2025 is set to be remembered as a defining year for artificial intelligence (AI), marked by groundbreaking research across reasoning, advanced multimodality, model efficiency, and AI hardware.It was also the year AI researchers were thrust into the spotlight, with many of them being aggressively pursued by tech giants such as Meta to staff newly formed research labs. Eye-popping salaries and million-dollar signing bonuses were offered to secure top AI talent, who were being poached and traded with the same intensity as professional athletes.
From vibe-coding to the definition of general intelligence, several research themes also sparked intense debates this year. A growing number of researchers also questioned whether training and scaling large language models (LLMs) alone can lead us to artificial general intelligence (AGI) – a hypothetical state of intelligence where AI systems perform tasks on par or better than humans.
As Ilya Sutskever, one of the founders of OpenAI, recently said, “…in some sense we are back to the age of research.” Here are the highlights of the year’s AI breakthroughs and discoveries that could set the stage for an even more game-changing 2026.
1. DeepSeek’s cost and energy-efficient models
In late January 2025, Chinese AI startup DeepSeek launched its open-weight AI model, DeepSeek-R1, triggering a massive sell-off in AI stocks as investors worried that it could perform as well as models developed by frontrunners OpenAI and Google while using fewer resources.
Nvidia’s shares plummeted by 17 per cent in a single session, and close to $600 billion was wiped off from the chipmaking giant’s market cap, making it the largest-ever single-day drop for a US company.
But beyond the market reaction, the launch of DeepSeek R1 sent shockwaves across Silicon Valley because it challenged the conventional perception of how investment-intensive the “training” process of an LLM is, before the next stage — “inference” — can be achieved.Story continues below this ad
Inference refers to the process by which AI models generate responses to queries after they have gone through training on vast volumes of data scraped off the internet. It marked a major inflection point in AI development, as the company found ways to extract more compute from less advanced Nvidia H20 GPUs by combining advanced machine learning techniques such as distillation, mixture of experts (MoE), and multi-head latent attention (MLA).
DeepSeek’s breakthrough was also celebrated because its V3 and R1 reasoning models were open-weighted, meaning anyone could locally deploy them on their own hardware. Its researchers also published a technical paper detailing the model development process.
2. Gold medal-winning AI models at IMO 2025
While LLMs are widely known for their ability to churn out essays and other forms of text in seconds, researchers have also been developing AI models to crack math problems for several years.
2025 saw a clearer sign of progress as two AI models – developed by OpenAI and Google DeepMind – achieved scores high enough to win gold medals at the International Mathematical Olympiad (IMO) 2025, a prestigious math competition for high school students.Story continues below this ad
It was the first time any AI model achieved such a high level of success on these kinds of problems, and could accelerate progress in pure mathematics and even help crack long-standing, unsolved research challenges in fields like cryptography and space exploration.
3. AI image generator behind viral Studio Ghibli art trend
AI image generators improved dramatically in 2025, moving well beyond the era of abstract renderings and glitchy amalgamations to produce visuals that are far more coherent, detailed, and realistic.
In March, OpenAI released a new feature called ‘Images for ChatGPT’, unwittingly sparking a frenzy of people using the AI tool to transform their selfies, family portraits, wedding pictures, and photos of their pets into animated images in a familiar aesthetic known as Ghibli art.
The viral social media trend led to a record surge in ChatGPT users, with OpenAI CEO Sam Altman joking that their GPUs were melting. Average weekly active users breached the 150 million mark for the first time this year, driven by the Ghibli art trend, according to data from market research firm Similarweb.Story continues below this ad
The native image generator within the chatbot was powered by OpenAI’s GPT-4o model. It churned out high-quality images because GPT-4o took a step-by-step, top-to-bottom autoregressive approach to image rendering, rather than the diffusion process used by other AI models. Another noteworthy feature was the model’s ability to use an existing photo or uploaded image as a starting point to generate artwork.
4. Embrace of the model context protocol (MCP)
Several AI companies in 2025 shipped browsers, shopping tools, and other products centred on AI agents, LLM-powered autonomous systems capable of completing tasks on their own. While major enterprises have adopted AI agents to boost productivity, their widespread use for everyday, consumer-related tasks has yet to take off.
One of the reasons AI agents are still slow and less reliable is that they are tasked with navigating and pulling information from the web, which is designed for humans rather than machines. This is where the Model Context Protocol (MCP) comes in. It allows chatbots and AI agents to securely connect to external data sources and take real actions.
Although the MCP was developed by Anthropic in 2024, the AI startup behind Claude donated the MCP standard to the Linux Foundation in December 2025, where it will be managed by the newly formed Agentic AI Foundation (AAIF). AAIF comprises Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg.Story continues below this ad
Competing big tech companies such as Microsoft and Google also announced support for MCP for their services this year as part of a broader, fundamental shift in how the internet operates.
5. First AI model trained in space
In December 2025, Nvidia-backed AI startup Starcloud announced that it had successfully trained the first generative AI model using GPUs aboard a satellite in low Earth orbit. The model is a fine-tuned variant of Gemma, Google’s open-weight small language model (SLM). It has also been integrated with the satellite’s telemetry sensors, measuring its altitude, orientation, location, and speed. This lets users on Earth query the chatbot about the satellite’s location and receive updates such as ‘I’m above Africa and in 20 minutes, I’ll be above the Middle East.’
In addition to Gemma, Starcloud said it used the space-based H100 chip to train NanoGPT, an SLM created by OpenAI founding member Andrej Karpathy, on the complete works of William Shakespeare.
Amid criticism that big tech’s insatiable demand for data centres and AI infrastructure is straining Earth’s resources, tech companies are exploring out-of-the-box solutions, including outer space, to build data centres and run AI models without drawing millions of litres of water daily or producing substantial greenhouse gas emissions on Earth.Story continues below this ad
By demonstrating that AI models can be trained and run on GPUs aboard solar-panel-fitted satellites in orbit, Starcloud shows that the science-fiction-like concept is not as outlandish as it seems. It could mark the birth of an entirely new industry. However, there is still a long path ahead and several bottlenecks to resolve along the way.