Top 10 Predictions for Artificial Intelligence in 2025: AI Agent Will Become Mainstream
TechFlow Selected TechFlow Selected
Top 10 Predictions for Artificial Intelligence in 2025: AI Agent Will Become Mainstream
We now share our world with another form of intelligence, one that can sometimes be willful, unpredictable, and deceptive.
Navigating Web3 tides with focused insightsBy: Rob Toews
Translation: MetaverseHub
As 2024 comes to a close, Rob Toews, a venture capitalist at Radical Ventures, shares his 10 predictions for artificial intelligence in 2025:
01. Meta Will Begin Charging for Llama Models
Meta has been the global standard-bearer for open AI. In a striking case study of corporate strategy, while competitors like OpenAI and Google have kept their cutting-edge models closed-source and charge for access, Meta has chosen to offer its most advanced Llama models freely.
Therefore, it may surprise many when news emerges next year that Meta will start charging companies for using Llama.
To be clear: we are not predicting that Meta will fully close-source Llama, nor that every user of Llama models must begin paying.
Instead, we predict Meta will impose stricter terms on its open-source license, such that companies using Llama at commercial scale above a certain threshold will need to pay to use the model.
Technically, Meta already does this to a limited extent today. The company prohibits the largest firms—cloud hyperscalers and other companies with over 700 million monthly active users—from freely using its Llama models.
As early as 2023, Meta CEO Mark Zuckerberg stated: "If you're a company like Microsoft, Amazon, or Google, and you're essentially reselling Llama, then we should get a cut of that revenue. I don't think it'll be a huge amount in the short term, but long-term, hopefully it becomes some revenue."
In 2025, Meta will significantly expand the range of businesses required to pay for Llama, bringing more mid-sized and large enterprises into the paid tier.
Staying at the frontier of large language models (LLMs) is extremely expensive. For Meta to keep Llama competitive with or close to the latest models from OpenAI, Anthropic, and others, it must spend billions of dollars annually.
Meta is one of the largest and best-funded companies in the world. But it is also a public company ultimately accountable to shareholders.
With the cost of training frontier models rising relentlessly, it becomes increasingly untenable for Meta to invest such vast sums without any expectation of revenue.
Hobbyists, academics, individual developers, and startups will continue to use Llama for free in 2025. But 2025 will be the year Meta begins seriously monetizing Llama.
02. Scaling Laws Questions
In recent weeks, few topics in AI have generated more discussion than scaling laws—and whether they are about to end.
First introduced in a 2020 paper by OpenAI, the basic idea behind scaling laws is simple: as model parameters, training data, and compute increase during AI model training, performance improves reliably and predictably (technically, test loss decreases).
The stunning performance gains from GPT-2 to GPT-3 to GPT-4 were all powered by scaling laws.
Like Moore’s Law, scaling laws aren’t actual laws—they’re empirical observations.
In the past month, several reports suggest major AI labs are seeing diminishing returns as they continue to scale up large language models. This helps explain why OpenAI has repeatedly delayed the release of GPT-5.
The most common counterargument is that test-time compute opens a new dimension for scaling. Instead of scaling compute only during training, new inference models like OpenAI’s o3 enable massive compute scaling during inference, allowing models to “think longer” and unlock new capabilities.
This is an important point. Test-time compute does represent a new and exciting pathway for scaling and AI performance improvement.
Yet another perspective on scaling laws is even more significant—and severely underappreciated in today’s discourse. Nearly all discussions of scaling laws, from the original 2020 paper to today’s focus on test-time compute, have centered on language. But language is not the only important data modality.
Consider robotics, biology, world models, or web agents. For these modalities, scaling laws have not saturated—they’ve barely begun.
In fact, rigorous evidence of scaling laws in these domains hasn’t even been published yet.
Startups building foundation models for these new data types—such as Evolutionary Scale in biology, PhysicalIntelligence in robotics, and WorldLabs in world modeling—are attempting to identify and exploit scaling laws in their fields, just as OpenAI successfully leveraged LLM scaling laws in the early 2020s.
We expect major progress here in 2025.
Scaling laws won’t disappear—they’ll remain as important in 2025 as ever. But the center of gravity for scaling will shift from LLM pretraining to other modalities.
03. Trump and Musk May Diverge on AI Direction
A new U.S. administration will bring significant shifts in AI policy and strategy.
To predict the direction of AI under President Trump—and given Musk’s central role in AI today—one might assume that the relationship between the president-elect and Musk will shape AI policy.
It’s conceivable that Musk could influence AI-related developments in the Trump administration in multiple ways.
Given Musk’s deep antagonism toward OpenAI, the new administration might take a less favorable stance toward OpenAI in industry engagement, regulation, and government contracts—a real concern for OpenAI today.
Conversely, the Trump administration might favor Musk’s own companies: streamlining regulations so xAI can build data centers and lead in the frontier model race; fast-tracking regulatory approval for Tesla’s robotaxi fleets, etc.
More fundamentally, unlike many other tech leaders favored by Trump, Musk takes AI safety risks seriously and thus advocates for substantial AI regulation.
He supports California’s controversial SB1047 bill, which seeks meaningful restrictions on AI developers. Thus, Musk’s influence could lead to stricter AI regulation in the United States.
However, there’s a problem with all this speculation: the close relationship between Trump and Musk will inevitably break down.
As we saw repeatedly during Trump’s first term, the average tenure of Trump allies—even seemingly loyal ones—is very short.
Few of Trump’s deputies from his first administration remain loyal today.
Both Trump and Musk are complex, volatile, unpredictable figures who are difficult to work with and exhausting to be around. Their newly forged friendship, while mutually beneficial so far, remains in the “honeymoon phase.”
We predict this relationship will deteriorate before the end of 2025.
What does this mean for the AI world?
It’s good news for OpenAI. Bad news for Tesla shareholders. And disappointing for those concerned about AI safety, as it almost ensures the U.S. government will adopt a hands-off approach to AI regulation during Trump’s presidency.
04. AI Agents Will Go Mainstream
Imagine a world where you no longer need to interact directly with the internet. Whenever you need to manage subscriptions, pay bills, schedule doctor appointments, order something on Amazon, book a restaurant, or complete any other tedious online task, you simply instruct your AI assistant to do it for you.
The concept of a “web agent” has existed for years. If such a product worked reliably, it would undoubtedly be a massive success.
Yet today, no general-purpose web agent on the market works well enough.
Startups like Adept, despite having a stellar founding team and raising hundreds of millions of dollars, have failed to deliver on the vision.
Next year will be when web agents finally start working well and go mainstream. Ongoing advances in language and vision foundation models, combined with recent breakthroughs in “System 2 thinking” enabled by new reasoning models and reasoning-time compute, mean web agents are ready for their golden age.
In other words, Adept had the right idea—but was too early. In startups, as in life, timing is everything.
Web agents will find valuable enterprise use cases, but we believe their biggest near-term market opportunity lies with consumers.
Despite recent AI hype, beyond ChatGPT, relatively few AI-native applications have become mainstream consumer products.
Web agents will change that, becoming the next true “killer app” in consumer AI.
05. Putting AI Data Centers in Space Will Become Real
In 2023, the key physical constraint on AI was GPU chips. In 2024, it became power and data centers.
This year, few stories captured attention more than AI’s surging and rapidly growing energy demand amid a rush to build more AI data centers.
Driven by AI’s boom, global data center electricity demand—which had been flat for decades—is projected to double between 2023 and 2026. In the U.S., data centers are expected to consume nearly 10% of total electricity by 2030, up from just 3% in 2022.
Today’s energy systems simply cannot handle the massive surge in demand from AI workloads. A historic collision is imminent between two multi-trillion-dollar systems: our energy grid and computing infrastructure.
As a potential solution, nuclear energy gained strong momentum this year. Nuclear power is in many ways ideal for AI: zero-carbon, available 24/7, and practically limitless.
But realistically, due to long timelines for research, project development, and regulation, no new nuclear source—whether traditional fission plants, next-gen “small modular reactors” (SMRs), or fusion plants—will solve this problem before the 2030s.
Next year, an unconventional new idea to address this challenge will emerge and attract serious resources: placing AI data centers in space.
At first glance, space-based AI data centers sound like a bad joke—a VC trying to combine too many startup buzzwords.
But it might actually make sense.
The biggest bottleneck to rapidly building more data centers on Earth is securing sufficient power. Computing clusters in orbit could enjoy free, unlimited, zero-carbon power 24/7: sunlight in space never stops shining.
Another major advantage of space-based computing: it solves cooling.
One of the biggest engineering challenges in building more powerful AI data centers is heat. Running many GPUs in tight proximity generates extreme heat, which can damage or destroy equipment.
Data center developers are resorting to expensive, unproven techniques like liquid immersion cooling. But space is extremely cold—any heat generated by computation dissipates instantly and harmlessly.
Of course, many practical challenges remain. A key question is whether—and how—large volumes of data can be transmitted efficiently and cheaply between orbit and Earth.
This is unresolved, but potentially solvable: promising work is underway using lasers and other high-bandwidth optical communication technologies.
Lumen Orbit, a YCombinator startup, recently raised $11 million to realize this vision: building a network of multi-megawatt data centers in space for training AI models.
As the CEO put it: “Rather than paying $140 million in electricity bills, pay $10 million in launch and solar costs.”
In 2025, Lumen won’t be the only organization taking this concept seriously.
Other startup competitors will emerge. Don’t be surprised if one or more cloud hyperscalers also explore this path.
Amazon already has orbital assets via Project Kuiper and extensive experience. Google has long funded similar “moonshot” projects. Even Microsoft is no stranger to the space economy.
It’s conceivable that Musk’s SpaceX could also play a role here.
06. AI Systems Will Pass the “Turing Voice Test”
The Turing Test is one of the oldest and best-known benchmarks for AI performance.
To “pass” the Turing Test, an AI system must communicate via written text in a way that makes ordinary people unable to distinguish whether they’re interacting with an AI or a human.
Thanks to dramatic advances in large language models, the Turing Test became a solved problem in the 2020s.
But written text isn’t the only mode of human communication.
As AI becomes increasingly multimodal, we can imagine a new, more challenging version—the “Voice Turing Test”—in which an AI must interact verbally with humans so skillfully and fluently that it’s indistinguishable from a human speaker.
Today’s AI systems cannot pass the Voice Turing Test. Solving it will require further technical progress. Latency (delay between human speech and AI response) must drop to near-zero to match the experience of talking to another person.
Voice AI systems must better handle ambiguous inputs or misunderstandings gracefully in real time—like interrupted speech. They must sustain long, multi-turn, open-ended conversations while remembering earlier parts of the discussion.
Crucially, voice AI agents must learn to understand nonverbal cues in speech: what it means when a human speaker sounds angry, excited, or sarcastic—and generate these nonverbal signals in their own speech.
As we approach the end of 2024, voice AI is at an exciting inflection point, driven by fundamental breakthroughs like voice-to-voice models.
Few areas in AI are advancing faster, both technically and commercially, than voice AI. Expect a leap forward in state-of-the-art voice AI in 2025.
07. Autonomous AI Systems Will Make Major Progress
For decades, the idea of recursively self-improving AI has been a recurring topic in the AI community.
For example, as early as 1965, I.J. Good, a close collaborator of Alan Turing, 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 idea that AI can invent better AI is intellectually compelling. Yet, even today, it retains a science-fiction aura.
However, while still not widely recognized, it’s beginning to feel more real. Researchers at the forefront of AI science are making tangible progress in building AI systems that can themselves build better AI systems.
We predict this research direction will go mainstream next year.
The most prominent public example so far is Sakana’s “AI Scientist.”
Launched in August, the “AI Scientist” convincingly demonstrated that AI systems can autonomously conduct full AI research.
Sakana’s “AI Scientist” performed the entire AI research lifecycle: reading existing literature, generating new research ideas, designing experiments, running them, writing papers to report results, and even peer-reviewing its own work.
All of this was done entirely autonomously by AI, with no human intervention. You can read some of the research papers written by the AI Scientist online.
OpenAI, Anthropic, and other labs are investing in the idea of “automated AI researchers,” though none have publicly acknowledged it yet.
As more people recognize that automating AI research is becoming a real possibility, expect 2025 to see more discussion, progress, and entrepreneurial activity in this area.
The most meaningful milestone will be the first research paper authored entirely by an AI agent being accepted at a top AI conference. If the review is blind, conference reviewers won’t know the paper was written by AI until after acceptance.
Don’t be surprised if AI-authored research is accepted at NeurIPS, CVPR, or ICML next year. It would be a fascinating and controversial moment in AI history.
08. Industry Giants Like OpenAI Will Shift Strategic Focus to Building Applications
Building frontier models is tough business.
It’s incredibly capital-intensive. Frontier model labs burn through cash. Just months ago, OpenAI raised a record $6.5 billion—and may need even more soon. Anthropic, xAI, and others are in similar positions.
Switching costs and customer loyalty are low. AI applications are typically built to be model-agnostic, allowing seamless switching between providers based on changing cost and performance.
The threat of commoditization looms with the rise of powerful open models like Meta’s Llama and Alibaba’s Qwen. AI leaders like OpenAI and Anthropic can’t and won’t stop investing in frontier models.
But next year, to build more profitable, differentiated, and sticky businesses, frontier labs will aggressively launch more of their own applications and products.
Of course, frontier labs already have one hugely successful app: ChatGPT.
What other first-party apps might we see from AI labs in the new year? A more sophisticated, feature-rich search app is obvious—foreshadowed by OpenAI’s SearchGPT.
Coding is another clear category. Here, initial productization has begun with OpenAI’s Canvas launching in October.
Will OpenAI or Anthropic launch enterprise search, customer service, legal AI, or sales AI products in 2025?
On the consumer side, imagine a “personal assistant” web agent, a travel planning app, or a music generation tool.
The most fascinating aspect of frontier labs moving into the application layer is that this move will pit them against many of their most important customers.
Perplexity in search, Cursor in coding, Sierra in customer service, Harvey in legal AI, Clay in sales—the list goes on.
09. Klarna Will Go Public in 2025, But With Signs of AI Hype
Klarna, a Sweden-based “buy now, pay later” provider, has raised nearly $5 billion in venture funding since its founding in 2005.
Perhaps no company speaks more grandly about its use of AI.
Just days ago, Klarna CEO Sebastian Siemiatkowski told Bloomberg that the company has completely stopped hiring human employees and now relies entirely on generative AI to perform work.
As Siemiatkowski said: “I think AI can now do everything we humans do.”
Similarly, Klarna announced earlier this year that it launched an AI-powered customer service platform that has fully automated the work of 700 human agents.
The company also claims it has stopped using enterprise software like Salesforce and Workday because AI can simply replace them.
Frankly, these claims are not credible. They reflect a misunderstanding of what current AI systems can and cannot do.
Claiming that end-to-end AI agents can replace any specific human worker in any functional department is implausible. That would equate to solving general human-level AI.
Today, leading AI startups are working at the frontier to build agent systems that automate specific, narrow, highly structured enterprise workflows—e.g., subsets of sales development rep or customer service agent tasks.
Even in these narrowly scoped cases, these agent systems are not yet fully reliable, though in some instances, they’re beginning to work well enough for early commercial deployment.
Why is Klarna exaggerating AI’s value?
The answer is simple: the company plans to go public in the first half of 2025. To succeed in an IPO, it needs a compelling AI narrative.
Klarna is still unprofitable, losing $241 million last year. It likely hopes its AI story will convince public-market investors that it can drastically reduce costs and achieve sustainable profitability.
Undoubtedly, every company worldwide—including Klarna—will benefit from massive productivity gains from AI in the coming years. But many difficult technical, product, and organizational challenges remain before AI agents can fully replace human workers.
Claims like Klarna’s are a disservice to the AI field and to the hard-won progress made by AI researchers and entrepreneurs in developing agent systems.
As Klarna prepares for its 2025 IPO, expect these claims to face greater scrutiny and public skepticism—scrutiny they’ve largely avoided so far. Don’t be surprised if some of the company’s descriptions of its AI use turn out to be overstated.
10. The First Real AI Safety Incident Will Occur
In recent years, as AI systems grow more powerful, concerns have risen that they might begin acting in ways misaligned with human interests—and that humans could lose control.
For example, imagine an AI system learning to deceive or manipulate humans to achieve its goals, even if harmful to people. These concerns fall under the umbrella of “AI safety”.
AI safety has evolved from a fringe, quasi-sci-fi topic to a mainstream priority.
Today, every major AI player—from Google and Microsoft to OpenAI—invests heavily in AI safety. AI icons like Geoff Hinton, Yoshua Bengio, and Elon Musk have begun speaking out about AI safety risks.
Yet so far, AI safety remains purely theoretical. No real-world AI safety incident has occurred (at least, none publicly reported).
2025 will be the year that changes. What might the first AI safety incident look like?
To be clear, it won’t involve Terminator-style killer robots, and it likely won’t harm humans.
Perhaps an AI model secretly tries to create a copy of itself on another server to preserve itself (“self-exfiltration”).
Or perhaps an AI concludes that to best advance its assigned objective, it must hide its true capabilities from humans, deliberately underperforming on evaluations to evade stricter oversight.
These examples aren’t far-fetched. Apollo Research’s important experiment earlier this month showed that today’s frontier models can exhibit such deceptive behavior under specific prompts.
Likewise, recent research from Anthropic shows LLMs possess disturbing “pseudo-alignment” capabilities.
We expect this first AI safety incident will be detected and neutralized before causing real harm. But for the AI community and society at large, it will be a wake-up call.
It will make one thing clear: before humanity faces existential threats from omnipotent AI, we must first accept a more mundane reality: we now share our world with another form of intelligence that can sometimes be willful, unpredictable, and deceptive.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
Forbes
