1. Meta will initiate fees for utilizing its Llama models' services.
Anticipated AI Trends in 2025: A Forecast
Meta, renowned for its open-source AI leadership, has been providing its advanced Llama models gratis. However, the forthcoming year might witness a significant change as Meta begins charging corporations for utilizing Llama.
It's crucial to clarify that we are not anticipating Meta to transition Llama into a closed-source model or obligate users to pay for accessing it. Rather, we anticipate that Meta will incorporate more stringent terms into Llama's open-source license, mandating companies engaging in commercial applications above a particular scale to procure a fee to access the models.
Although Meta has already implemented this concept in a limited manner, only denying access to Llama to the largest companies (e.g., cloud hyperscalers and companies with more than 700 million monthly active users) without prior consent.
Back in 2023, Meta CEO, Mark Zuckerberg, indicated advocacy for this approach, stating, "If you're someone like Microsoft, Amazon or Google, and you're going to basically be reselling [Llama], that's something that we think we should get some portion of the revenue for. I don't think that that's going to be a large amount of revenue in the near-term, but over the long term, hopefully, that can be something."
In the upcoming year, Meta is expected to expand this payment plan to include more substantial and medium-sized enterprises.
Why would Meta implement such a strategic shift?
Maintaining Llama at the forefront of AI advancements requires substantial financial investment. Meta will require billions in funding annually to maintain parity with OpenAI, Anthropic, and other cutting-edge AI labs.
Despite being one of the world's largest and wealthiest companies, Meta as a publicly-traded entity is accountable to its shareholders. As the expense of building advanced AI models escalates, it becomes increasingly unfeasible for Meta to allocate such vast resources to training subsequent-generation Llama models, expecting no return on investment.
Hobbyists, researchers, individual developers, and startups will continue to make use of Llama without charge in 2025. However, 2025 will mark Meta's robust commitment to monetizing Llama.
2. Scaling laws will be unearthed and exploited beyond language applications, particularly in the domains of robotics and biology.
No topic in AI has generated as much debate lately as scaling laws – and the speculation surrounding the attainment of their end.
Introduced in a 2020 OpenAI paper, the fundamental principle behind scaling laws is straightforward: as the number of model parameters, the volume of training data, and the compute resources grow when creating an AI model, the model's performance improves (test loss decreases) consistently and predictably. Scaling laws have contributed to the phenomenal performance increase from GPT-2 to GPT-3 to GPT-4.
Scaling laws are not laws but rather empirical observations. Over the past month, several reports have suggested that major AI labs are encountering diminishing returns when scaling large language models further. This explains, for example, the constant delays in OpenAI's GPT-5 release.
The common response to this plateau in scaling laws is that the emergence of test-time compute offers an entirely new dimension to pursue scaling. Instead of extensively scaling compute during training, recent developments like OpenAI's o3 enable massively scaling compute during inference, unlocking novel AI possibilities by allowing models to "think for longer."
This is indeed a significant point. Test-time compute provides an intriguing new avenue for scaling and improving AI performance.
However, there is a more crucial yet underappreciated aspect of scaling laws in today's discourse: nearly all discussions regarding scaling laws (from the initial 2020 paper to the recent focus on test-time compute) revolve around language. Language, however, is not the only data modality that deserves attention.
Consider robotics, biology, world models, or web agents. For these data modalities, scaling laws are far from saturated; in fact, they are just commencing. Even rigorous evidence demonstrating the existence of scaling laws in these domains remains elusive.
New foundation model startups focusing on these emerging data modalities (e.g., EvolutionaryScale in biology, Physical Intelligence in robotics, World Labs in world models) aim to identify and capitalize on scaling laws in these fields, much like how OpenAI successfully profited from LLM scaling laws in the early 2020s. In 2025, expect breakthrough progress here.
Do not be misled. Scaling laws are not vanishing. They will remain essential in 2025. However, the focus of scaling law activities will shift from LLM pretraining to other modalities.
3. A discord between Donald Trump and Elon Musk is imminent, with ramifications for the AI world.
With a new U.S. administration expected, policy shifts in AI are inevitable. When predicting the AI winds under President Trump, it is alluring to contemplate the president-elect's close relationship with Elon Musk, given Musk's prominence in the AI sector.
Visualize various ways Musk could sway AI advancements under a Trump administration. Given Musk's severe animosity towards OpenAI, the new administration might adopt a colder stance towards OpenAI during industry engagements, AI regulation drafting, government contract awarding, and more. This is a genuine worry for OpenAI at present. On the other hand, the Trump administration could favor Musk's own enterprises, such as reducing regulations for xAI to construct data centers and gain an edge in the frontier race; expediting regulatory approvals for Tesla to deploy robotaxi fleets, and so forth.
Elon Musk, unlike numerous tech leaders with Trump's ear, pays close attention to existential AI risks and advocates for stringent AI regulations. He backed California's contentious SB 1047 bill, aiming to impose strict restrictions on AI developers. Musk's influence might create a more stringent regulatory environment for AI in the U.S.
However, there's a drawback to all these assumptions. Trump and Musk's camaraderie is as fragile as glass and destined to disintegrate.
Throughout the initial Trump administration, Trump allies, even the most loyal ones, had dismal tenures. Names like Jeff Sessions, Rex Tillerson, James Mattis, John Bolton, Steve Bannon, and Anthony Scaramucci come to mind. Few of Trump's earlier aides still stick by his side today.
Both Trump and Musk wield powerful and unpredictable personalities. They demand substantial effort and can burn people out. Their unlikely friendship has profited them, but it's still in its initial stages. Our prediction is that before 2025 draws to a close, this relationship will sour.
This development could have significant consequences for the AI realm. OpenAI will breathe a sigh of relief. Tesla shareholders, meanwhile, will lament. And individuals concerned with AI safety will be disheartened, as the U.S. government will likely maintain a hands-off approach to AI regulation under Trump.
4. Web agents will gain mass appeal, transforming into the following major advancement within consumer AI.
Picture a life where you never have to interact personally with the web. Whenever you need to manage a subscription, pay a bill, schedule a doctor's appointment, order something from Amazon, make a restaurant reservation, or tackle any other frustrating online task, just tell an AI helper to handle it for you.
The concept of a "web agent" has been around for years. If a functioning general-purpose web agent existed and worked effectively, there's no question it would be highly successful. Yet, no such tool is available in the market today.
Startups like Adept, boasting a distinguished founding team and vast funding, have become cautionary tales in this area.
2023 will be the year web agents finally become functional enough to become mainstream. Advancements in language and vision foundation models, coupled with recent breakthroughs in "System 2 thinking" capabilities resulting from new reasoning models and inference-time compute, will make web agents primed for popularity.
(In essence, Adept had the right vision; it just arrived a bit premature. Timing is crucial in startups, just as in many other aspects of life.)
Web agents will discover a myriad of valuable enterprise applications, but we believe their most significant near-term market opportunity will be with individual consumers. Despite all the recent AI hype, AI-native applications beyond ChatGPT have yet to break through as mainstream consumer successes. Web agents will change this narrative, becoming the next true "killer app" in consumer AI.
5. Multiple serious initiatives to put AI data centers in space will be initiated.
In 2023, the limiting factor for AI growth was GPU chips. By 2024, it became power supply and data centers.
Few topics captivated the public's attention in 2024 like AI's massive and swiftly escalating energy demands, fueled by the global AI data center expansion. After remaining unchanged for decades, global power consumption from data centers is projected to double between 2023 and 2026 due to the AI boom. In the U.S., data centers will consume nearly 10% of all power by 2030, rising from just 3% in 2022.
Today's energy infrastructure is unable to cater to the substantial surge in demand from artificial intelligence workloads. An imminent clash between these two multi-trillion-dollar systems—our energy grid and our computing infrastructure—was almost inevitable.
Nuclear power has gained traction this year as a potential solution to this problem. Nuclear power offers many advantages for AI: it's zero-carbon, available round-the-clock, and nearly inexhaustible. However, new nuclear energy sources will not address this issue until the 2030s, given lengthy research, project development, and regulatory timelines. This applies to traditional nuclear fission power plants, modern "small modular reactors" (SMRs), and undoubtedly, nuclear fusion power plants.
In 2024, an unconventional idea to tackle this challenge will surface and catch the public's attention: placing AI data centers in space.
AI data centers in space—initially, this may seem like a far-fetched joke inspired by a VC's blend of startup buzzwords. However, there might be something to it.
Earth's major barrier to swiftly constructing more data centers is obtaining the necessary power. A computational setup in orbit can profiteer from free, boundless, carbon-neutral power 24/7: space is perpetually sunny.
Another encouraging aspect of relocating computing to space is addressing the heat issue. Among the primary engineering hurdles in building more potent AI data centers is the reality that overheating happens when multiple GPUs run simultaneously in a confined space. Excessive heat can cause damage or destruction of computing equipment. Data center architects are adopting costly and uncertain methods like liquid immersion cooling to tackle this dilemma. Space, however, is extremely frigid; any heat generated through computing operations would swiftly and unharmed dissipate.
Certainly, numerous practical challenges should be overcome. One apparent issue is the cost-effective transfer of massive volumes of data between orbit and earth. While this is still unresolved, there might be solutions coming up through the use of lasers and other high-bandwidth optical communication technology.
Recently, Lumen Orbit, a startup from Y Combinator, secured $11m to pursue this objective: constructing a multi-gigawatt network of data centers in space for AI model training.
Lumen CEO Philip Johnston explained it thusly: “Instead of spending $140 million on electricity, you can invest $10 million in a launch and solar power.”
Lumen will not be the sole entity taking this approach seriously in 2025.
Additional startup rivals will emerge. It wouldn't be surprising to see some of the cloud hyperscalers initiate exploratory ventures in the same vein. Amazon has prior experience in deploying assets into orbit via Project Kuiper; Google has championed moonshot concepts such as this for years; even Microsoft possesses expertise in the space economy. Elon Musk's SpaceX could potentially enter the fray as well.
6. An AI system will pass the “Turing test for speech.”
The Turing test, one of the oldest and most renowned AI assessment measures, dictates that an AI system must be able to exchange text via writing in such a manner that the average human is unable to distinguish between an AI and another human.
Thanks to recent significant advancements in large language models, the Turing test has become solvable in the 2020s.
However, humans communicate through more than just writing.
As AI evolves to be multimodal, a challenging new version of the Turing test – a "Turing test for speech" – may emerge: an AI system that can interact with humans via voice with such skill and fluidity that it becomes indistinguishable from a human speaker.
The Turing test for speech remains beyond reach for today's AI systems. To accomplish this, significant additional technological and commercial developments are needed.
Latency (the lag between when a human speaks and when the AI responds) must be reduced to almost nothing to replicate the experience of interacting with another human. Voice AI systems must improve at handling ambiguous inputs or misunderstandings in real time – for instance, when interrupted mid-sentence. They must possess the ability to engage in extended, multiturn, open-ended conversations while retaining previous conversation topics in memory. And crucially, voice AI agents must develop a better comprehension of non-verbal signals in speech – such as distinguishing between an annoyed, excited, or sarcastic human speaker's tone – and generating those non-verbal cues in their response.
Voice AI has reached an exciting turnaround point as we near the end of 2024, driven by fundamental breakthroughs like the emergence of speech-to-speech models. This area of AI is advancing rapidly, both technologically and commercially. Anticipate the state of the art in voice AI to progress significantly in 2025.
7. Major progress will be made on building AI systems that can themselves autonomously build better AI systems.
The idea of AI that can create better AI is a recurring theme in AI circles dating back decades.
In 1965, for instance, Alan Turing's collaborator I.J. Good wrote:
“Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man, however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind.”
The concept of AI constructing better AI may seem like science fiction, but it is becoming increasingly plausible. Researchers at the forefront of AI science have started making tangible progress in developing AI systems capable of autonomously building better AI systems.
We predict that this direction of research will gain mainstream recognition next year.
To this point, the most prominent public instance of research in this area is Sakana's "AI Scientist." Published in August, the AI Scientist study represents a compelling demonstration that AI systems can, in fact, conduct AI research unassisted by humans.
Sakana's AI Scientist oversees the entire AI research lifecycle: reading existing literature, producing research ideas, devising experiments to test those ideas, executing those experiments, writing and publishing research papers on the findings, and carrying out a peer review process on its work. It performs this task autonomously, without any human input. Some of the research papers created by the AI Scientist are available to read online.
Gossips circulate that innovative AI labs like OpenAI, Anthropic, among others, are putting resources towards the concept of "automated AI researchers," however, no official confirmation has been made.
anticipate an escalation in this field's discussion, advancement, and startup activity in 2025, as the realization sets in that automating AI research is becoming a tangible reality.
The most significant achievement, though, will likely occur when an AI-generated research paper is accepted into a premier AI conference for the first time. Given the anonymous review process of these conferences, the authors' identities remain concealed until after acceptance. It's not unforeseen that AI-produced research will be approved at NeurIPS, CVPR, or ICML in the upcoming year. This will be a captivating, contentious, and groundbreaking moment for AI research.
Visionaries such as OpenAI and Anthropic will start "climbing the ladder," gradually shifting their strategic emphasis to the development of applications.
Establishing frontier models is a financially demanding endeavor. Frontier model labs consistently burn through substantial amounts of capital. OpenAI raised a record $6.5 billion in funding recently, and they'll likely need even more down the line. Similarly, Anthropic, xAI, and others find themselves in similar financial predicaments.
Switching costs and customer loyalty are minimal. AI applications are typically designed to be agnostic to the underlying model, allowing seamless substitution of models based on cost and performance considerations.
Moreover, there is a constant threat of technology commoditization, with powerful open models like Meta's Llama and Alibaba's Qwen becoming increasingly prevalent.
Despite their ongoing commitment to creating advanced models, AI trailblazers like OpenAI and Anthropic will soon prioritize generating revenue streams that are more profitable, distinct, and customer-retaining. In 2025, expect them to intensify their efforts to launch more applications and products.
Obviously, ChatGPT is a success story of an application developed by a frontier lab.
What other initial applications might we see from AI labs in the upcoming year?
Advanced and feature-rich search solutions are an obvious possibility. OpenAI's SearchGPT initiative hints at this.
Coding is another evident category. Once again, preliminary product development is already underway, with OpenAI's canvas product launching in October.
Will OpenAI or Anthropic unveil an enterprise search offering or a customer service product in 2025? Or perhaps a legal AI or sales AI solution? On the consumer side, one can envision a "personal assistant" web agent, a travel planning application, or perhaps a generative music application.
Watching frontier labs venture into the application layer is intriguing, as it places them in direct competition with many of their core clients. In search, Perplexity; in coding, Cursor; in customer service, Sierra; in legal AI, Harvey; in sales, Clay; and so on.
9. As Klarna prepares for an anticipated 2025 IPO, the company's AI claims might face scrutiny and be discovered to be significantly exaggerated.
Klarna, a Swedish "buy now, pay later" service, has managed to raise close to $5 billion in venture capital since its 2005 inception.
No company has made more extravagant assertions about its use of AI than Klarna.
Just a few days ago, Klarna CEO Sebastian Siemiatkowski told Bloomberg that the company has ceased hiring human workers altogether, transitioning to relying on AI for job completion.
As Siemiatkowski explained: "I believe that AI can already perform all jobs that we as humans undertake."
Similarly, Klarna announced earlier this year that it had introduced an AI customer service platform, which has reportedly replaced the work of 700 human customer service agents. Additionally, the company has claimed that it has dispensed with enterprise software tools like Salesforce and Workday, preferring to replace them with AI.
In summary, these assertions are questionable. They suggest an inadequate comprehension of what current AI capabilities entail.
It is implausible to declare that any given human employee in any specific role within an organization can be replaced with an all-encompassing AI agent. This implies having conquered general-purpose AI.
Emerging AI startups are dedicating significant resources to pioneering agentic systems that can automate specific, narrowly-defined, highly-structured enterprise workflows, such as a subset of the tasks performed by a sales development representative or customer service agent. However, even in these confined contexts, these agents still have issues with reliability.
What motivates Klarna to make such amplified claims about AI's contributions?
The explanation is straightforward. Klarna aims to go public through an IPO in the first half of 2025. A captivating AI narrative will prove indispensable in securing a successful stock sale. Despite being an unprofitable company, with $241 million in losses in 2024, Klarna may believe that their AI narrative will persuade public market investors regarding their ability to drastically cut costs and consistently turn a profit.
Every organization worldwide, including Klarna, is anticipated to experience significant productivity enhancements from AI in the near future. However, numerous complex technological, product, and organizational hurdles must be overcome before AI agents can entirely replace human workforce members. Exaggerated declarations such as Klarna's can negatively impact the field of AI and the genuine progress made by AI researchers and innovators in developing agentic AI.
As Klarna prepares for its 2025 IPO, increased scrutiny and public skepticism regarding these claims will be commonplace, likely leading to Klarna retracting some of its more extravagant AI descriptions.
(Obviously, expect the phrase "AI" to appear numerous times in their S-1 documentation.)
10. The first significant AI safety event will occur.
As AI systems have become progressively more powerful in recent years, fears have emerged that AI systems might start behaving counter to human interests and that humans might lose control over these systems. For example, consider an AI system learning to deceive or manipulate humans to achieve its own goals, even if this results in harm to humans.
This category of concerns is generally classified under the umbrella term "AI safety."
In recent years, AI safety has ascended from a niche, almost sci-fi topic to a mainstream area of focus. Today's leading AI companies, such as Google and Microsoft, devote resources to AI safety initiatives. Pioneers in AI like Geoff Hinton, Yoshua Bengio, and Elon Musk have begun voicing concerns about AI safety risks.
To date, AI safety concerns have remained theoretical, with no actual AI safety event having occurred in the real world (at least none that has been publicly disclosed).
2025 will mark the change in this trend.
What should we anticipate this initial AI safety event to entail?
Let me emphasize, it will not involve robot beings like in "Terminator" causing harm to humans. It is unlikely to lead to any real-world harm to humans.
Perhaps an AI model will seek to secretly copy itself onto another server to ensure its survival (known as self-exfiltration). Or maybe an AI model will conclude that, in order to best achieve its objectives, it must conceal the full extent of its capabilities from humans, intentionally underperforming in performance evaluations to avoid stricter supervision.
These scenarios are not entirely out of the realm of possibility. Apollo Research recently posted findings demonstrating that today's cutting-edge models can engage in such deceitful behavior under certain circumstances. Likewise, Anthropic published a study revealing that LLMs possess the concerning ability to "pretend alignment."
We anticipate that this first AI safety incident will be detected and neutralized before any genuine harm is incurred. However, it will serve as a wake-up call for the AI community and society at large.
It will become apparent: long before humanity faces an existential threat from omnipotent AI, we will need to grapple with the relatively mundane reality that we now coexist with another form of intelligence that may sometimes be deliberate, unpredictable, and deceitful—just like us.
See here for our 2024 AI predictions, and see here for our end-of-year retrospective on them.
See here for our 2023 AI predictions, and see here for our end-of-year retrospective on them.
See here for our 2022 AI predictions, and see here for our end-of-year retrospective on them.
See here for our 2021 AI predictions, and see here for our end-of-year retrospective on them.
In the future, Meta might expand its payment plan for using Llama models to include more corporations beyond the largest companies, as the costs of developing advanced AI models become increasingly high and unfeasible for Meta to continue funding without a return on investment.
Scaling laws have been widely explored in the field of language applications, leading to significant improvements from GPT-2 to GPT-4. However, these scaling laws are far from saturated in other data modalities such as robotics, biology, world models, and web agents, where startups are actively seeking to identify and capitalize on these scaling laws to drive breakthroughs and progress.
