ChatGPT launches
Dialogue-tuned AI goes mainstream overnight, resetting expectations for search, productivity and customer service in one stroke.
The Story of AI So Far
How AI evolved since
ChatGPT launched.
On 30 November 2022, a chatbot demo changed the trajectory of software. This is the journey from that moment to today's multimodal, agentic, regulation-shaped AI: the milestones that mattered, why they mattered, and what they mean for your business.
Chapter zero
Why November 2022 changed the AI market
AI research had been accelerating for a decade, but it lived in papers and demos. ChatGPT did something different: it put a frontier model behind a text box anyone could use. One hundred million people arrived within months, the fastest adoption of any consumer product in history.
What followed wasn't just bigger models. It was a category change: AI became a platform, woven into search, browsers, code editors, phones and business workflows. The story below tracks that transformation in four chapters.
2022 – 2023
The Spark
On 30 November 2022, ChatGPT turned a research curiosity into the fastest-adopted consumer product in history. Within a year, AI had moved from the lab into search engines, browsers and boardrooms, and the race was on.
-
-
AI comes to Bing & Edge
The first major search and browser integration signals that AI will live inside existing software, not just separate chat apps.
-
GPT-4 arrives
The first widely deployed multimodal model raises the ceiling on reasoning and reliability, and accepts images, not just text.
-
Function calling
AI learns to use tools. The design pattern behind every modern agent, automation and structured workflow starts here.
-
Llama 2 goes open
A serious open-weight model family becomes free for commercial use, igniting the open-source AI ecosystem.
-
Mistral 7B punches up
A compact European model outperforms its size class, shifting the conversation to efficiency and deployability.
-
Bletchley Declaration
Twenty-eight governments put frontier AI safety on the geopolitical agenda for the first time.
-
Gemini 1.0 launches
Google repositions its entire model stack around a natively multimodal family: Ultra, Pro and Nano.
2024
The Acceleration
Models stopped being chatbots. They learned to see, hear, speak in real time, read entire books in one go. And the law finally caught up, with the EU AI Act becoming the first binding framework in a major market.
-
Gemini 1.5: the long-context leap
Million-token context windows turn models from short-form assistants into document and codebase workers.
-
Claude 3 family
The Haiku / Sonnet / Opus tiering arrives: pick your balance of speed, cost and intelligence.
-
AlphaFold 3
Proof the AI wave is bigger than chat: model advances begin reshaping drug discovery and biology itself.
-
GPT-4o goes real-time
Live voice and vision interaction makes AI feel less like typing into a box and more like talking to a colleague.
-
EU AI Act in force
The first large-scale, binding legal framework for AI. Governance shifts from principles to obligations.
-
Claude learns to use a computer
AI moves from describing interfaces to operating them: clicking, typing and completing tasks on screen.
-
ChatGPT search
Source-linked web search lands inside the chat box, and the line between chatbot and search engine dissolves.
2025
The Age of Agents
The market stopped asking "what can it write?" and started asking "what can it do?" Reasoning models, coding agents and open challengers from unexpected places made AI that plans, acts and verifies the new normal.
-
DeepSeek-R1 shocks the market
A fully open reasoning model at a fraction of expected cost forces every lab to rethink its economics.
-
Claude Code & hybrid reasoning
Coding agents enter real developer workflows; models learn to decide when to think hard and when to answer fast.
-
Agents SDK
Agent development gets standard building blocks: unified tools, state and web search out of the box.
-
Llama 4
Open-weight, natively multimodal mixture-of-experts models push context length and efficiency higher.
-
Claude 4
Long-running autonomous tasks and agent workflows become the headline capability, not the demo.
-
GPAI obligations apply
General-purpose model rules take effect: compliance becomes a real product constraint, not a future worry.
2026
Where We Are Now
Agentic AI is no longer a frontier feature. It is the product. Million-token context, computer use and any-input-to-any-output generation are becoming standard, and the question for businesses has changed from "should we?" to "how, exactly?"
-
Claude Sonnet 4.6
Agentic coding, planning and million-token context windows become standard frontier features, not headlines.
-
Gemini 3.5 & Gemini Omni
"Agentic Gemini": stronger action-taking and any-input-to-any-output media generation, announced to 8.5 million monthly developers.
-
Sovereign, self-hosted agents
Command A+ and Mistral Medium 3.5 make multilingual, self-hostable agentic AI practical without hyperscaler compute.
The bigger picture
Five shifts that explain everything above
From text to everything
Models stopped being text-only. Today's systems see images, hear speech, hold live conversations and generate any output from any input. The interface barrier between humans and software is dissolving.
From paragraphs to libraries
Context windows grew from a few pages to over a million tokens. AI went from answering questions to reading your entire contract pile, codebase or case history in one sitting.
From answers to actions
Tool use, function calling and computer use turned models from talkers into doers: planning, searching, clicking, executing and reporting back. This is the shift that makes business automation real.
From exotic to economical
Efficiency breakthroughs (sparse architectures, small high-performance models, prompt caching) collapsed the cost of serving powerful AI. Capability that cost pounds per query in 2023 now costs fractions of a penny.
From one lab to an ecosystem
Open-weight models from Meta, Mistral, DeepSeek and others created real choices: on-premise deployment, data residency, customisation and freedom from any single vendor. Picking well is now a strategic decision.
So what?
What the next phase of AI means for your business
The winners of this era aren't the companies with the cleverest prompts. They're the ones who turn fast-moving capability into reliable systems: lead capture that never sleeps, documents that process themselves, follow-up that never slips.
The technology is ready, and every milestone above proves it. The hard part is choosing the right combination of models, tools and guardrails for a real workflow, and building it so it stays reliable. That's the work we do.
Book an AI strategy session Free consultation: where AI genuinely pays in your business, and where it doesn't. Get an AI workflow audit We map where your hours go and identify the automations with the fastest payback. Build a pilot with our team An assistant, agent or automation, built, tested and live in your business.
Questions
Frequently asked questions
What happened after ChatGPT launched?
Within months, AI moved from a standalone chatbot into search engines, browsers and office software. GPT-4 brought multimodality, function calling brought tool use, and by 2024–2026 the focus had shifted to agentic AI (systems that plan and act, not just answer) alongside the first binding regulation such as the EU AI Act.
When did multimodal AI become mainstream?
GPT-4 accepted images in March 2023, but the mainstream moment came across 2024: Google's Gemini family was natively multimodal, Claude 3 added vision, and GPT-4o made real-time voice and vision interaction a standard consumer experience.
Which AI models are open-source or open-weight?
Meta's Llama family, Mistral's models, Google's Gemma and DeepSeek's R1 are the best-known open-weight releases. "Open-weight" is the precise term: the weights are downloadable, but licence terms vary, which matters when choosing models for commercial deployment.
What does the EU AI Act mean for businesses using AI?
The Act entered into force in August 2024, with prohibited-practice rules applying from February 2025 and general-purpose AI obligations from August 2025. Most UK businesses using AI tools face indirect effects through suppliers, but anyone deploying AI in or into the EU should understand their risk category.