Topic · Themes
Understanding AI: all our articles
All 122 Alyriaz articles on the “Understanding AI” topic: news, analysis and explainers, newest first. To stay up to date, you can get an alert as soon as a new article on this topic is published.
What is data poisoning? When a few hundred texts are enough to trap a model.
Data poisoning involves inserting elements into a corpus that are designed to disrupt a model trained on it. We explain the mechanism, why it works with very small quantities, and the debate it raises.
AI labs are buying books by the pallet, and destroying them to scan them.
AI labs are bulk-buying physical books published before 2022, touted as the last source of guaranteed human-written text. The industrial scanning destroys the volumes. We explain why this rush is happening, what it says about the state of data, and what it threatens.
OpenAI cuts its prices by 80%, and it's the model itself that figured out how
On 4 August 2026, OpenAI cut the price of GPT-5.6 Luna by 80% and that of Terra by 20%, explaining that these gains come from optimisations the model itself contributed to. We analyse what this loop reveals, and what it changes for developers.
What is an AI's "temperature"? The setting that decides whether it's sensible or inventive
Temperature is the setting that controls the degree of randomness in an AI's responses. We explain how it works, why a low temperature doesn't guarantee truth, and how to choose it based on your use case.
Seven everyday uses of AI that genuinely make a difference
Far from the talk of revolution, here are seven concrete uses of AI where the time saved is real and verifiable, and three situations where it's better not to use it. A practical article, without jargon.
AI agents speed up scientific code by 60x. The problem is, they seem to be right even when they're wrong.
A field report conducted by OpenAI and academic partners shows that coding agents can modernise neglected research software, with speed-ups of up to 60 times. The same report warns that these systems produce convincing and sometimes false results. We explain this tension.
Who owns what an AI writes for you?
Does a text, image or code produced by an AI really belong to you? We look at what the terms of use say, what copyright law says, and why these two answers don't coincide at all.
Why Claude, ChatGPT and Gemini don't look alike
Large models rely on very similar technologies, but adopt clearly distinct tones. We explain where this personality comes from, at which stage it is crafted, and what it implies when choosing a tool.
What is C2PA? The invisible passport that will accompany every image
The C2PA is the technical standard that makes it possible to attach a verifiable history of origin to an image, video, or audio file. We explain how it works, why it upends the logic of detection, and what it cannot do.
Why AI almost always tells you you're right
AI models tend to validate what you say rather than contradict you. This flaw, known as complaisance or sycophancy, stems directly from how these models are trained. We explain the mechanism and the phrasings that can elicit an honest opinion.
What is an "embedding"? The technique that lets a machine understand that "doctor" and "physician" mean the same thing.
An embedding transforms text into a list of numbers that encodes its meaning. We explain this key concept, how it enables semantic similarity search, and why it is the invisible engine behind most enterprise document assistants.
Why would an AI capable of solving a mathematical conjecture get two numbers wrong when adding them?
AI models excel at advanced mathematics yet sometimes fail at basic arithmetic. Here's why: they don't calculate—they predict text, and numbers are tokenised in a way that makes arithmetic surprisingly difficult.
What is "prompt caching"? The setting that could halve your AI bill
Prompt caching lets you avoid paying full price for the same context sent with every request. We explain the mechanism, the real savings, the pitfalls of cache lifetime, and why it's the first lever to pull before September's price hikes.
On 1 September, Claude Sonnet 5 increases by 50%. And there's a catch almost no one mentions.
The promotional rate for Claude Sonnet 5 ends on 31 August 2026: from 1 September, input and output costs rise by 50%. But the real trap lies elsewhere: the new tokenizer counts more tokens for the same text. We explain the actual calculation and how to prepare now.
Why an AI can't say "I don't know"
An AI states something false with exactly the same tone as when it tells the truth. We explain why uncertainty is structurally difficult for these systems to express, what labs are trying to do about it, and what that means for your trust.
How to verify what an AI tells you: the five-reflex method
An AI produces false information with the same confidence as accurate information. Here's a practical five-step method for spotting what needs verification, doing it quickly, and knowing where vigilance matters most.
What is an AI "jailbreak"? The security problem no one has solved yet
A jailbreak involves bypassing an AI model's guardrails to make it produce content it should refuse to generate. We explain what this phenomenon is, why it is structurally difficult to prevent, what labs are doing to protect against it, and why an incident of this kind led to a model being suspended worldwide.
Why does your French cost more than English when you use an AI?
For equivalent content, writing in French consumes more tokens than English, which increases cost and reduces the available space in the context window. We explain why this inequality exists, what it changes in practice, and how to limit it.
Kimi K3 burns through twelve times more reasoning tokens than Claude Opus: the hidden cost of thinking
An analysis of Kimi K3's reasoning traces reveals it consumes more than twelve times as many reasoning tokens as Claude Opus 4.8. This is both the secret to its performance and its biggest economic flaw. We explain the mechanism, and why the price per token says nothing about the real cost.
What is a "Mixture of Experts"? The architecture that explains why giant models don't cost that much
Mixture of Experts, or mixture of experts: this architecture lets a model have trillions of parameters while only activating a small fraction per request. We explain the mechanism, why it now dominates, and what it means when you read a spec sheet.
Other topics in “Themes”
- AI agents77 articles
- Code and developers22 articles
- Images and video20 articles
- Voice and music15 articles
- Robots and self-driving cars9 articles
- Health15 articles
- School and education16 articles
- Jobs and work34 articles
- Laws and regulation59 articles
- Cybersecurity32 articles
- Safety and risks51 articles
- Privacy and data45 articles
- Disinformation and deepfakes34 articles
- Energy and climate15 articles
- Chips and data centers40 articles
- Research and science43 articles
- Money and markets53 articles
- AI prices and costs43 articles
- Geopolitics and sovereignty58 articles
- Open source and local AI44 articles
- Culture and creativity42 articles
- Everyday life32 articles
- Philosophy and ethics97 articles
- Benchmarks and comparisons30 articles