Home/Blog
Articles
Worth reading first
Swipe for more
Evaluation & evidence
The EU delayed the part with no standards
Three AI Act obligations take effect today and the widely reported headline says the opposite. What was deferred, what was not, and why the split falls exactly where…
August 2, 2026
Safety & governance
2.6 million robotic surgeries, none of them autonomous
July 30, 2026
Foundations
Nine subjects, and the model was almost never the blocker
August 2, 2026
Language & meaning
Prompt engineering in 2026: what still works and what changed
July 18, 2026
Browse by subject
Foundations
29
Theory, history and the parts that will still be true in twenty years.
Language & meaning
21
How language models handle language, and the places they do not.
Agents
11
Building agents, deploying them, and why most attempts do not reach production.
Evaluation & evidence
73
Measurement, benchmarks, and how to tell whether a claim is worth anything.
Building with AI
18
Retrieval, cost, infrastructure and the practical business of shipping something.
Safety & governance
30
Alignment, security, regulation and who is accountable when it goes wrong.
The three foundational guides
The long answers to the questions everything else builds on.
GuideWhat is artificial intelligence?
AI is the field of making machines do things that seem to require intelligence, and that definition has moved every…
GuideWhat is machine learning?
Machine learning is programming with examples instead of rules, an approach that took over because writing the rules…
GuideWhat is AGI?
AGI, artificial general intelligence, is a system with broad human-level capability across domains. It's also a term…
Evaluation & evidence
72 seconds, or 30 minutes, and both are trials
Territory 11 opens on the first subject this corpus has examined where the evidence is genuinely good. Registered trials,…
August 2, 2026
Evaluation & evidence
The 12% was non-inferior, and P was 0.41
MASAI is the best-evidenced AI deployment in medicine and the headline everyone quoted describes a result the trial did…
August 2, 2026
Evaluation & evidence
Refactoring fell from 25% to 3.8%
Survey evidence says AI improves code quality. Repository telemetry says the opposite. They are measuring different…
August 1, 2026
Evaluation & evidence
Phase I improved. Phase II did not.
AI-designed drugs clear safety trials at well above industry rates. At the stage that tests whether a drug works, the…
August 1, 2026
Evaluation & evidence
Adding the doctor to the model changed nothing
A randomised trial found physicians did better with an LLM than with conventional resources. Its second comparison,…
August 1, 2026
Evaluation & evidence
The pilot failed and the staff deployed it anyway
Enterprise AI is measured by what organisations sanctioned. A separate literature measures what their employees actually…
August 1, 2026
Evaluation & evidence
The pilot ran on data the production system will never see
Enterprise AI pilots are built on a curated slice, pre-cleaned, with limited users and manual review. Then production data…
July 31, 2026
Evaluation & evidence
The framework that exists produced 1.6%
The FDA has two AI tracks. One is final, has authorised over 1,350 devices, and is the regime under which almost none of…
July 31, 2026
Evaluation & evidence
One trial, a waitlist control, and a letter
The best evidence for AI mental health support is a single randomised trial of a purpose-built clinical tool. Its own…
July 31, 2026
Evaluation & evidence
The lock-in is the prompts, not the API
Switching costs used to require board approval to incur. AI switching costs accumulate through ordinary engineering…
July 31, 2026
Evaluation & evidence
621,000 robots installed, virtually no humanoids
Territory 7 opens on the question article 117 left unresolved. Industrial robotics is enormous and growing. The…
July 30, 2026
Evaluation & evidence
220 million miles, inside a boundary Waymo drew
The strongest safety evidence in physical autonomy, and the methodology that makes it honest is also what limits what it…
July 30, 2026
Written something like this?
Artifipedia takes long-form pieces from people who work in or study the subject. A documented failure, evidence from your own domain, or a widely repeated claim that does not survive checking. Unpaid, properly edited, permanent byline.
Read the contributor guidelinesFoundations29
- Nine subjects, and the model was almost never the blocker
- Where the attention mechanism actually came from
- I spent a month writing down what every AI concept depends on
- Why AI aces the test and fails the variation
- Why neural networks generalize when theory says they can't
- Who invented deep learning, and why it took so long
- What an AI confidence score actually means
- What AI cannot do, no matter how capable it gets
- Where AI has not landed: 77% report no use case
- What actually caused the AI winters
- Regularization: L1, L2, dropout and what unites them
- What linguistics predicts about where AI fails
- Why AI is bad at math, and which failures are permanent
- What is reinforcement learning? Learning from reward
- Neurosymbolic AI: what the symbolic half gets right
- Does it understand? The argument, properly stated
- How to tell if your AI actually works
- How neural networks work: the idea under all of it
- AI scaling laws: why bigger wins, and whether it's ending
- The model was right. Acting on it would have killed people.
- What is generative AI? The complete guide
- What is deep learning? The complete guide
- What are embeddings? How AI turns meaning into numbers
- Why does deep learning work? The generalization mystery
- What is computer vision? How machines learn to see
- Supervised vs unsupervised learning: the four types
- AI vs machine learning vs deep learning: the difference
- What can AI do, and what can't it? A predictive map
- What is the Turing test? And why it stopped mattering
Language & meaning21
- Prompt engineering in 2026: what still works and what changed
- How AI generates images: from noise to a picture
- How to read a model release
- Why AI answers the question you asked, not the one you meant
- How AI generates video: from noise to motion
- Why AI thinks hot and cold mean the same thing
- How a sentence becomes an answer: an LLM end to end
- Context engineering: the skill that replaced prompt engineering
- How reasoning models work: AI that thinks before answering
- Why AI hallucinates: the confident lie is a feature, not a bug
- How transformers work: the architecture that ate AI
- Why AI gives different answers: sampling and temperature
- What is multimodal AI? One model that sees, reads, and hears
- What is a context window? Why bigger isn't better
- What is a token? How AI reads text in chunks
- Who said that? Why diarization is harder than transcription
- Why AI works worse in your language
- What is a large language model (LLM)? Complete guide
- What is natural language processing (NLP)?
- Why long context windows fail: the lost-in-the-middle problem
- How to reduce AI hallucinations: what actually works
Agents11
- Why AI agents fail: the seven failure modes
- Agent memory: what should persist, and what should not
- How AI agents actually work: the loop behind the hype
- What to measure before you deploy an agent
- Human in the loop is weaker than it sounds
- What is MCP? The standard that wired AI into everything
- Your agent works in the demo because the demo is the easy case
- Most agent projects will be cancelled. This is why.
- How AI memory works, and why the model remembers nothing
- Multi-agent AI gets worse as you add agents
- Will AI take my job? What the evidence shows
Evaluation & evidence73
- The EU delayed the part with no standards
- 72 seconds, or 30 minutes, and both are trials
- The 12% was non-inferior, and P was 0.41
- Refactoring fell from 25% to 3.8%
- Phase I improved. Phase II did not.
- Adding the doctor to the model changed nothing
- The pilot failed and the staff deployed it anyway
- The pilot ran on data the production system will never see
- The framework that exists produced 1.6%
- One trial, a waitlist control, and a letter
- The lock-in is the prompts, not the API
- 621,000 robots installed, virtually no humanoids
- 220 million miles, inside a boundary Waymo drew
- Robots weed ten million acres. They still cannot pick.
- Every chatbot query on earth is 2% of AI's power
- Ten domains, and the specification moved in every one
- Same GPUs, same month, opposite depreciation
- The best-documented humanoid deployment is seven units
- A million deliveries, and the drone never lands
- Every AI chip passes through one company's machines
- A million trajectories, 85% from four robots
- Same company, same year, revenue figures 63% apart
- 0.03% of construction spend, so the work moved indoors
- 32 million vacuums, and 38% on household tasks
- AI in science: 380,000 predicted, 736 actually made
- Testing a system that answers differently every time
- How to check an AI claim before you believe it
- Fixing the seed does not make it reproducible
- AI in software: 19% slower, and they felt 20% faster
- AI incidents: two registers, 1,460 and 14,530
- LLM-as-a-judge: when a model can grade another model
- What a model card should say and usually does not
- AI support: 90% deflected, 40% actually resolved
- AI in translation: the field that retired its own metric
- How to read an AI paper
- AI in journalism: 45% of news answers had a flaw
- Superintelligence: the empirical record is zero
- Why machine learning does not do error bars
- Inference prices fell 9x a year. Also 900x.
- The water bottle was per 10 to 50 responses
- The 13% traveled. The authors' caveat did not.
- Three years in: no disruption, and one 20% hole
- AI in education: students feel twice the gain they get
- Why AI models get worse: forgetting, collapse, and drift
- A $3.5bn guarantee book and a $250bn commitment
- 232 studies, and 1.3% recorded skin type
- The binding constraint is a transformer, not a chip
- 1 to 2% of the chips, about 30% of the tokens
- The grader rewards what the detector flags
- Slop scores premium 70% of the time
- The good numbers all came from obligations
- Worse than chance means bias, not noise
- Zero-click is 60%, or 22.4%, from one provider
- Generation takes seconds. Debunking takes hours.
- Collapse needs you to throw the old data away
- Twenty-four hours of silence is a billable resolution
- One bug revoked every photo those cameras signed
- Fraud doubles in 18 months. Retraction takes 40.
- Clinicians override 49% to 96% of alerts
- 99.98% is a tracking rate, and the field is animation
- Eight subjects, one ratio, and it is not quality
- The same PDF says 83% and nobody quotes it
- 4.7% at one attempt, 63% at a hundred
- 61% once, 25% eight times running
- Lesson planning is 60 to 99% of the usage
- Good evidence, and eight different failures
- The flags land on lower prior attainment
- Plus 48% with the tool, minus 17% without it
- Why AI benchmarks mislead: contamination, gaming, saturation
- Your agent returned 200 OK and did the wrong thing
- 67% against 39%, and the gap is training
- How do we measure AI progress? The benchmark problem
- Do AI detectors work? Accuracy, bias, and false positives
Building with AI18
- Beyond vector search: how RAG actually works in 2026
- Your RAG system isn't hallucinating. It never found the answer.
- How AI models are trained: from raw text to a system that helps
- RAG vs fine-tuning: the decision, honestly
- What is fine-tuning? How to specialize an AI model
- When not to use an agent
- How LLM inference works: why it's bound by memory, not compute
- Do you actually need a vector database?
- How quantization shrinks AI models without breaking them
- What an agent costs, and where the money actually goes
- Speculative decoding: faster LLM generation, same output
- Knowledge distillation: how small models learn from big ones
- What is synthetic data? Training AI on AI-made data
- What is a mixture of experts (MoE)? Bigger, cheaper AI
- How much energy does AI use? Training vs inference
- Why does AI need GPUs? Parallelism and the memory wall
- Run an LLM locally: how much VRAM do you need?
- Open weights vs open source AI: what's actually released
Safety & governance30
- 2.6 million robotic surgeries, none of them autonomous
- Nine cases, and none was fixed by a better model
- Robodebt: losing quietly to avoid losing publicly
- Horizon: the law presumed the computer was right
- The robots removed the walking. It was also the rest.
- The sepsis model caught 7% of what clinicians missed
- Williams v Detroit: the match was not the failure
- AI in finance: the regulator looked and stepped back
- AI in government: 126 use cases, 65 not made public
- Data sovereignty: the question that decides your AI architecture
- The Dutch benefits scandal: the rule, not the model
- Amazon's hiring AI: the case with no primary source
- AI in medicine: 1,524 devices, 1.6% with trial data
- Zillow Offers: $304 million, in the audited filing
- COMPAS: both sides of the dispute were correct
- Moffatt v Air Canada: what the $650 ruling settled
- The Tempe crash: it saw her for 5.6 seconds
- AI alignment and safety, without the hype or dismissal
- AI in hiring: 18 bias audits from 391 employers
- AI in law: 1,313 filings sanctioned in 106 countries
- 61% flagged for non-native writers, 3% for native
- Agent permissions: the question nobody asks until afterwards
- What is prompt injection, and why is it unsolved?
- Mechanistic interpretability: opening the AI black box
- AI bias and fairness: why 'fair' has no single answer
- The secret language that never was, and the escape that did
- What is AI sycophancy? Why AI tells you what you want
- What is AI jailbreaking? Why safety can be talked around
- How to secure an LLM application: risks and defenses
- AI and copyright: the three questions people confuse