AI trends in 2025
January 2025 updates:
- Nvidia at CES talked about how inference needs to scale for reasoning models as the current tokens/second output, while suitable for humans is too slow for models to chatter with each other.
- Nvidia at CES also talked about Physical AI and introduced an open source Cosmos model.
- Google Deepmind published a paper for inference time scaling for diffusion models.
Written by me and reviewed by gpt-4o
A breakthrough could happen while you are reading this and this whole article may need revision. Remember the times we are in w.r.t. AI
Welcome to 2025!
AI is here to stay, and it’s changing how we work, communicate, and explore the world. Let’s dive into the key trends shaping AI this year and what they mean for all of us.
At its core, large language models (LLMs) rely on three main ingredients:
- Architecture: The “blueprint” for how these models work (think Transformers or other innovative designs).
- Data: The stuff they learn from—real-world and AI-generated data.
- Hardware: The powerful tools (GPUs, cloud computing, and specialized chips) that make it all happen.
I will stick to analysing what could happen with these 3 ingredients in 2025 and how that would shift the AI world.
1. Changes in Model Architecture
The way LLMs are designed is a big deal. In 2025, the Transformer architecture is still king, with small tweaks making models faster, smarter, and cheaper. But hey, who knows? A groundbreaking new design could pop up and shake things up!
Why does this matter? Even tiny changes in architecture can make a huge difference in how accurate, efficient, and accessible these models are. If a revolutionary design does emerge, it could completely change the game.
2. The Big Players: What Are They Up To?
Everyone’s chasing the same thing: more data, bigger neural networks, and better hardware. But how much more do they need to build the next-gen models? And do we even have enough compute and data to make it happen?
Here’s a quick look at the major players:
OpenAI/Microsoft & Anthropic: These are in a race to build the most advanced models. Why? To charge the highest price per token. But here’s the catch: once open-source models catch up, their pricing power drops. So, they’re doubling down on compute, data, and even lobbying for regulations that make it harder for open-source to compete.
At the start of 2025, OpenAI has shifted focus from releasing cutting-edge models to “reasoning” models (more on that later). These models give better responses but come with higher costs for both them and their users. They will most likely stick to this shift through the year while working on the next frontier model.
Anthropic, on the other hand, has been quietly doing its thing—building solid models and experiences that users love. Their next set of frontier models have been delayed too and it will not be a surprise if they do a variation of what OpenAI did in 2025 to improve model performance. Their challenge has been expanding their user base as OpenAI keeps capturing the narrative in popular culture. In 2025, they will have to figure out how to expand their user base.
Google: Fundamentally, Google’s not really a model provider—they’re a product company. Their challenge has been balancing exposing-AI-models-as-tools with using models to improve their own services (like Search and YouTube). Gemini 2 is already released (limited release as of now) and the reviews have been great. They have also launched deep-research tool (again, limited release as of now) that uses search and LLMs to do online research for users.
2025 will be the year where Google finds their balance as both a model provider and a product company.
Meta: By open-sourcing models like LLaMA, Meta is making AI more accessible and affordable. There’s this idea that Meta wants the future to be all about AI-generated content, but I think it’s more about using AI tools to help humans create more content. Meta’s goal? Build these tools and run them at scale.
Mistral and others: These smaller players are doing their thing with open-source models and different business strategies. They’re not trying to compete at the frontier level but are carving out niches in less demanding use cases. This competition is great for us—it pushes prices down and gives us more options.
For end-users like us, all this competition means better tools, more choices, and maybe even lower prices. Who doesn’t love that?
3. Bigger, Better Models: The Role of Data & Hardware
Ever wonder what powers these massive AI models? It’s a mix of cutting-edge hardware and a ton of data. But here’s the problem: we’re running out of new data to train them.
In 2025, expect companies to hunt for more data (from wherever they can get it) and invest in synthetic data. One cool trend is test-time compute—a way to boost a model’s performance without retraining it.
Here’s how it works: the model generates multiple responses for a query, picks the best one, and repeats the process until it gets the best possible answer. For users, this means better results without needing a bigger model.
This cycle—AI generating data, training better models, attracting more users, and repeating—will drive the rise of free tools this year. Companies will offer generous free tiers or ad-supported models to get more users on board. Providers will start with ad-supported models to keep the tools free.
4. Smarter Models with Better Reasoning
AI models are getting smarter, thanks to “reasoning” capabilities. These from the outside seem like model-systems - like the test time compute architecture we discussed above.
However there’s a trade-off: reasoning models take longer to run, which means more hardware is needed to handle the same number of queries. So, expect data-centres to keep expanding.
In 2025, we could see small models use these techniques to perform better and take away use-cases from bigger models.
5. Announcements, Announcements, Announcements
One annoying trend? Overhyped announcements. Companies like OpenAI, Apple, and Google love to hold flashy events and promise features that take months (or never) to show up.
Meanwhile, open-source players and Anthropic are quietly gaining users by focusing on useful features instead of hype.
6. Enterprise AI
AI Agents are all the buzz, but I prefer to think of them as AI systems. These systems combine multiple ML techniques with LLMs to solve real business problems. Agents are definitely a big part of it - but they come with their own set of governance challenges. However, in 2025, expect enterprises adopt copilot model whole-heartedly.
Businesses will focus on governance, guardrails and will figure out which models work best for their workloads. Open source models would find their place in enterprises and self-hosting models to reduce costs should become commonplace. Technologies to manage models (k8s for models) should see growing investment.
It seems like most of the value in the information management will be captured by the provider which already manages enterprise data - Microsoft, Google, etc. Businesses on providers who fail to capture value there, will probably migrate to ones who do. This points to consolidation in enterprise data management towards the big players.
7. Consumer AI
Consumer AI still has a long way to go. The biggest challenges? Building trust, ensuring privacy, improving data access, creating novel APIs, and keeping costs reasonable. After the hype and dud that was Apple Intelligence, consumers will be skeptical of any tall claims with respect to "AI features". In 2025, incremental features will continue and there will be space for third party apps to build first class AI experiences on top of the iOS and Android. It seems the "AI first OS layer" will take some time to materialise (if it ever does).
8. China and Global Competitiveness
China is making big moves in AI, despite sanctions. They’re releasing open-source models like Deepseek and Qwen, and they’re catching up (and even surpassing) US companies in some areas.
In 2025, expect China to keep pushing ahead and disrupting the global AI landscape.
9. Robotics and Accelerated Science
AI is speeding up progress in other fields like robotics, medicine, media, and self-driving cars. This year, robots might finally become cost-effective enough to take on real jobs. Again, it seems China is winning this race too. This is the most exciting, but diverse space in the AI landscape that requires careful studying.
Conclusion: A Year of Steady Progress
2025 is shaping up to be a year of steady progress. Businesses will start seeing real value from generative AI, and while we might not get a groundbreaking revolution, the improvements in architecture, data, and hardware will keep us moving forward.