The State of Large Language Models in 2025: What Has Actually Changed
Context windows grew, prices fell, and the benchmark scores moved less than the marketing implies. A look at which changes are real and which are rounding errors.
Articles
What Inference Actually Costs, and How to Stop It Surprising You
Per-token pricing looks trivial until it multiplies by traffic. The bill is controllable, and most of the control is in decisions made early.
Risk Tiers: The Shape Almost Every AI Regulation Is Taking
Jurisdictions differ on the detail and agree on the structure: obligations scale with what the system is used for, not with how it was built.
Who Owns the Output? The Copyright Question in Two Halves
There are two separate legal arguments running, they have different answers, and conflating them is why the debate goes in circles.
Automation and the Jobs That Actually Change
The useful unit of analysis is the task, not the occupation. Almost no job is fully automatable, and almost every job contains tasks that are.
Bias Is a Product Decision Before It Is a Technical One
Most of the harm attributed to biased models is decided upstream: what the system is asked to predict, and what data was available to predict it from.
Building AI Products for Users on Bad Connections
Streaming responses, hopeful reconnects and generous payloads all assume a connection that a great many users do not have.
The Data Work Behind the Models, and Who Does It
Annotation, moderation and evaluation are done by people, a great many of them in East and West Africa, and the conditions attached deserve more attention than they get.
Retrieval-Augmented Generation, Explained Properly
The idea is simple and the failure modes are not. Most disappointing implementations fail at retrieval, not at generation.
Evaluating an AI Feature Before You Ship It
You cannot unit-test a language model, which is not the same as being unable to test it. A small hand-built evaluation set is worth more than any public benchmark.
What a Longer Context Window Actually Buys You
Context windows have grown enormously. The ability to use everything in them has not grown at the same rate, and the gap matters.
Why Models Make Things Up, and What Genuinely Reduces It
Confident invention is a consequence of what these systems are trained to do. It can be reduced substantially and it cannot be eliminated by prompting.
The AGI Conversation Is Crowding Out Better Questions
Arguing about a hypothetical future system is more comfortable than arguing about the deployed ones. That is largely why it is so popular.
Benchmarks Are Not Progress
A number that improves every quarter is reassuring and largely decorative. Measuring what a system is good at is much harder than ranking it.