The 2025 AI year in review

The year AI learned to act

Welcome to our end of year edition.

If it feels like AI moved faster this year than at any point before, you are not imagining it. 2025 was the year AI stopped being something you tried occasionally and started being something you could hand real work to.

This newsletter is our attempt to make sense of what actually mattered. Not every launch or headline, but the shifts that changed how people work, decide, and build. We have pulled together the moments that shaped the year, along with a few practical lessons we have seen in real client work.

If you only skim one thing before the break, make it this.

If 2023 was the year of the chatbot, and 2024 the year of scale, 2025 will be remembered as the year AI moved from answering questions to completing work.

Not prompts. Tasks.

Whole pieces of work, from research and analysis to coding and operations.

This shift is already changing how work gets done, how companies compete, and where value is created. Below is a clear, practical summary of what mattered this year, and why.

1. The efficiency shock and the reasoning race

The first real surprise of 2025 came from China, not Silicon Valley.

In January, DeepSeek released R1, a reasoning model that matched top tier performance at a reported training cost of around six million dollars. That release shattered a core assumption, that progress required ever larger, ever more expensive compute clusters.

It proved something else mattered just as much, algorithmic efficiency.

Western labs accelerated work on reasoning systems. xAI released Grok 3, Anthropic introduced hybrid reasoning in Claude 3.7, and OpenAI released GPT-5 with automatic routing between fast responses and deeper reasoning.

2. From chatbots to agents

The biggest change was not smarter answers, it was AI taking action.

Agents arrived, software that can use tools, browse the web, click buttons, fill forms, and complete tasks end to end.

OpenAI launched Operator and later Atlas. Amazon pushed enterprise reliability with Nova Act. Anthropic pushed agents into software development with Claude 4, which could open repos, run tests, and deploy fixes with minimal human input.

3. Power became the real bottleneck

As models improved, compute stopped being the main constraint. Energy did.

Project Stargate signalled the start of the gigawatt era, and the industry began treating grid access as a competitive advantage.

Nvidia’s dominance grew further, alongside a clear message for 2026 planning, AI progress is now limited by grids, not models.

4. AI enters the physical world

2025 was also the year AI gained a body.

Robotics made visible progress, and wearables moved beyond novelty. Intelligence started moving out of screens and into the world.

Meta’s AI enabled glasses signalled hands free interaction at scale, and new form factors began competing with the smartphone as the default interface.

5. Society caught up

The cultural and political impact of AI became impossible to ignore.

Regulation accelerated, deepfakes went mainstream, and the debate shifted from “is this coming” to “who is responsible”.

AI was no longer abstract. It was personal.

Signals that mattered for leaders

Model progress felt AGI like, but advantages copied fast.
Major features were copied in weeks. Differentiation shifted toward pricing, UX, ecosystem fit, and trust.

Agents moved from experiments to workflow defaults.
Agent tools changed how people worked day to day, and multi hour task delegation became normal.

Context engineering replaced heavy prompting.
Connectors, memory, and runtime context reduced copy paste, while trust and control became core design constraints.

AI became a GDP scale force.
Capital flowed into infrastructure. Most spend still went to hardware, and individuals often saw impact before organisations.

Search and buying behaviour shifted.
AI mediated discovery became normal. Companies now sell to humans and to AI systems acting on their behalf.

What this means

2025 closed the experimentation phase. Reasoning is reliable. Agents can act.

Intelligence is now a utility you call through an API, wear on your face, or watch walk across a factory floor.

The real question for 2026 is no longer whether AI works. It is how fast organisations, governments, and individuals adapt to a world where delegating to machines is normal.

A practical tip to upgrade your AI results...

In our client training, we teach RTFC (Role, Task, Format, Context) as a simple, reliable framework for prompting in order to get the best results from AI. The limitation is obvious, it is hard to think of all the right context upfront.

The upgrade is letting the AI help you build the prompt.

Start with by typing your best effort RTFC prompt, then hit shift+enter and add the following text:

“The above is a prompt I want you to improve. Ask me five questions, one at a time, to improve this prompt. Once I have answered, rewrite the prompt and then run it.”

The same pattern works with transcripts. Instead of going straight to meeting notes, ask the AI to ask you three questions first, then generate the summary.

We are learning not just to ask better questions of AI, but to get AI to ask better questions of us.

 Looking ahead for Fambirai

For Fambirai, looking ahead to 2026, we will be releasing the first in a series of interactive online courses.

The first course is designed for teachers, with others planned for different roles and contexts. If you know any teachers who would like early access in return for their feedback, please share the link below with them or get them to email mattd@fambirai.com

Early access to Fambirai's AI for Teachers course

Welcome to the agentic era.

Thanks for reading.

If there is something specific you would like us to cover in future newsletters, or if you have feedback on this edition, let us know.

If you would like help with your AI strategy, or training for your team, get in touch. You can also view past clients and testimonials on our website.

Cheers,
The Fambirai Team