ARTIFICIAL INTELLIGENCE
The technology reshaping how we think, work, create, compete, and live.
Artificial intelligence is software that learns patterns from very large amounts of data, then uses those patterns to predict, generate, decide and act. It is not magic and it is not a mind. It is statistics, hardware and engineering at a scale that only recently became possible — and that scale is now rearranging industries, research and geopolitics.
These are orientation figures, not a live feed. Anything on this page that is a number, a ranking or a market claim carries a label: FACT ANALYSIS ESTIMATE FORECAST.
What is AI, in plain language
Ten ideas that cover most of what people mean when they say "AI". Open any card for the longer version and an everyday example. No prior knowledge assumed.
Each step is a narrower idea sitting inside the one before it. Every large language model is a deep learning system; not every deep learning system is a language model.
Seventy-five years, in steps
AI is not new. It has moved through waves of optimism, collapse and revival. Select a milestone to read what happened and why it mattered.
Why AI suddenly got good
Nothing about the mathematics changed overnight. Six things arrived at once, and each one multiplied the others.
The uncomfortable part for researchers: much of the recent progress came less from clever new ideas than from doing a small number of known things far bigger than anyone had tried. That pattern may not continue forever, and several labs now argue the next gains come from reasoning and data quality rather than raw size.
Not all models do the same job
"AI model" covers a dozen different kinds of system. Here is what each type is for, with well-known families as examples. Named families are illustrative, not a ranking — capability claims change monthly and belong in benchmark reports with dates on them.
The organisations
Fourteen of the groups that matter most, filtered by what they primarily do. Select any card for a full profile. Founding dates and headquarters are matters of record; strategy descriptions are editorial analysis.
The AI race
The competition is not really about chatbots. It runs across eight layers at once — models, chips, cloud capacity, consumer reach, agents, robotics, open weights and enterprise distribution. A company can lead one column and be absent from the next.
From chatbots to agents
A chatbot answers. An agent is given a goal, decides on its own steps, uses tools to carry them out, checks the result and tries again. That difference is where most of the current engineering effort sits.
What an agent actually has
Where it goes wrong
Animated illustration of a typical agent cycle. Real systems repeat the middle four steps many times before returning a result.
What's happening now
The active fronts, grouped by where the work is happening. Select a card to read what is actually changing and what is still unresolved.
AI news
A newsroom layout wired to a swappable data layer. It is seeded with dated, checkable entries from the public record — not a live feed, and nothing here is invented. Connect a real feed where the code says so and the same cards will render today's stories.
NEWS_SEED with a fetch
from an RSS-to-JSON service or your own endpoint — the integration point is commented in the script
(loadNews()). Until then, treat this as a layout demonstration.
The compute war
Training a frontier model means running trillions of arithmetic operations across thousands of specialised chips for weeks, in buildings that consume as much electricity as a small town. Compute is the hard constraint on modern AI, which makes it the real battleground.
The physical layer nobody sees
Who makes the silicon
The stack, bottom to top
Every chatbot reply rests on ten layers, starting with a power station. Select a layer to see what happens there and who operates it.
The AI economy
Where the money goes, in shape rather than in dollars. The charts below show relative proportions and directions of travel, not audited figures.
Where AI capital tends to concentrate
The pattern of an infrastructure build-out
Every large infrastructure cycle — railways, fibre, mobile networks — has spent heavily in advance of returns. Some paid back extraordinarily; some did not. Which one AI turns out to be is genuinely unsettled.
Will AI take our jobs?
The honest answer is: it depends on the task, not the job title. Most roles are bundles of tasks, and AI absorbs some of them while making others more valuable. Here is where a range of work currently sits between full automation and human-plus-machine.
ANALYSIS Positions are editorial judgement based on how much of the work is routine, verifiable and low-stakes. Labour-market outcomes also depend on regulation, cost, liability and how quickly organisations actually change — historically the slowest variable of all.
The other side of AI
Not doom, and not dismissal. These are the problems that researchers, regulators and the companies themselves treat as real, along with what is currently being done about each.
The global AI race
AI capability now sits alongside energy and semiconductors as a matter of national strategy. Select a region for its position on research, chips, regulation and infrastructure.
Open weights, closed models
One of the field's sharpest disagreements, and one where reasonable people land in different places. Note that "open" in AI usually means downloadable weights, not full open source: training data and code are often still withheld.
Closed / proprietary
The model runs on the provider's servers. You reach it through an API or product; the weights never leave.
ARGUMENTS FOR
- Misuse can be monitored, rate-limited and revoked after release
- Safety mitigations can be updated centrally for every user at once
- Costly training runs have a clearer route to being paid for
- Tight integration with products and infrastructure
ARGUMENTS AGAINST
- Outside researchers cannot inspect what they cannot download
- Dependence on a vendor's pricing, availability and policy changes
- Capability concentrates in a small number of organisations
- Data leaves your environment unless specific arrangements are made
Open / open-weight
The trained weights are published for download. Anyone can run, inspect, fine-tune or modify the model on their own hardware.
ARGUMENTS FOR
- Independent scrutiny of behaviour, bias and failure modes
- Runs locally — useful for regulated data, offline work and low latency
- Specialisation on domain data without sending it anywhere
- Lowers costs and spreads capability beyond a few firms
ARGUMENTS AGAINST
- Safety training can be removed by anyone with the weights
- Release is irreversible: a published model cannot be recalled
- Running and securing it becomes the deployer's responsibility
- Licences vary widely; several "open" models carry real restrictions
In practice most serious organisations use both — closed models where capability and support matter, open weights where control, cost or privacy matter. The debate is unresolved because it turns on empirical questions about misuse that nobody has settled yet.
The next five years
FORECAST Everything in this section is speculation. It is informed by current research directions, but the AI field has an unbroken record of surprising forecasters in both directions. Read it as a set of open questions, not predictions.