Technology, explained

How large language models actually work

A language model does one thing: predict the next chunk of text given everything before it. Trained on enormous quantities of text, it learns statistical structure — grammar, common facts, argument shapes, code patterns — without ever storing a database it can look things up in.

That single design choice explains almost all of its behaviour. It is fluent on every subject, because fluency is what it learned. It is reliable on well-covered material, because that material appeared often enough for the patterns to be strong. And it is confidently wrong at the edges, because a low-confidence answer and a high-confidence one come out in the same authoritative voice. There is no internal alarm that fires when it does not know.

The practical rule follows directly: treat model output as a strong first draft written by someone widely read but occasionally mistaken, and who never says "I'm not sure." Verify anything load-bearing — numbers, names, dates, citations, legal or medical specifics — against a primary source.

What models are genuinely good at

• Transformation. Rewriting, summarising, translating, reformatting, changing tone. The source material is in front of it, so there is little to invent.
• First drafts. Getting from a blank page to something to react to.
• Explanation. Restating a well-established concept at a different level of expertise.
• Structured extraction. Pulling fields out of messy text into a consistent shape.

Where they fail

• Precise recall. Exact figures, citations and quotations are where fabrication concentrates.
• Current events. Training has a cutoff; without a live search tool the model is describing the past.
• Arithmetic at length. Multi-step calculation drifts unless the model can run actual code.
• Knowing its own limits. Confidence in the output does not track accuracy.

Buying a laptop or phone without overpaying

Marketing pushes attention toward the processor, because it is the easiest thing to put a number on. It is rarely what makes a device feel old.

• Memory and storage age faster than the chip. Software grows to fill both. Buy more RAM before a faster processor at the same price.
• Battery health, not battery size, decides year three. Cells degrade by charge cycle. A replaceable or cheaply serviceable battery is worth more than a slightly larger sealed one.
• The screen is the thing you look at all day. It is also the component least often compared. Brightness and colour accuracy matter more than resolution beyond a certain density.
• Last year's flagship usually beats this year's mid-range at the same price, on build quality, screen and camera alike.
• Check the support window before you buy. A phone with four more years of security updates is a longer-lived device than one with faster benchmarks and two.

The four habits that stop most attacks

Almost all consumer compromise traces back to one of four things: a reused password, missing two-factor authentication, unpatched software, or a convincing message. Three of the four are solved by tooling you set up once.

1. A password manager. Its real value is not strength, it is uniqueness. One breached site then cannot open any other account you own.
2. App-based two-factor authentication. An authenticator app or hardware key, not SMS — SIM-swap attacks defeat text messages specifically.
3. Automatic updates. Most exploited vulnerabilities were patched before they were widely abused. The gap that gets people is the delay.
4. A pause. The fourth cannot be automated. Legitimate organisations never need your one-time code, and manufactured urgency is the oldest tell in fraud. Slowing down is the control.

Notice what is missing from that list: antivirus software, VPNs and browser extensions. They are not useless, but they are far down the list of things that would have prevented a typical breach.

Why everything moved to the cloud — and some of it moved back

Cloud computing converted a capital expense into an operating one. Instead of buying servers, you rent capacity by the hour and give it back when you are done. For an unpredictable workload that is decisive: you stop paying for a peak you only hit twice a year.

The trade is that rented capacity carries a margin. For a workload that is steady, heavy and always on, that margin compounds into a large recurring bill, and owned hardware can cost materially less. This is why some mature companies now move part of their infrastructure back — not a reversal of the cloud, but a correction where the original economics never applied.

The right answer is almost always a mix, decided per workload rather than per company.