2025 marked a pivotal year as breakthroughs in large language models and multimodal systems propelled artificial intelligence from speculation to real-world deployment, sparking a global race in infrastructure, investment, and societal impact.
In 2025 the promise of artificial intelligence moved irreversibly from speculative fiction into operational reality, reshaping technology, capital flows, employment and everyday life. According to 36Kr’s year‑end review, the year saw large language models (LLMs) and multimodal systems make leaps in reasoning and agentic behaviour that transformed industry conversations about the proximity of artificial general intelligence (AGI) and the first signs of artificial superintelligence (ASI). [1]
Technical progress was conspicuous across several fronts: reasoning, multimodal processing and autonomous agents. Industry benchmarks captured that shift, Stanford’s AI Index reported models surpassing human baselines on a range of tests, and specialised benchmarks such as MMMU, designed to probe interdisciplinary, image‑and‑text reasoning at university level, showed rapidly rising scores from major models. According to 36Kr, improvements from Google’s Gemini and OpenAI’s o3 series were particularly notable. [1]
That technical advance was matched by a surge in funding and capital expenditure. Industry data show generative AI attracted roughly $33.9 billion in investment in 2025 while tech giants pushed capital expenditure toward an estimated $400 billion, prompting debate over bubbles and energy consumption. Morgan Stanley observed a parallel explosion in demand for custom silicon and ASIC development as reasoning workloads increased computational needs, underscoring that model progress is tightly coupled with infrastructure investment. [1][2]
The shift from chatbots to agentic systems also accelerated the practical deployment of AI. Goldman Sachs forecast that AI would evolve into “hybrid workers” collaborating with humans across industries, a trend visible in 2025 as agents and robotics moved into production, healthcare and knowledge work. Forbes and Global X further framed these agentic advances as logical stepping stones toward broader AGI capabilities and widespread, semi‑autonomous task performance. [3][4][7]
Open‑source and non‑Western actors played a consequential role in this new landscape. 36Kr highlighted DeepSeek, whose R1 model passed peer review and drew attention in Nature, as a major dark horse, while community efforts around LLaMA, Mistral and other toolchains reduced the barrier to entry for engineering, fine‑tuning and local deployment. At the same time, forum discussions showed the limits of distilled, locally runnable models versus cloud‑scale variants, emphasising a continuing divide between accessible tools and state‑of‑the‑art systems. [1]
The human and economic implications were stark. The OECD found nearly 40% of global employment is exposed to AI, with higher exposure in advanced economies where cognitive tasks predominate; policymakers and firms face the twin tasks of enabling productivity gains while managing workforce transition and inequality risks. Industry leaders’ rhetoric reflected both ambition and urgency: according to 36Kr, figures including Mark Zuckerberg and Elon Musk framed huge investments as preferable to falling behind, Zuckerberg saying he would rather “risk misinvesting hundreds of billions of dollars” than lag, and other CEOs offered a spectrum of timelines for AGI from imminent to multi‑year horizons. [5][1]
Voices inside the field remain divided over how far current architectures can go. While results on benchmarks and the proliferation of agentic capabilities have led some executives to declare AGI effectively achieved in narrow senses, prominent researchers such as Yann LeCun cautioned that autoregressive LLMs have limitations and will need richer sensory grounding and alternative architectures to reach true generality. The result is a contested but fast‑moving research agenda, with leading labs releasing major model updates at a cadence measured in weeks rather than years. [1][6]
If 2025 rewired expectations, the coming years will test whether those expectations are sustainable and societally manageable. Investors and corporates are doubling down on specialised hardware, agent platforms and expert models even as governments, academics and civil society push for governance, safety and equitable transition policies. According to reports from Morgan Stanley, Goldman Sachs, Forbes, the OECD and Global X, the race is now as much about infrastructure, labour strategy and regulation as it is about algorithms, and the winners will be those who align technical advance with economic resilience and public legitimacy. [2][3][4][5][7]
📌 Reference Map:
##Reference Map:
- [1] (36Kr) – Paragraph 1, Paragraph 2, Paragraph 5, Paragraph 6, Paragraph 7
- [2] (Morgan Stanley) – Paragraph 3, Paragraph 8
- [3] (Goldman Sachs) – Paragraph 4, Paragraph 8
- [4] (Forbes) – Paragraph 4, Paragraph 8
- [5] (OECD) – Paragraph 6, Paragraph 8
- [6] (Techaimag) – Paragraph 6
- [7] (Global X) – Paragraph 4, Paragraph 8
Source: Noah Wire Services
Noah Fact Check Pro
The draft above was created using the information available at the time the story first
emerged. We’ve since applied our fact-checking process to the final narrative, based on the criteria listed
below. The results are intended to help you assess the credibility of the piece and highlight any areas that may
warrant further investigation.
Freshness check
Score:
8
Notes:
The narrative presents a comprehensive overview of AI developments in 2025, with specific references to recent events and publications, such as DeepSeek’s peer-reviewed model in Nature and Meta’s challenges. The inclusion of recent data and events suggests a high level of freshness. However, the article’s extensive coverage and depth may indicate a synthesis of existing information, potentially reducing originality. The numerous citations to 36Kr’s own publications also raise questions about content originality.
