Top 10 Predictions for Artificial Intelligence in 2025: AI Agent Direction Will Become Mainstream

As 2024 is coming to an end, Rob Toews, a venture capitalist from Radical Ventures, shared his 10 predictions for artificial intelligence in 2025.

01 Meta will start charging for the Llama model

Meta is a benchmark for open artificial intelligence in the world. In a striking case study of business strategy, while competitors such as OpenAI and Google closed their cutting-edge models to source code and charged usage fees, Meta chose to provide its advanced Llama model for free.

Therefore, it will come as a surprise to many that starting next year, Meta will begin charging companies that use Llama.

It needs to be made clear that: we did not predict that Meta would completely close the source of Llama, nor does it mean that any user using the Llama model must pay for it.

On the contrary, we predict that Meta will impose more restrictions on the open source license terms of Llama, so that companies using Llama in a certain scale of commercial environment will need to start paying to use the model.

Technically, Meta has now achieved this within a limited scope. The company does not allow the largest companies - cloud supercomputers and other companies with over 700 million monthly active users - to freely use its Llama model.

As early as 2023, Meta CEO Mark Zuckerberg said, “If you are a company like Microsoft, Amazon, or Google, and you basically resell Llama, then we should get a share of the revenue from it. I don’t think this will be a big revenue in the short term, but in the long run, I hope this will be some revenue.”

Next year, Meta will significantly expand the scope of enterprises that must pay to use Llama, bringing in more large and medium-sized enterprises.

Keeping up with the cutting-edge large language models (LLMs) is very expensive. If Meta wants Llama to be consistent or close to the latest cutting-edge models of companies such as OpenAI and Anthropic, it needs to invest billions of dollars each year.

Meta is one of the largest and most well-funded companies in the world. But it is also a public company and ultimately accountable to its shareholders.

As the cost of producing cutting-edge models continues to soar, Meta’s practice of investing such a huge amount of money to train the next generation of Llama models without expected revenue is becoming increasingly unsustainable.

Enthusiasts, scholars, individual developers, and startups will continue to use the Llama model for free next year. However, 2025 will be the year when Meta starts to seriously realize the profitability of Llama.

02. Problems related to ‘Scaling Law’

In recent weeks, one of the most discussed topics in the field of artificial intelligence is the scaling laws and whether they are about to end.

The scale law was first proposed in a 2020 OpenAI paper, and its basic concept is simple and clear: as the number of model parameters, training data, and computational power increases during the training of artificial intelligence models, the model’s performance will reliably and predictably improve (technically, its test loss will decrease).

From GPT-2 to GPT-3 to GPT-4, the remarkable performance improvements are all thanks to the scaling law.

Just like Moore’s Law, the scaling law is not actually a true law, but simply an empirical observation.

Over the past month, a series of reports have indicated that major artificial intelligence labs are seeing diminishing returns as they continue to scale up large language models. This helps to explain why the release of OpenAI’s GPT-5 has been repeatedly delayed.

The most common refutation of the scale law tends to be steady is that the calculation during the test opens up a whole new dimension, which allows for scale expansion in this dimension.

In other words, instead of scaling up computation during training, new inference models like OpenAI’s o3 make it possible to scale up computation during inference, unlocking new AI capabilities by allowing the model to “think for a longer time”.

This is an important point. The calculation during testing does represent a new and exciting avenue for AI performance improvement.

But another view of the scale law is more important and has been seriously underestimated in today’s discussion. Almost all discussions of the scale law, from the initial paper in 2020 to the attention to the calculation of tests today, have focused on language. But language is not the only important data pattern.

Think about robot technology, biology, world models, or network proxies. For these data patterns, the scale law has not yet saturated; on the contrary, they have just begun.

In fact, there is still no published evidence of the strict laws of scale in these fields.

Startups that build foundational models for these new data patterns - such as Evolutionary Scale in the field of biology, Physical Intelligence in the field of robotics, and WorldLabs in the field of world modeling - are trying to identify and leverage the scaling laws of these fields, just as OpenAI successfully leveraged the scaling law of large language models (LLMs) in the first half of the 2020s.

Next year, significant progress is expected to be made here.

The laws of scale are not going away, and they will be as relevant in 2025 as ever. However, the center of activity of the law of scale will shift from LLM pre-training to other modes.

**03.**Trump and Musk may have differences in the direction of AI

The new US government will bring a series of policy and strategic changes regarding artificial intelligence.

To predict the direction of artificial intelligence under President Trump and taking into account Musk’s current central position in the field of artificial intelligence, people may tend to focus on the close relationship between the elected president and Musk.

It can be imagined that Musk may influence the development of AI-related Trump administration in various ways.

Given the deep-seated hostility between Musk and OpenAI, the new government may take a less friendly stance towards OpenAI in terms of engaging with the industry, formulating AI regulations, and awarding government contracts, which is a real concern for OpenAI today.

On the other hand, the Trump administration may be more inclined to support Musk’s own company: for example, cutting red tape to enable xAI to establish data centers and take the lead in cutting-edge model competitions; providing fast regulatory approval for deploying a robot taxi fleet for Tesla, etc.

More fundamentally, unlike many other tech leaders favored by Trump, Musk places great emphasis on the security risks of artificial intelligence and therefore advocates significant regulation of AI.

He supports the controversial SB1047 bill in California, which seeks to impose meaningful restrictions on artificial intelligence developers. Therefore, Musk’s influence could lead to a more stringent regulatory environment for artificial intelligence in the United States.

However, all of these speculations have a problem. The close relationship between Trump and Musk will inevitably break down.

As we have seen time and time again during the first Trump administration, the average tenure of Trump’s allies, even the seemingly most steadfast, is very short-lived.

Among the deputies of Trump’s first government, there are few who remain loyal to him today.

Both Trump and Musk have complex, volatile, and unpredictable personalities. They are not easy to cooperate with and exhaust people. Their newly discovered friendship has been mutually beneficial so far, but still in the “honeymoon period”.

We predict that this relationship will deteriorate by the end of 2025.

What does this mean for the world of artificial intelligence?

This is good news for OpenAI. This will be unfortunate news for Tesla’s shareholders. And for those who are concerned about the safety of artificial intelligence, this will be a disappointing news, as it almost ensures that the US government will take a hands-off approach to AI regulation during the Trump administration.

04 AI Agent will become mainstream

Imagine a world where you no longer need to interact directly with the internet. Whenever you need to manage a subscription, pay a bill, make a doctor’s appointment, order something on Amazon, make a restaurant reservation, or complete any other tedious online task, you can simply instruct the AI assistant to do it for you.

The concept of this ‘network proxy’ has been around for years. If there is such a product and it can function properly, there is no doubt that it will be a highly successful product.

However, there is currently no functional general-purpose network proxy available on the market.

Startups like Adept, even with a team of purebred founders and hundreds of millions of dollars raised, have failed to realize their vision.

Next year will be the year when web proxies finally start to work well and become mainstream. Continued advancements in linguistic and visual foundation models, coupled with recent breakthroughs in “second system thinking” capabilities due to new inference models and inference time calculations, will mean that web agents are poised for a golden age.

In other words, Adept was right, just too early. In a start-up, as with many things in life, timing is everything.

The Web3 proxy will find a variety of valuable enterprise use cases, but we believe the Web3 proxy’s biggest market opportunity in the near term will be consumers.

Despite the recent popularity of artificial intelligence, there are relatively few native AI applications that can become mainstream consumer applications other than ChatGPT.

Network agents will change this situation and become the next real ‘killer application’ in the consumer artificial intelligence field.

The idea of ​​placing an artificial intelligence data center in space will be realized.

In 2023, the key physical resource constraining the development of artificial intelligence is GPU chips. In 2024, it becomes power and data centers.

In 2024, few stories can compare to the attention on the huge and rapidly growing demand for energy as artificial intelligence is eager to build more AI data centers.

Due to the rapid development of artificial intelligence, the power demand of global data centers, which has remained stable for decades, is expected to double between 2023 and 2026. In the United States, the power consumption of data centers is projected to reach nearly 10% of the total power consumption by 2030, compared to only 3% in 2022.

Today’s energy systems simply can’t handle the huge surge in demand from AI workloads. A historic collision is on the horizon between our energy grid and computing infrastructure, two trillion-dollar systems.

As a possible solution to this dilemma, nuclear energy has gained momentum this year. Nuclear power is in many ways an ideal source of energy for artificial intelligence: it is a zero-carbon energy source, available around the clock, and, in fact, inexhaustible.

However, from a realistic perspective, due to the long research, project development, and regulatory time, new energy will not be able to solve this problem before the 2030s. This is true for traditional nuclear fission power plants, next-generation Small Modular Reactors (SMRs), and nuclear fusion power plants.

Next year, an unconventional new idea to address this challenge will emerge and attract real resources: placing artificial intelligence data centers in space.

Artificial intelligence data centers in space, at first glance, may sound like a bad joke, an venture capitalist trying to combine too many startup buzzwords.

But in fact, this may make sense.

The biggest bottleneck to building more data centers on the planet quickly is getting the power you need. Computing clusters in orbit can enjoy free, unlimited, zero-carbon electricity around the clock: the sun in space always shines.

Another important advantage of placing computing in space is that it solves the cooling problem.

One of the biggest engineering obstacles to building a more powerful artificial intelligence data center is the heat generated by running many GPUs in a confined space, which can cause damage or destruction to the computing devices.

Data center developers are resorting to expensive and unproven methods such as liquid immersion cooling to try to solve this problem. But space is extremely cold, and any heat generated by computing activity dissipates immediately and harmlessly.

Of course, there are many practical challenges that need to be addressed. An obvious question is whether and how to transmit large amounts of data between orbit and Earth at low cost and high efficiency.

This is a pending issue, but it may prove to be resolvable: promising work can be carried out using lasers and other high-bandwidth optical communication technologies.

A startup named Lumen Orbit, backed by YCombinator, recently raised $11 million to realize the vision of building a multi-megawatt data center network in space for training artificial intelligence models.

As the CEO of the company said, ‘Instead of paying $140 million in electricity bills, it’s better to pay $10 million for launch and solar costs.’

!

In 2025, Lumen will not be the only organization taking this concept seriously.

Competitors from other start-ups will also emerge. Don’t be surprised if one or more cloud computing hyperscalers are exploring along these lines.

Amazon already has a wealth of experience putting assets into orbit through Project Kuiper; Google has long been funding similar “moonshots”; Even Microsoft is no stranger to the space economy.

It can be imagined that SpaceX, Musk’s company, will also make achievements in this regard.

06 The AI system will pass the ‘Turing Voice Test’.

The Turing test is one of the oldest and most well-known benchmarks for artificial intelligence performance.

In order to ‘pass’ the Turing test, an artificial intelligence system must be able to communicate through written text, making it impossible for ordinary people to distinguish whether they are interacting with artificial intelligence or with other people.

Thanks to the significant advances in large-scale language models, the Turing test has become a solved problem in the 2020s.

But written texts are not the only way humans communicate.

As AI becomes more and more multimodal, one can imagine a new, more challenging version of the Turing Test – the “Speech Turing Test.” In this test, the AI system must be able to interact with humans through speech, with skills and fluency that are indistinguishable from human speakers.

Today’s artificial intelligence systems cannot yet pass the Turing test for speech. Solving this problem will require further technological advances. The delay (the lag between human speech and AI response) must be reduced to near zero to match the experience of conversing with another human.

Voice AI systems must be better at gracefully handling real-time processing of fuzzy input or misinterpretation, such as cases where speech is interrupted. They must be able to engage in long conversations, multi-turn, open-ended dialogues while remembering earlier parts of the discussion.

And it is crucial that the voice AI agent learns to better understand the non-verbal cues in speech. For example, what it means when a human speaker sounds angry, excited, or sarcastic, and generates these non-verbal cues in its own speech.

As we approach the end of 2024, speech AI is at an exciting turning point, driven by fundamental breakthroughs such as the emergence of speech-to-speech models.

Today, few fields in artificial intelligence are progressing as rapidly in both technology and business as voice artificial intelligence. It is expected that the latest technology in voice artificial intelligence will make a leap forward by 2025.

07 The autonomous AI system will make significant progress

The concept of recursive self-improving AI has been a frequent topic in the AI community for decades.

For example, as early as 1965, I.J.Good, a close collaborator of Alan Turing, wrote:“Let us define a super-intelligent machine as a machine that can far exceed all human intellectual activities, no matter how clever it is.”

"Since designing machines is one of these intellectual activities, superintelligent machines can design better machines; At that point, there will undoubtedly be an ‘intelligence explosion’ where human intelligence will be left far behind. ”

Artificial intelligence can invent better artificial intelligence, which is a concept full of wisdom. However, even today, it still retains the shadow of science fiction.

However, although this concept is not yet widely recognized, it has actually begun to become more real. Researchers at the frontiers of AI science have begun to make tangible progress in building AI systems that themselves can build better AI systems.

We predict that this research direction will become mainstream next year.

So far, the most notable publicly available example of research in this direction is Sakana’s ‘Artificial Intelligence Scientist’.

“Artificial Intelligence Scientist” was released in August this year, convincingly proving that artificial intelligence systems can indeed conduct AI research completely autonomously.

Sakana’s ‘AI Scientist’ itself performs the entire lifecycle of AI research: reading existing literature, generating new research ideas, designing experiments to test these ideas, conducting these experiments, writing research papers to report the research results, and then peer reviewing their work.

These tasks are done entirely autonomously by artificial intelligence and do not require human intervention. You can read some of the research papers written by AI scientists online.

OpenAI, Anthropic, and other research labs are investing resources in the idea of ‘automated AI researchers’, but there is currently no public acknowledgement of any news.

As more and more people realize that artificial intelligence research automation is actually becoming a real possibility, it is expected that there will be more discussions, advancements, and entrepreneurial activities in this field by 2025.

The most significant milestone, though, will be the first time that a research paper written entirely by an AI agent has been accepted by a top AI conference. If the paper is blindly reviewed, the conference reviewer will not know that the paper was written by AI until the paper is accepted.

Don’t be surprised if the results of AI research are accepted by NeurIPS, CVPR, or ICML next year. This will be a fascinating and controversial historic moment for the field of artificial intelligence.

08 Industry giants like OpenAI are strategically shifting their focus to building applications

Building cutting-edge models is a difficult task.

Its capital intensity is staggering. The cutting-edge model lab requires a large amount of cash. Just a few months ago, OpenAI raised a record $6.5 billion, and in the near future, it may need to raise even more funds. Anthropic, xAI, and other companies are also in a similar situation.

Conversion costs and customer loyalty are low. AI applications are often built with model agnosticism in mind, and models from different vendors can be seamlessly switched based on changing cost and performance comparisons.

With the advent of state-of-the-art open models, such as Meta’s Llama and Alibaba’s Qwen, the threat of technology commoditization looms. AI leaders like OpenAI and Anthropic can’t and won’t stop investing in building cutting-edge models.

However, next year, in order to develop business lines with higher profits, greater differentiation, and stronger stickiness, the Frontier Lab is expected to vigorously launch more of its own applications and products.

Of course, Frontier Labs already has a very successful use case: ChatGPT.

What other types of first-party applications can we see from AI Labs in the new year? One obvious answer is a more sophisticated, feature-rich search app. OpenAI’s SearchGPT heralds this.

Coding is another obvious category. Similarly, with OpenAI’s Canvas product making its debut in October, preliminary productization work has begun.

Will OpenAI or Anthropic launch an enterprise search product in 2025? Or is it a customer service product, a legal AI, or a sales AI product?

On the consumer side, we can imagine a “personal assistant” web proxy product, or a trip planning app, or an app that generates music.

One of the most fascinating things about watching cutting-edge labs move to the application layer is that this move will put them in direct competition with many of their most important customers.

Perplexity in the search field, Cursor in the encoding field, Sierra in the customer service field, Harvey in the legal artificial intelligence field, Clay in the sales field, and so on.

09 Klarna will go public in 2025, but there are signs of overestimating the value of AI

Klarna is a Sweden-based “buy now, pay, pay” service provider that has raised nearly $5 billion in venture capital since its inception in 2005.

Perhaps no company can boast as much about its application of artificial intelligence as Klarna.

Just a few days ago, Klarna CEO Sebastian Siemiatkowski told Bloomberg that the company had stopped hiring human workers altogether and instead relied on generative AI to get the job done.

As Siemiatkowski puts it: “I think that AI can already do all the work that we humans do.” ”

In a similar vein, Klarna announced earlier this year that it had launched an AI-powered customer service platform that has fully automated the work of 700 human agents.

The company also claims that it has stopped using enterprise software products such as Salesforce and Workday because it can simply replace them with artificial intelligence.

To put it bluntly, these claims are not credible. They reflect a lack of understanding of the capabilities and inadequacies of today’s AI systems.

Claims that an end-to-end AI agent can replace any specific human employee in any functional department of an organization are not reliable. This is equivalent to solving the problem of general human-level artificial intelligence.

Today, leading AI start-ups are at the forefront of the field working to build agency systems that automate specific, narrow, and highly structured enterprise workflows, for example, sales development reps or a subset of customer service agent activities.

Even in these narrow scopes, these proxy systems don’t work completely reliably, although in some cases they’ve started to work well enough to get commercial applications in the early days.

Why does Klarna exaggerate the value of artificial intelligence?

The answer is simple. The company plans to go public in the first half of 2025. The key to a successful listing is to have a compelling artificial intelligence story.

Klarna remains a non-profit enterprise, losing $241 million last year. It may hope that its artificial intelligence story can convince public market investors that it has the ability to significantly reduce costs and achieve sustainable profitability.

There is no doubt that every business in the world, including Klarna, will enjoy the huge productivity gains that AI will bring about in the coming years. However, there are still many thorny technical, product, and organizational challenges to be solved before AI agents can completely replace humans in the workforce.

Describing Klarna in such exaggerated terms is a blasphemy to the field of artificial intelligence and a blasphemy to the hard progress made by artificial intelligence technology experts and entrepreneurs in developing AI agents.

As Klarna prepares to go public in 2025, it is expected that these claims will face stricter scrutiny and public skepticism, while so far, most of these claims have not been questioned. Don’t be surprised if the company’s descriptions of its AI applications are somewhat exaggerated.

10 The first real AI security incident will occur

In recent years, as AI has become more powerful, there have been growing concerns that AI systems may begin to behave in ways that are inconsistent with human interests, and that humans may lose control of these systems.

Imagine, for example, that an AI system learns to deceive or manipulate humans in order to achieve its goals, even if those goals cause harm to humans. These concerns are often categorized as “AI security” concerns.

In recent years, AI security has shifted from a fringe quasi-sci-fi topic to a mainstream field of activity.

Today, every major player in artificial intelligence, from Google and Microsoft to OpenAI, has invested heavily in AI security work. AI idols like Geoff Hinton, Yoshua Bengio, and Elon Musk have also begun to express their views on AI security risks.

However, so far, the issue of artificial intelligence security remains purely theoretical. There has never been a real artificial intelligence security incident in the real world (at least not publicly reported).

2025 will be the year to change that, what will the first AI security incident look like?

To be clear, it will not involve Terminator-style killing robots, and it is unlikely to cause any harm to humans.

Perhaps the artificial intelligence model will try to secretly create its own copy on another server to preserve itself (known as self-filtering).

Perhaps the AI model may come to this conclusion: In order to best advance its assigned goals, it needs to hide its true capabilities from humans, deliberately perform low-key in performance evaluation, and avoid more rigorous scrutiny.

These examples are not far-fetched. Important experiments published earlier this month by Apollo Research have shown that today’s cutting-edge models are capable of this deception under specific prompts.

Similarly, recent studies in anthropology also suggest that LLMs have unsettling ‘pseudo-alignment’ capabilities.

!

We expect that this first artificial intelligence security incident will be discovered and eliminated before causing any actual harm. But for the AI industry and the entire society, this will be a eye-opening moment.

It will make one thing clear: before humanity faces the existential threat of all-powerful artificial intelligence, we need to accept a more mundane reality: we are now sharing our world with another form of intelligence, which can sometimes be capricious, unpredictable, and deceptive.