overview

powered byBuild Club

Welcome to The State of AI.

AI moves faster than any technology before it. This is everything you need to know to keep up: where it came from, where it stands today, where it's heading, and what you can actually do about it.

Last updated 23 hours ago

It's been three and a half years since ChatGPT launched.

In that time, we went from the iPhone 14 to the iPhone 17.

But we also went from a chatbot that wrote poems and made up facts to systems that operate computers better than humans, work autonomously for hours, and assemble their own teams of AI agents to complete projects overnight.

Late 2022

ChatGPT launches.

The world's first experience with generative AI. Hit 100 million monthly users in two months.

2023

AI learns to see, hear, and create.

Starts passing medical licensing exams, went from 10th percentile to 90th percentile in bar exam. Able to create and see images, audio, and video.

2024

AI learns to reason.

Inference costs for GPT-3.5-level intelligence fall 280x in 18 months. DeepSeek, a small Chinese lab, open-sources frontier-class models.

2025

AI starts doing things

AI could browse the web, write software, and operate apps on your behalf. For the first time, you could delegate real work to it.

Early 2026

Today

Late 2022

ChatGPT launches.

The world's first experience with generative AI. Hit 100 million monthly users in two months.

2023

AI learns to see, hear, and create.

Starts passing medical licensing exams, went from 10th percentile to 90th percentile in bar exam. Able to create and see images, audio, and video.

2024

AI learns to reason.

Inference costs for GPT-3.5-level intelligence fall 280x in 18 months. DeepSeek, a small Chinese lab, open-sources frontier-class models.

2025

AI starts doing things

AI could browse the web, write software, and operate apps on your behalf. For the first time, you could delegate real work to it.

Early 2026

Today

Most technology moves predictably. You can skip five years of iPhone updates and pick up the latest one without missing much.

AI doesn't work like that. Every few months, something that was previously impossible becomes routine, and the pace of innovation isn't slowing down; it's speeding up.

In the last three months alone, we saw:

Staying up to date has never been more important, or more overwhelming.

We built this to fix that.

And the good news is: you're still early.

How many organisations are actually scaling AI?

Each dot represents 0.1% of organisations.

No AI adoption (12%)
Pilots & experimentation (55%)
Scaling select functions (26%)
Scaled enterprise-wide (7%)

Source: McKinsey, “The state of AI in 2025” (Nov 2025, n=1,993). McKinsey’s follow-up “State of AI trust in 2026” reaffirms the broad shape: 88% of organisations now use AI in at least one function, but only around 6% are capturing more than 5% EBIT impact from it.

Current landscape#

Everything above was built by a handful of companies. Three years ago, only one mattered. Today there are at least six, all producing models within striking distance of each other, backed by more funding than any technology sector in history, and being worked on by the smartest minds in the world.

And these companies aren't slowing down. At the frontier, we're already seeing glimpses of what's next: Coordinated teams of agents that manage entire workflows together, AI that works autonomously for days instead of hours, machines that can see and interact with the physical world, and AI that accelerates the development of the next generation of AI.

These models power hundreds of products you can use today. ChatGPT, Claude, and Microsoft Copilot are what most people know, but they only scratch the surface of whats possible. There are hundreds of AI tools purpose-built for specific work, from legal research to sales outreach to financial modelling. The chatbot is the best place to start understanding what's possible, but it shouldn't be the end point.

Impact#

AI is already reshaping every industry and every job function.

Adoption (US firms with 250+ employees)

Source: US Census Bureau Business Trends and Outlook Survey AI Supplement, filtered to firms with 250+ employees. Reference period Nov 2025 – Feb 2026.

Sales and Marketing share a figure because BTOS combines them into one category. Design & Creative isn’t included in BTOS’s 15-function breakdown — deep-dive coverage on a separate page.

Adoption has moved faster than almost any previous technology cycle. AI spending is on track to roughly double again this year, from about 0.8% to 1.7% of revenue across surveyed enterprises, per BCG's 2026 AI Radar, even as the share of firms capturing measurable value remains in the low single digits.

The research is increasingly clear on what separates the winners from everyone else: it's not which tools they bought. It's whether they redesigned how their teams actually work. According to McKinsey's State of AI survey, the small group of high performers is nearly three times more likely than other organisations to say they have fundamentally redesigned individual workflows around AI.

Buying the tools is only one small part of the equation. Redesigning how you work is where the real leverage is.

Takeaways#

Everything above is a tool, but tools aren't leverage.

When spreadsheets first arrived, the companies that benefited most weren't the ones that bought Excel. They were the ones that stopped running their business on paper. The same thing is happening with AI. Leverage comes from finding the right tools, reinventing your workflows around them, and making sure everyone around you is keeping up. That way, as the tools improve, so does your work.

01

Try everything.

Start using AI for real work, as often as you can.

  • Pay for the premium tiers. LLMs are expensive to run and the most useful features are behind a paywall for good reason.
  • Don't overthink which tool to start with. The leading models are within a few percentage points of each other, so it genuinely doesn't matter.
  • Approach with optimism, not pessimism. Different products have different strengths. Be curious, figure out what each one is good at, and don't write anything off after one attempt.
  • Remember: today's models are the worst they'll ever be. Something that didn't work three months ago might work today. Keep going back.
  • If you lead a team, make space for experimentation. If you don't, experiment anyway.
02

Think in systems and observe your work.

Most companies bolt AI onto processes that were designed for humans doing everything manually, then wonder why the results are underwhelming.

  • Look at how you actually spend your days and ask yourself what you would hand to a capable assistant if you had one.
  • Anything repetitive, tedious, or process-heavy is worth testing against the tools that exist today.
  • If something can't be automated yet, note it anyway. What's impossible this month might be routine by next quarter.
  • Pick one workflow, rebuild it from scratch with AI in the loop, measure the result, and do it again.
03

Invest in your people, not just your tools.

The organisations seeing real results aren't just investing in AI tools, they're investing in helping their people actually use them.

  • Your team are the experts in their own work, given the space and the budget to experiment. They're the ones best placed to figure out where AI creates the most leverage.
  • The best results come from individuals who are given the space and the budget to experiment, not from top-down mandates to "use AI."
  • Find the people who are already experimenting and give them room to show others what's possible.
  • Treat upskilling as a line item, not an afterthought.

Conclusion#

The technology behind all of this was originally built to translate between languages. It turned out the most productive use was translating human language into software, and in three and a half years that has fundamentally reshaped how work gets done.

The models will keep getting smarter. But smarter tools don't create leverage on their own.

The companies pulling ahead aren't the ones using better tools. They're the ones that changed how they work.

If you want to chat about AI in your workforce or workflows, contact us.

AI Resources

A growing collection of our favourite resources to master AI.

Microsoft 365 Copilot

AI woven directly into Word, Excel, Outlook and Teams.

Want to go deeper?

These guides are free, and always will be. If you want hands-on help rolling Claude out across your team, get in touch with us.

Get in touch
the landscape

the landscape

powered byBuild Club

This section is designed to give you a zoomed-out view of the AI landscape as it stands today. We’ve broken it down into the areas we think matter most: the frontier labs building the most powerful models, the money flowing in, what’s happening to the workforce, how governments are responding, and where real adoption is actually happening.

Frontier Labs

Frontier labs are the companies building the models that power everything else in AI. Three years ago, the only one that mattered was OpenAI. Today there are at least six, all producing models within striking distance of each other, backed by hundreds of billions in capital.

Tap any logo to go deeper.

Since ChatGPT launched, frontier AI has been improving exponentially. The length of tasks AI can reliably complete on its own is doubling every four months. Two years ago, the best models topped out at tasks that took about ten minutes. Today, they can handle work that would take a human expert most of a day. And what runs at the absolute frontier reaches consumer hardware within 6 to 12 months.

Doubling time
~4 months
Length of tasks AI can complete
Late 2022 (ChatGPT launch)
~36 seconds
GPT-3.5
Today (May 2026)
~12 hours
Claude Opus 4.6 (highest publicly released)
OpenAIAnthropicGoogleAbove 16h: unreliable

* Mythos was not released publicly. Anthropic deemed it too dangerous to deploy and gave restricted access only to security partners. METR was granted evaluation access. METR also notes that measurements above 16 hours are at the edge of their current task suite’s reliability (hatched band).

1GPT-5.5 Pro160(157164)
2GPT-5.5158(156163)
3GPT-5.4 Pro158(156162)

Source: METR Horizon benchmark v1.1, 50% reliability threshold (p50 horizon length, raw in minutes). Linear Y-axis in hours — note the unmistakably exponential climb. All 26 evaluated models shown; state-of-the-art releases are connected and labelled. Doubling time since 2023 is approximately four months (128 days, per METR’s own fit; 188 days across the full all-time series). The benchmark measures how long a task can be before frontier AI’s chance of completing it correctly drops below 50%. Mythos Preview added to the METR dataset on 8 May 2026. Top-10 leaderboard: Epoch AI Capabilities Index (ECI), late May 2026 — a composite of ~40 benchmarks scaled so Claude 3.5 Sonnet = 130 and GPT-5 = 150; CI values are 90% confidence intervals. Five ECI points correspond roughly to a doubling of METR’s autonomous-task time horizon. Live data at epoch.ai/eci.

No single model leads everywhere. As of May 2026, this is where each one is strongest. It shifts with every major release.

The Workforce

In 2023, 25% of workers were using AI regularly. By early 2025, KPMG found 87% of American workers were using it weekly. 51% daily. The adoption curve has been steep. What companies are doing about it has not kept pace.

Adoption is near-universal
87%
of US workers use AI at least weekly
KPMG, early 2025
75%
of knowledge workers use AI at work
Microsoft Work Trend Index — 31K workers, 31 countries
88%
of organisations are using AI in some form
McKinsey, 2025
Organisational readiness is not
7%
have fully scaled AI across the enterprise
McKinsey, 2025
16%
of companies have redesigned jobs around AI
Deloitte — 3,235 leaders, 24 countries
<50%
are making significant adjustments to talent strategies
Deloitte, 2025

Broad but shallow#

Microsoft’s Work Trend Index, covering 31,000 workers across 31 countries, found that 75% of knowledge workers now use AI at work. Nearly half started less than six months ago.

Deloitte surveyed 3,235 business leaders across 24 countries and found that workforce access to AI tools expanded by 50% in a single year, from under 40% to around 60% of workers with sanctioned access. But among those with access, fewer than 60% actually use it in their daily workflow. That gap has barely moved since last year.

McKinsey found 88% of organisations are using AI. 7% have fully scaled it across the enterprise.

The skills premium#

PwC’s AI Jobs Barometer analysed close to a billion job ads across six continents. Workers with AI skills command a wage premium of over 30%, up from 25% the year before. In roles most exposed to AI, skills are changing 66% faster than in other occupations, up from 25% faster a year earlier.

The World Economic Forum estimates 39% of existing skills will be disrupted by 2030. Its data suggests AI will create 97 million new jobs while displacing 85 million. Whether the people losing the jobs are the same people gaining the new ones is a different question.

The redesign gap#

Deloitte found that 84% of companies have not redesigned jobs around AI. The biggest barrier to AI integration isn’t the technology. It’s insufficient worker skills. Yet fewer than half of companies are making significant adjustments to their talent strategies. Most are educating employees on AI fluency. Far fewer are rearchitecting roles, workflows, or career paths.

36% of companies expect at least 10% of their jobs to be fully automated within a year. 82% expect that within three years. But the jobs being automated first — data entry, reconciliation, first-level customer support — are often the entry point for longer careers. Automating them without creating alternative pathways doesn’t just remove tasks. It removes the bottom rung of the ladder.

Where the results are showing#

McKinsey found that the 6% of companies seeing real financial impact from AI are 3.6x more likely to be aiming for transformative change, 2.8x more likely to have redesigned workflows, and 3x more likely to have strong senior leadership commitment.

Deloitte found that 34% of companies are using AI to deeply transform their businesses. Another 30% are redesigning key processes. The remaining 37% are using AI at a surface level with little change to how they operate. All three groups report productivity gains. Only the first group is building something structurally different.

37%
30%
34%
Surface-level useRedesigning processesDeeply transforming

All three groups report productivity gains. But productivity alone isn’t translating to financial impact for most.

6%
high performers
Report >5% of earnings attributable to AI
McKinsey defines these as organisations reporting both significant value and measurable EBIT impact. The remaining 94% are using AI but not yet transforming with it.
Aim for transformative change, not just efficiency
New business models and revenue streams, not incremental cost cuts
3.6x
more likely
Senior leadership actively drives AI adoption
Role-modelling use, not just approving budgets
3x
more likely
Redesigned workflows end-to-end around AI
55% of high performers vs ~20% of others
~3x
more likely
Invest >20% of their digital budget in AI
Over one-third of high performers allocate at this level
5x
more likely
25%
Report transformative impact, doubled from last year
84%
Have not redesigned jobs around AI capabilities
#1 barrier
Insufficient worker skills, not the technology itself

Sources: Deloitte, “State of AI in the Enterprise” (Jan 2026, n=3,235 across 24 countries). McKinsey Global Survey on the State of AI (Nov 2025, n=1,993 across 105 countries).

Adoption

88% of organisations say they use AI, up from 78% a year ago. But most are still running pilots. Only a third have begun scaling beyond experiments. Just 39% report any measurable impact on profits, and only 6% attribute more than 5% of their operating earnings to AI. The gap between companies buying AI tools and companies getting results from them is wide, and it hasn’t closed as fast as the adoption numbers suggest.

88%of organisations use AI
Using AI
At least one business function
~55% still experimenting or piloting
33%
Scaling
Begun deploying beyond pilots
Scaling doesn’t guarantee financial results
39%
Seeing profit impact
Any effect on operating earnings
Most of those report less than 5% of profits
6%
Transformative results
>5% of profits from AI + significant value
7%
Have fully scaled AI across the entire organisation
79%
Use genAI specifically, up from 33% in 2023
23%
Scaling agentic AI in at least one function

Source: McKinsey Global Survey on the State of AI (Nov 2025), 1,993 participants across 105 countries. “Profit impact” refers to enterprise-level EBIT. “High performers” defined as >5% EBIT attributable to AI plus reported significant value.

The pattern in the data is consistent. The companies seeing real returns aren’t the ones that bought the most tools. They’re the ones that redesigned how work gets done. McKinsey found that high performers are 2.8x more likely to have restructured workflows around AI, not just added AI to existing ones. Most companies are still bolting AI onto processes that were designed for humans working alone. That’s why the tools are everywhere and the results aren’t.

Where results are real#

The functions seeing the strongest cost savings from AI are software engineering, manufacturing, and IT, where many organisations report reductions of 10 to 20%. Revenue gains are highest in marketing and sales, strategy and corporate finance, and product development, where a significant share report uplift above 10%.

But across the board, the scale is still modest. Most cost savings sit under 10%. Most revenue gains are less than 5%. Only 39% of organisations report any measurable impact on operating earnings, and just 6% attribute more than 5% of those earnings to AI. The results are spreading across more functions every year. For most companies, they haven’t yet reached a scale that transforms the business.

Strongest cost savings
Software engineering
10–20% reductions reported
Manufacturing
10–20% reductions reported
IT
10–20% reductions reported
Strongest revenue gains
Marketing & sales
Significant share report >10%
Strategy & corporate finance
Significant share report >10%
Product development
Significant share report >10%
39%
of organisations report any measurable impact on operating earnings from AI
6%
attribute more than 5% of those earnings to AI — McKinsey’s “high performers”

Source: McKinsey & Company, “The State of AI in 2025,” Nov 2025 (n=1,993, 105 countries), Exhibits 5–8.

The pilot trap#

Only 25% of companies have moved 40% or more of their AI experiments into production. A pilot can run with a small team using clean data. Production requires infrastructure, integration, security reviews, compliance, and ongoing maintenance. Three-month timelines stretch to eighteen months. Because pilots are cheap and low risk, companies keep launching new ones rather than scaling what already works.

The agents wave#

The history section of this article traced agents from AutoGPT’s failed promise in 2023 to the working systems of 2025. The adoption data shows what happened next.

Deloitte found that 23% of companies are using agentic AI at least moderately today. Within two years, that number is expected to reach 74%. McKinsey found the same 23% figure independently. Customer support is the highest-impact use case, followed by supply chain, R&D, and cybersecurity. 85% of companies expect to customise agents to fit their specific business needs.

But governance is lagging behind deployment. Only 21% of companies report having a mature model for governing autonomous agents. Unlike traditional AI that recommends actions for humans to take, agents act directly. They make purchases, send communications, modify systems. The companies moving fastest on agents are starting with lower-risk use cases and building governance frameworks before scaling. The ones rushing to deploy widely without those foundations are taking on risk they may not fully understand.

85%
expect to customise agents to their business
42%
still developing their agentic strategy roadmap
35%
have no formal agentic AI strategy at all

Sources: Deloitte AI Institute, “The State of AI in the Enterprise: The Untapped Edge,” Jan 2026 (surveyed Aug–Sep 2025, n=3,235 across 24 countries). Risk percentages, use case rankings, 85% customisation, and 21% governance from the same report. Strategy roadmap data (42%, 35%) from Deloitte 2025 Emerging Technology Trends study. McKinsey independently confirmed 23% agentic scaling (Nov 2025, n=1,993).

What companies are spending#

Enterprise spending on generative AI was $1.7 billion in 2023. It grew to $11.5 billion in 2024. It hit $37 billion in 2025. A 22x increase in two years. It now represents 6% of the entire global SaaS market, the fastest-growing software category in history.

Coding and developer tools account for $7.3 billion of that, with half of all developers now using AI daily. At least 10 AI products now generate over $1 billion in annual recurring revenue. More than 50 have crossed $100 million. Three years ago, almost none of these products existed.

Gartner estimates total worldwide AI spending across all categories will reach $1.5 trillion in 2025, grow to $2 trillion in 2026, and hit $3.3 trillion by 2029.

2023 spend
$1.7B
Year of experimentation
2025 spend
$37B
3.2x year-on-year
2-year growth
22x
Fastest category in history
Applications $19BInfrastructure $18B
6%
Rest of SaaS market ~$370B

$37B of ~$400B total SaaS — achieved in 3 years since ChatGPT’s launch

Source: Menlo Ventures, “State of Generative AI in the Enterprise” (2023, 2024, 2025). SaaS market size per Gartner / Statista / Precedence Research.

What’s being bought#

Coding and developer tools lead at $7.3 billion, with half of developers now using AI daily. General-purpose copilots account for $8.4 billion. Industry-specific solutions account for $3.5 billion, led by healthcare, which is adopting AI at 2.2 times the rate of the broader economy.

In 2024, 47% of AI solutions were built internally. By 2025, 76% were purchased. Ready-made tools are reaching production faster than custom builds. AI software conversion rates are double traditional SaaS. Deals close at twice the speed.

Horizontal$8.4B
Copilots
$7.2B
Agent platforms
$750M
Productivity tools
$450M
Departmental$7.3B
Coding
$4.0B55% of dept. spend
IT operations
$700M
Marketing
$660M
Customer success
$630M
Design + HR
$870M
Vertical$3.5B
Healthcare
$1.5B2.2x economy rate
Legal
$650M
Other verticals
$640M
Creator tools
$360M
Government
$350M
2024
53%
47%
PurchasedBuilt internally
2025
76%
24%
PurchasedBuilt internally
47%
AI deal conversion rate, vs 25% for traditional SaaS
50%
of developers now use AI coding tools daily
$550M→$4B
Coding tools spend in one year (7.3x growth)

Source: Menlo Ventures, “2025: The State of Generative AI in the Enterprise” (Dec 2025), survey of 495 U.S. enterprise AI decision-makers.

The trajectory#

Enterprise AI spending tripled in a single year. The share of companies reporting transformative impact doubled from 12% to 25%. 84% of organisations are increasing AI investments. Revenue from AI remains mostly aspiration: 74% hope to grow revenue through AI, only 20% are doing so today.

The pattern is familiar from every previous technology wave. Early adoption is broad and shallow. A smaller group figures out how to use it properly and gains a compounding advantage. Then everyone else scrambles to catch up. The difference this time is the speed.

market map

powered byBuild Club

market map

AI adoption looks uneven across functions. Some teams have raced ahead, others have barely started, and even the leaders are still figuring out how to convert tools into results.

Adoption (US firms with 250+ employees)

Source: US Census Bureau Business Trends and Outlook Survey AI Supplement, filtered to firms with 250+ employees. Reference period Nov 2025 – Feb 2026.

Sales and Marketing share a figure because BTOS combines them into one category. Design & Creative isn’t included in BTOS’s 15-function breakdown — deep-dive coverage on a separate page.

DepartmentSummary
Engineering
Spending grew sevenfold in a year, but experienced developers got slower.
Finance
20.5%
Usage quadrupled in a year. Only 14% of CFOs can point to clear returns.
Marketing
Teams doubled their AI content output. Consumer willingness to engage with it halved.
Sales
A third of B2B companies cut their sales teams last year. Win rates fell anyway.
Product
Entry-level PM hiring dropped 73%. Every enterprise product team adopted AI tools.
Support
AI handles most routine queries now. The companies that celebrated loudest are hiring humans back.
HR
Adoption nearly doubled. Cost-per-hire went up, not down.
Legal
10.5%
Four in five lawyers use AI. Court sanctions for fake AI citations went from 10 to 700.
Operations
More companies abandoned their AI initiatives in 2025 than completed them.
Design
Half of creatives use AI daily. 42% of companies scrapped most of their AI projects this year.

It's late 2022. You get a message from a friend.

chatgpt.com
You have to try this

You're skeptical but you give it a try.

And it does it.

It actually rhymes. It's funny. It sounds weirdly human.

You share it with friends, with colleagues.

You get it to write a work email. It's better than anything you'd write yourself.

Until the cracks show.

You ask it a math question and it gives you a confident answer that's completely wrong.

You ask it to cite a source and it invents one out of thin air.

AI headline (1 of 9)AI headline (2 of 9)AI headline (3 of 9)AI headline (4 of 9)AI headline (5 of 9)AI headline (6 of 9)AI headline (7 of 9)AI headline (8 of 9)AI headline (9 of 9)

The headlines shift.

You scope it down to grammar checks, the occasional email and a funny poem

Fast forward 3 years

A child born when ChatGPT launched hasn't even started preschool yet.

We went from the iPhone 13 to the 16.

A week ago, someone in Sydney used AI to design a personalised cancer vaccine for his dying dog. The tumour shrank 75% within a month.

Hundreds of thousands of people are texting an AI assistant that reads their emails, makes phone calls, books restaurants. It runs in the background 24/7, doing things before they even ask.

Its been 3 years, and it feels like it's just getting started. How on earth did this happen?

Let's start at the beginning.

Late 2017

They published it openly. For anyone to use.

And a small nonprofit in San Francisco saw something Google missed.

OpenAI took Google's paper and had an idea: if you trained it on enough text — books, websites, conversations, code, the entire written output of civilisation — it might learn to do anything expressible in language.

So they built the first Large Language Model (LLM)

Every AI system in this story is built on this idea. And it's simpler than you'd think.

Feed it an enormous amount of text and train it to predict the

That's it. Text in, text out. Always predicting the next most likely word.

OpenAI's bet was that once the model had seen enough of how humans write and think — next-word prediction might start to look like understanding.

So they tested it.

One year laterJune, 2018

GPT-1

117 million parameters.

A research experiment. Nobody noticed.

Eight months laterFebruary, 2019

GPT-2

1.5 billion parameters.

Thirteen times bigger. It could write semi-coherent paragraphs.

OpenAI deemed it “too dangerous to release.”

One year laterJune, 2020

GPT-3

175 billion parameters.

A hundred times bigger again. It could write essays, generate code, hold conversations. But it was a developer tool.

The general public had no idea it existed

Two and a half years laterNovember, 2022

A simple chat interface.

Sam Altman announcing ChatGPT

OpenAI put GPT-3.5 in a simple chat interface. Free. No setup. Just a text box.

Internally, they called it a “research preview.” They expected maybe a million users and didn’t expect it to matter very much.

And here we are in late 2022: your friend's message, the poems, the amazement — and then the disappointment when you realised it would confidently make things up and couldn't do basic maths.

But while the magic was fading for you, the world was reorganising.

The technology you'd scoped down to grammar checks was about to receive more money and attention than almost anything in modern history.

And OpenAI now had the resources to push it to its absolute limit.

The model you'd been using was GPT-3.5.

Four months laterMarch, 2023

GPT-4 Releases

Suddenly, next word prediction felt like actual intelligence.

It passed medical licensing exams. It outperformed 93% of human candidates on coding interviews.

And people were scared.

Future of Life Institute open letter calling for a six-month pause on AI development, signed by over 33,000 researchers

Over 30,000 researchers signed an open letter calling for an immediate six-month pause on AI development.

Nobody paused.

The opposite happened. More money poured in. More businesses launched.

Within months, every major company had an AI strategy, or was panicking because they didn't.

But it was still just text.

You typed words, and the AI typed words back.

Six months laterSeptember, 2023

AI learns to see — and create.

By late 2023, AI could see. Photograph your fridge and ask what to cook, upload a chart and have it explained.

And it could generate images, audio, video.

It turned out the same approach that predicted the next word could predict the next pixel.

The technology had a name that was suddenly everywhere: generative AI.

And a whole new wave of businesses exploded around this.

September 2023 – September 2024

We had reached the plateau.

New models kept coming. Slightly faster, slightly cheaper, but they were all fundamentally the same.

The consensus formed quickly: we've hit the ceiling.

That was a completely reasonable conclusion based on everything available at the time.

September, 2024

AI Learns To Think

There had been a clue. Researchers noticed that if you asked these models to "think step by step," they performed significantly better. Like a student who shows their working on an exam instead of jumping straight to the answer.

But the models weren't actually thinking. They were producing text that looked like reasoning because you'd asked them to.

Then OpenAI released a model called o1.

For the first time a model had been trained to genuinely reason — to break down problems, test approaches, backtrack when something wasn't working, and think internally, sometimes for minutes, before responding.

And just like that, the ceiling shattered.

The ceiling had been real. But it was a ceiling on one approach.

Reasoning was an entirely new one.

And suddenly the "just next word prediction" argument that sceptics had relied on for two years was dead overnight.

Not because they were wrong about what they'd observed. But because they'd assumed those limitations were permanent.

But reasoning had two big problems. It was way too expensive for everyday use.

And nobody outside OpenAI knew how to build it. It still required billions of dollars and the most advanced chips on Earth to compete.

Four months laterJanuary, 2025

Reasoning Becomes Open Source and Commoditised

DeepSeek R1 release announcement

A Chinese AI lab called DeepSeek quietly released a new model.

DeepSeek wasn't a tech giant. It was a spinoff from a hedge fund in Hangzhou, China. Its founder, Liang Wenfeng, had stockpiled NVIDIA chips back in 2021, before US export controls made it illegal to sell the best hardware to China.

With those older chips, a fraction of the budget, and zero venture capital, they built a reasoning model that matched OpenAI's o1 on major benchmarks.

Then they did something that none of the major labs had.

First page of DeepSeek-R1 paper

They published the full methodology. Open-source. For anyone on earth to replicate.

7 days laterJanuary 27, 2025

The Market Responds

It turned out that making AI cheaper didn't shrink the market. It exploded it. When the cost of intelligence drops, demand doesn't fall. It compounds. Companies that couldn't afford AI suddenly could. Use cases that had been too expensive to justify suddenly made sense.

Just days later, OpenAI released o3-mini.

It was 93% cheaper than o1, and it actually

The pressure DeepSeek created had made the cost problem collapse almost overnight.

After reasoning, everything accelerated.

The shift was subtle at first. AI helped people work faster. Draft an email, summarise a document, get unstuck on a problem.

Then at some point it became something else. People weren't just using it to speed things up. They were relying on it. And the idea of going back to working without it was starting to feel impossible.

But no matter how smart, how cheap, or how capable these models got, the interaction was always the same.

You type. It types back. You decide what to do next. It could think. It could reason. It could be built by anyone.

But it couldn't do anything.

Early 2025

AI started doing things.

The idea wasn't new. Back in 2023, a project called AutoGPT went viral with the promise of AI that could act autonomously.

It couldn't.

The models weren't good enough.

But by early 2025, everything that had been built over the previous two years converged at once.

The models could see well enough to read a screen, reason well enough to plan, code well enough to build tools on the fly and hold enough in working memory to stay on task for hours instead of seconds.

In January, OpenAI launched Operator, an AI that could open a browser and complete tasks on your behalf.

In February, Anthropic released Claude Code, an AI that could write, test, and ship entire software projects. Within six months it was generating over $1 billion a year in revenue.

Then there was Manus. A startup that gave AI an entire computer it controlled however it wanted. Two million people joined the waitlist in days. Invite codes resold for thousands of dollars. Nine months later, Meta acquired it for over $2 billion.

The word "agentic" entered the lexicon overnight.

A year earlier, you'd give an AI a task and check back in 30 seconds.

Now you could give it a complex project and check back tomorrow, and the work would be done.

Not a report about the work. The work itself. Code written, emails sent, databases queried, tools built from scratch when no existing one fit. All while you slept.

By early 2026, the leading models were completing tasks that take human experts over fourteen hours.

And that brings us to right now.

It's been three and a half years since ChatGPT launched.

In that time, AI went from writing bad poems to passing the bar exam. From text-only to generating photorealistic video. From needing billions of dollars to being replicated on restricted hardware for a fraction of the cost. From answering questions to assembling its own teams of AI agents and completing projects overnight, without being asked.

The examples from the beginning of this piece don't seem so impossible anymore.