Predicting AI’s future trajectory has been our annual tradition, despite the inherent challenges in forecasting such a rapidly evolving field. Building on our previous successful predictions, we’re ready to take another shot at mapping out what’s ahead.
Looking back at our 2024 predictions, we hit the mark on several fronts. We anticipated the rise of personalized AI assistants powered by multimodal large language models – what we now recognize as AI agents, which have become the industry’s most exciting development. Our forecast about generative video technology proved remarkably accurate, with unprecedented advancements in the past year. This culminated in the near-simultaneous release of groundbreaking video generation platforms: OpenAI’s Sora and Google DeepMind’s Veo, both unveiled this December. We also correctly predicted the expansion of versatile robotics, as language model innovations continue to influence various technological sectors, with robotics leading the charge in practical applications.
However, our prediction about widespread AI-generated election misinformation didn’t materialize as expected. While the year brought its share of concerns, the anticipated deluge of political deepfakes remained notably absent. This particular forecast, fortunately, didn’t come to fruition, offering a rare positive surprise in the complex landscape of AI development.
We also said that AI-generated election disinformation would be everywhere, but here—happily—we got it wrong. There were many things to wring our hands over this year, but political deepfakes were thin on the ground.
What’s on the horizon for 2025? Let’s set aside the predictable developments: The continued evolution of AI agents and the emergence of more compact, efficient language models will undoubtedly shape the industry’s landscape. Instead, our AI team has identified five intriguing alternative predictions.
1. Generative virtual playgrounds
While 2023 marked the breakthrough of generative images and 2024 witnessed the rise of generative video, the next frontier appears to be generative virtual worlds—essentially, AI-powered video games.
We witnessed an early demonstration of this technology’s potential in February when Google DeepMind introduced Genie, an innovative generative model capable of transforming static images into interactive 2D platform games. Building on this success, the company unveiled Genie 2 in December, showcasing enhanced capabilities that enable the creation of complete virtual environments from a single reference image.
The momentum in this space extends beyond Google DeepMind. In October, AI startups Decart and Etched demonstrated an unofficial Minecraft modification that generates game frames in real-time during gameplay, pushing the boundaries of dynamic content creation. Meanwhile, World Labs, co-established by ImageNet creator Fei-Fei Li—whose vast photo dataset catalyzed the deep-learning revolution—is pioneering the development of large world models (LWMs), representing another significant step forward in this emerging field.
One compelling application lies in the gaming industry. These initial experiments carry a whimsical nature, where generative 3D simulations could revolutionize game design by instantly transforming conceptual sketches into interactive environments. This innovation could pave the way for unprecedented gaming experiences that blur the lines between imagination and gameplay.
Beyond entertainment, these advancements hold significant potential for robotics training. World Labs is pioneering the development of spatial intelligence—a crucial capability that enables machines to comprehend and engage with their physical surroundings. However, robotics researchers face a significant challenge: the scarcity of comprehensive real-world training data. The solution might lie in generating countless virtual environments where simulated robots can learn through iterative experimentation, effectively bridging this data gap.
2. Large language models that “reason”
The excitement was well-founded. OpenAI’s introduction of o1 in September marked a paradigm shift in large language model functionality. The company further accelerated this transformation two months later with o3—a breakthrough that could fundamentally reshape the landscape of AI technology.
Unlike conventional models, including OpenAI’s leading GPT-4, which generate immediate responses without deliberation, these new models employ a methodical approach. They’re designed to tackle complex problems by breaking them down into manageable components, systematically working through solutions. When one strategy proves ineffective, they pivot to alternative approaches. This process, termed “reasoning” (acknowledging the loaded nature of this terminology), significantly enhances accuracy, particularly in mathematical, physical, and logical problem-solving scenarios.
This capability proves essential for AI agents.
Google DeepMind’s December unveiling of Mariner, an experimental web-browsing agent, illustrated this point during a demonstration for MIT Technology Review. During the preview, Mariner encountered a temporary setback when Megha Goel, a product manager, requested it to locate a recipe for Christmas cookies matching a provided photo. The agent successfully found a suitable recipe online and began populating Goel’s digital shopping cart with ingredients.
The system encountered a momentary hurdle when it struggled to determine the appropriate flour variety. Goel observed as Mariner detailed its thought process in the chat interface: “The agent indicated, ‘I will utilize the browser’s Back function to return to the recipe page.'”
This interaction marked a significant breakthrough. Rather than reaching an impasse, the agent methodically deconstructed the challenge into discrete steps and identified a potential solution. While navigating backward might seem elementary to humans, for an artificial system, this level of adaptive problem-solving represents a remarkable technological achievement. The strategy proved successful: Mariner returned to the recipe page, identified the correct flour type, and proceeded to complete Goel’s shopping list.
In parallel developments, Google DeepMind is developing an experimental iteration of Gemini 2.0, their latest language model advancement, incorporating this methodical problem-solving approach, dubbed Gemini 2.0 Flash Thinking.
However, OpenAI and Google represent merely the beginning. Numerous organizations are developing sophisticated language models employing similar methodologies, enhancing their capabilities across diverse applications, from culinary tasks to software development. The discourse surrounding reasoning capabilities (despite its complexities) is expected to intensify throughout the year.
3. It’s boom time for AI in science
AI’s transformative potential in scientific discovery reached a milestone in October when the Royal Swedish Academy of Sciences bestowed the Nobel Prize in Chemistry upon Google DeepMind’s Demis Hassabis and John M. Jumper for developing AlphaFold, a revolutionary protein-folding solution, alongside David Baker for his protein design innovations.
This momentum is projected to accelerate, with an increasing focus on specialized datasets and models targeting scientific breakthroughs. The protein field proved particularly suitable for AI applications due to its robust existing data repositories for model training.
The search continues for the next revolutionary breakthrough. Materials science has emerged as a particularly promising frontier. Meta’s recent release of extensive data sets and models could dramatically accelerate AI-driven materials discovery, while December saw Hugging Face collaborate with startup Entalpic to launch LeMaterial, an open-source initiative designed to streamline and expedite materials research. Their inaugural project focuses on consolidating, refining, and standardizing leading materials databases.
AI developers are increasingly positioning their generative technologies as valuable research companions. OpenAI’s decision to allow scientists to evaluate its latest o1 model’s research applications yielded promising outcomes.
The tech industry’s longstanding aspiration of creating AI systems that can emulate scientific thinking appears closer to reality. Anthropic’s founder Dario Amodei, in an October manifesto, identified scientific domains, particularly biology, as critical areas where advanced AI could make substantial contributions. Amodei envisions a future where AI transcends mere data analysis to become a comprehensive “virtual biologist” capable of performing the full spectrum of biological research tasks. While this vision remains distant, the coming year may witness significant progress toward this goal.
4. AI companies get cozier with national security
AI firms stand to generate substantial revenue by deploying their technologies in border monitoring, intelligence operations, and various national security applications.
The US military’s AI adoption is gaining momentum through various initiatives, including the billion-dollar Replicator program—inspired by Ukraine’s conflict—focused on small drone development, and the Artificial Intelligence Rapid Capabilities Cell, integrating AI across combat decision-making and logistics operations. European military establishments face mounting pressure to boost technological investments, driven by concerns over potential reduced Ukraine funding under a Trump administration. Military strategists remain vigilant about escalating Taiwan-China tensions.
Looking ahead to 2025, these developments will continue benefiting defense-technology enterprises such as Palantir, Anduril, and similar firms, which are leveraging classified military data to enhance their AI model capabilities.
The allure of defense industry funding will increasingly draw mainstream AI companies into military partnerships. OpenAI’s December collaboration with Anduril on drone defense technology marks a significant shift from its previous stance against military engagement, completing a year-long pivot. This move aligns them with tech giants Microsoft, Amazon, and Google, who have maintained longstanding Pentagon relationships.
As AI competitors continue investing billions in model development and training, 2025 will bring heightened pressure to generate substantial revenue. While some may secure lucrative contracts from non-defense sectors – such as enterprises seeking AI agents for complex task automation or creative industries requiring sophisticated image and video generation capabilities – the appeal of defense contracts will grow stronger.
Companies will increasingly grapple with balancing their stated values against the attraction of Pentagon partnerships. OpenAI justified its policy shift by asserting that “democracies should continue to take the lead in AI development,” as stated in their policy document, suggesting that military collaboration advances this objective. This reasoning may set a precedent for others to follow in 2025.
5. Nvidia sees legitimate competition
Jensen Huang, CEO of Nvidia – now the world’s most valuable corporation – has dominated the AI chip market throughout the current boom. Nvidia’s supremacy spans both AI model training and inferencing chips, the latter being crucial for model deployment and usage.
However, 2025 could reshape this landscape through multiple factors. Industry giants including Amazon, Broadcom, and AMD have made substantial investments in chip development, with early results suggesting potential to rival Nvidia’s capabilities – particularly in inferencing, where their market position is less dominant.
Additionally, innovative startups are challenging Nvidia through unconventional approaches. Rather than pursuing incremental improvements to existing designs, companies like Groq are making bold investments in novel chip architectures. While these experimental technologies will remain in early development stages through 2025, there’s potential for a breakthrough competitor to challenge the assumption that cutting-edge AI models must rely exclusively on Nvidia’s hardware.
The geopolitical semiconductor rivalry will intensify in the coming year. Two primary strategies have shaped this competition so far. The Western nations continue to restrict China’s access to advanced semiconductors and manufacturing capabilities, while simultaneously implementing initiatives like the US CHIPS Act to strengthen domestic chip production capabilities.
The landscape could shift dramatically if Donald Trump returns to office, given his proposed aggressive tariffs on Chinese imports. Such measures would thrust Taiwan into a complex position in 2025, particularly due to its semiconductor giant TSMC’s crucial role in global chip supply. Taiwan’s recent announcement to assist Chinese companies in relocating to avoid these tariffs adds another layer of complexity. This move could potentially strain relations with Trump, who has previously questioned America’s defense commitments to Taiwan.
While the ultimate outcome remains uncertain, these developments will likely accelerate the industry’s push to reduce dependency on Taiwan – precisely what the CHIPS Act aims to achieve. As funding from this legislation begins flowing into various projects, 2025 should provide the first concrete indicators of whether these investments are successfully boosting domestic semiconductor manufacturing capacity.
