If you understand one technical idea about AI, make it this one, because it explains both why these tools are astonishing and why they occasionally get things confidently wrong.
A prediction machine, not a database
A large language model is, at its core, a probabilistic model of language. It has been trained on an enormous amount of text, and what it learnt from all that reading is deceptively simple sounding: given everything written so far, what word is most likely to come next?
When you send a message, the model does not look up an answer in a database. It generates a response one small chunk at a time, each chunk chosen from a probability distribution over everything it could say next. Write "the capital of France is" and the probability mass lands overwhelmingly on Paris. Write "the best way to structure my landscaping quote is" and the model draws on every pattern it has absorbed about quotes, landscaping, and business writing to produce something genuinely useful that has never been written before.
Here is the part that surprises people: to get good at predicting text about the world, the model had to internalise a huge amount about how the world works. Predicting the next word in a physics explanation badly requires knowing no physics. Predicting it well requires something that functions a lot like understanding. That is why a "next word guesser" can debug your spreadsheet formula and draft your shareholder update.
Why they get things wrong
The same mechanism explains the failure mode. Because the model is generating plausible text rather than retrieving verified facts, it can produce something that sounds completely authoritative and is simply wrong. People call this hallucination. The practical implications for you:
- Give the model the source material when facts matter, paste in the document, the email thread, the data, rather than trusting its memory
- Verify anything important before it leaves your hands, especially numbers, names, dates, and legal or medical claims
- Treat it like a brilliant, fast, occasionally overconfident colleague, not an oracle
Modern models also mitigate this themselves by searching the web and citing sources, which helps a great deal, but the habit of verification is still yours to keep.
Why models are different from each other
If they all predict text, why do Claude, ChatGPT, and Gemini feel different? Because a finished model is shaped by hundreds of choices: what data it was trained on, how large it is, how much computing power went into training, how long its memory within a conversation is, what tools it can use, and crucially, the fine tuning that happens after the initial training, where the raw predictor is taught to be helpful, follow instructions, refuse harmful requests, and carry itself a certain way. That last stage is where a model's judgement and personality come from, and it differs a lot between labs. Models are not interchangeable, and the differences show up most on long, complex, multi step work.
Why I use Claude
I have used all the major models, and I still test them regularly. My daily driver, and the platform D3ploy standardises on, is Claude, made by Anthropic. My reasons are practical:
Judgement over long tasks. The work I care about is rarely one question and one answer. It is multi step: read this codebase, plan the change, make it, test it, fix what broke. Claude holds the thread over long, complex work better than anything else I have used, and that reliability compounds.
Claude Code. Anthropic's coding harness is, for my money, the best way to build software with AI that currently exists. I effectively live in it, and a later article covers how.
Writing quality. A large share of business value from AI is simply words: proposals, emails, documentation, strategy. Claude writes like a thoughtful adult rather than a press release.
Trust posture. When you deploy AI into other people's businesses, the boring things become the main things: how the provider handles data, how the model behaves at the edges, how honestly it flags uncertainty. Anthropic's whole positioning is safety and reliability first, and that aligns with putting my name behind client systems.
None of this means the others are bad, they are not, and healthy competition is why everything keeps improving. But you get compounding returns from going deep on one platform rather than shallow on four, and Claude is the one I chose. The next article breaks down the different Claude models and which one to use when.
Claude showing multiple plausible continuations of one phrase