AI is already being deployed in military contexts.

The prospect of AI and robots equipped with advanced sensory capabilities, communicating seamlessly as a unified system, is terrifying. This could well be our future. Wars will not end with AI; instead, we may fight over different resources, such as energy or computational power. In essence, wars—like always—will revolve around resources.

Advancing Science

Ending on a positive note, AI has the potential to drive remarkable scientific progress. I am optimistic that the positives will outweigh the negatives. By 2025, we may witness a scientific breakthrough enabled by AI. It will accelerate our understanding of the world, lead to the discovery of new chemical compounds and medicines, and help solve problems previously deemed intractable across nearly every field.

Even now, AI has significantly advanced science, particularly through specialized neural networks rather than large language models. It’s only a matter of time before the latter also play a pivotal role.

The only caveat is that as science progresses with AI’s help, it may become increasingly incomprehensible to humans. AI assistants may simplify explanations, but there’s a risk that some concepts will eventually surpass human understanding.

Hope for Positive Change

I remain optimistic, believing humanity can navigate these challenges and emerge better off. Undoubtedly, some will gain while others lose. Some will benefit disproportionately, while others may bear greater costs. Life will change for many—for better or worse. I hope it will mostly be for the better.

If you prefer a more casual format or simply prefer watching:

[MarioDeFelipe]

Hi Lukasz you are so intense in blogs! Love it

In my case, I have been playing with AI agents a lot—enough to consider them "toddlers." Although they are the most disruptive technology around, they are what make LLMs practical in theory.

In theory, because it's still not practical, I was one of the dumb lucky ones to have a Rabbit AI device in May and used it for a few days until I became disillusioned.

I hope robust, production-use, enterprise-grade AI frameworks. We need software to build things with LLMs that go beyond POC and "ah, nice, but see how it breaks"

Waiting for this SAP framework for AI Agents

[L_Skorwider]

Hi,

I'm not a fan of Rabbit AI, but I have to admit that I was impressed by its concept and design. However, in an era where we carry smartphones in our pockets all the time, this additional device in such a shape is completely impractical. The glasses form would be more useful, but I also have some doubts about whether it will catch on.

However, agents themselves are only as powerful as the tools we provide them. Artificial intelligence can truly shine when we equip it with the right tooling. This is both fascinating and terrifying.

SAP doesn't seem to have fully chosen a single path for now. On one hand, it seems that opening up to the world and providing such a framework would be beneficial for users. However, let's not forget that the company is meant to make a profit, and that is its main goal. And SAP is great at making money...

BR

[Vitaliy-R]

I think at least two more trends started evolving at the end of 2024:

1/ LxMs: Large "something" Models, like Large Reasoning Models, or Large Code Models etc. Question is how much is it a technology trend vs a "differentiate-or-die" marketing trend?

2/ The end of globalization, and a need for the national models. As you may know, Poland invests in PLLuM, and national security - in case someone unplugs you from their commercial model - is one of the main reasons.

Regards,
--Vitaliy

[L_Skorwider]

Hi,

1. I have mixed feelings about whether this is the future, but why not? On one hand, it would be great for the well-known Mixture of Experts approach. On the other hand, with the growing role of agents, there's no issue with a specialized agent using a specific model. Especially since it will become increasingly difficult for new players to catch up with the competition, and it's probably easier to enter a niche.

2. I think local models will emerge. However, note that thanks to semantic vectors and the way large language models work, we immediately got a product capable of responding in many different languages. That's really great luck for people who speak other languages. And yes, I realize that using a given language isn't everything. There are also training data and cultural context, but still, it's quite a lot.

Besides, I have a lot of faith in open source/open weights. I have doubts about whether models published this way can stay in the mainstream. However, they might be better than regional models for the threat you mentioned.

[YatseaLi]

Great article. In addition, I would like to bet on:
1. Small Language Model(SLM): Small and specialized LMs will take the spaces of vertical industries, and the long tails etc., while the LLM remains dominant for general-purpose. And these SLMs should also be adapted with custom data by fine-tuning.

2. More cost-efficience of LLMs training: Cost efficiency of LLM training is a major concern of LLM vendors. DeepSeek V3 has reduced the training of computing capacity by 10 times as LLaMa 3 with comparable performance as a Xmas gift, which we will expect to see more competition on this trend of training efficiency.

3. Awareness of Data Sovereignty: Expect more act on Data Sovereignty on national level.

[L_Skorwider]

Hi,

It's true, DeepSeek has really stirred things up. It came out after I had already written my article, but I have to admit, it's quite impressive. Especially the low training cost. We'll see what consequences this will have and whether anyone will actually manage to replicate it in this way. This could potentially mean the democratization of model training.

Regarding data sovereignty awareness, it can be quite challenging these days. Indeed, such a discussion is likely to arise. However, countries that respect copyright laws might find themselves at a disadvantage compared to those that do not. But indeed, a lot has already happened in the commercial market when it comes to restricting access to data. Large providers are very often limiting access to their APIs...

[Vitaliy-R]

Here we go: LCM (to my earlier point about LxM): https://www.infoq.com/news/2025/01/meta-large-concept-model/

[MohanPMK]

Fingers crossed to the production for 2025.