10 Big Moments in AI in 2025: the tl;dr of what happened

Happy Holidays! I hope you are cozy and well fed during the last couple of days of 2025.

If you were muddled by the numerous headlines in AI in 2025, here is the condensed version to fully bring you up to speed with what happened in the AI space so you can skip all the reading.

1. AI shifted from chat to agents

Without a doubt, agents dominated the airspace this year. The most consequential shift of 2025 was the move away from conversational AI toward agents that could take action. Instead of asking AI for answers, users increasingly delegated work. Advanced reasoning capabilities (e.g., chain-of-thought processing) were integrated into major models from OpenAI (o-series updates), Google, and others. These enabled step-by-step problem-solving, boosting performance in coding, math, and planning. Scheduling, coding, research, document review, purchasing, and workflow execution became core use cases. This reframed the entire industry. Success was no longer measured by eloquence or cleverness, but by reliability, permissions, and the ability to operate across tools and systems. AI stopped being reactive and started behaving like labor.

2. Multimodal became the default, not the differentiator

In 2025, multimodal AI crossed from impressive to expected. Image understanding was baseline. Voice interaction became normal. Video generation matured enough to be used in marketing, training, and media workflows rather than novelty demos. AI was no longer something you typed into. It became something you spoke to, showed things to, and increasingly watched create. Any model that could not handle multiple modalities felt incomplete.

3. China emerged as the global leader in open source AI

In January 2025, (can you believe it was almost a year ago?) Chinese startup DeepSeek launched the open-source reasoning model DeepSeek-R1, matching or exceeding top Western models at a fraction of the cost. This shocked the industry, crashed Nvidia's stock temporarily, and signaled China's rapid rise in open-source AI leadership. Though we have not seen another update from Deepseek as momentous as what happened in January, China’s dominance in open source became undeniable this year. Beyond DeepSeek, Chinese firms like Alibaba and Moonshot flooded the market with free, high-performance open-source models, eroding U.S. dominance and influencing global AI ecosystems. These models were optimized for real world use rather than benchmark prestige. For startups, governments, and enterprises outside the US, Chinese open source models often became the practical default. This marked a shift in global AI power away from frontier exclusivity and toward adoption at scale.

4. Infrastructure and compute became the real bottleneck

While software became more efficient, the hunger for hardware reached a fever pitch. Nvidia’s Blackwell architecture officially became the nervous system of the global AI economy, powering nearly every major data center. However, demand so far outstripped supply that it triggered a "Compute Crisis."

The bottleneck shifted from just chips to the physical infrastructure required to run them. The "Energy Wall" became the primary constraint on scaling, forcing hyperscalers like Microsoft, Amazon, and Google to invest over $100 billion collectively into energy solutions, including controversial revivals of nuclear power and massive grid modernizations.

5. Regulation moved from theory to execution

The EU AI Act became the world’s first comprehensive legal framework for artificial intelligence.

Its core philosophy is a "Risk-Based Approach": the stricter the risk to human rights or safety, the stricter the rules.

The Risk Pyramid

There is a long list of banned technologies (basically anything that is a Black Mirror episode technology, most likely banned in the EU). Labelled as Unacceptable Risk include:  

• Social Scoring: Governments cannot score citizens based on behavior or personality.

• Real-Time Biometric ID: Police cannot scan faces in public crowds in real-time (with narrow exceptions for terrorism or missing persons).

• Untargeted Scraping: Companies cannot scrape images from the internet or CCTV to build facial recognition databases (a direct hit to business models like Clearview AI).

• Emotion Recognition: Banned in workplaces and schools.

• Behavioral Manipulation: AI designed to exploit vulnerabilities (e.g., age, disability) or manipulate decisions subliminally

Remember when  GPT-4o’s Voice Mode was very sing-song and emotive? That was fully available in the US. In the EU, the "Emotion Recognition" ban in the AI Act has forced OpenAI to release a "flattened" version of the voice that is less expressive and cannot mimic human emotion as accurately. Companies that violate the policies face a fine of up to €35 million or 7% of global turnover for using prohibited AI practices.

6. Gemini 3 signaled a shift in market power

Google’s launch of Gemini 3 marked a turning point. In the beginning of the year ChatGPT had about a 80% market share in AI usage with Google’s less than 5%. As of Dec 2025, Gemini seems to have risen to about 20%. Inside the industry, this moment was widely viewed as code red for OpenAI, not because OpenAI fell behind technically, but because Google demonstrated how powerful AI becomes when paired with massive distribution. Paired with viral tools like Nano Banana (image generation/editing) and Veo 3 (video with audio), it drove massive user engagement and positioned Google strongly in consumer AI.

7. Claude reframed AI around Skills, not personality

Anthropic releases Skills for Claude, a folder-based system that lets organizations bundle workflows, procedures, and executable scripts into packages their AI assistant can autonomously access when handling specific tasks. Do you know the scene from the Matrix where Neo learns Kungfu  in a matter of seconds? That is similar to training the Skills features within Claude. By packaging expertise into folders rather than code, Skills can better infuse enterprise knowledge and workflows with real agentic capabilities.

8. OpenAI experimented with commerce and advertising

In parallel with advances in agents and multimodal capabilities, OpenAI began quietly exploring new business models. Signals around advertising, shopping, and product discovery suggested a willingness to position ChatGPT as a commerce interface. AI started influencing not just answers, but intent and transactions. This marked a philosophical shift. OpenAI began to look less like a research lab and more like a platform negotiating its place in the attention and commerce economy.

This holiday, I did a lot of my holiday shopping with ChatGPT. Lots to improve on! But it is much better than the Amazon browsing experience I’d tell ya.

9. The AI browser makes AI ambient

This year we saw the prominence of the AI browser. On October 21, 2025, OpenAI introduced ChatGPT Atlas, a web browser with ChatGPT built into its core. The browser allows ChatGPT to accompany users anywhere across the web, helping them directly in the window where they're working, understanding context, and completing tasks without requiring copy-pasting or leaving the page. A few months before that, we saw Comet from Perplexity.

Both companies have tried to reimagine how people interact with the web by embedding AI directly into browsing rather than treating it as a separate tool. This shift comes as traffic from bots is expected to surpass human traffic in coming years, fundamentally changing the economics and design of web browsing.

10. Hype Correction and Enterprise Reality Check

This, we also saw this major headline: about 95% of enterprise AI initiatives never deliver on their original business goals. Instead of transforming operations or delivering measurable ROI, most projects remained stuck in pilots, proof of concept phases, or produced outputs that were technically interesting but operationally unusable.

This statistic became a reality check for the entire industry because it came from multiple independent sources, including analyst firms, CIO surveys, and internal benchmarking studies, which all converged on a similar pattern. The takeaway was not that AI doesn’t work. It was that building AI capabilities is fundamentally harder than building models.

AI in 2025 Conclusion

So it’s been a wild year! Remember, as technology gets exponentially better, it’s hard for us humans to fathom the pace of progress, since we are more acclimated to linear paces of change. AI literacy will become more important than ever. If any of the above was confusing to understand, then here is the AI course I designed to bring you up to speed.

Hello! I’m Sharon Gai, a keynote speaker on AI and its effects on workers and the future of work.