Explainer
What does an AI agent actually do?
A practical guide to planning, tools, memory, permissions, and human review.
Evidence with its limits visible
Research
Structured findings, explainers, methods, and evidence designed to remain useful.
Latest from Research
20 published worksExplainer
What is a context window—and why is it not memory?
The text a model can see is different from what a product remembers about a person or organization.
Explainer
How does a GPU turn electricity into tokens?
A plain-language route through matrix operations, memory bandwidth, networking, and inference.
Explainer
What makes an AI model open?
Weights, code, data, licenses, and documentation can be open in different combinations.
Explainer
How are AI systems evaluated before release?
Benchmarks are only one layer in a release process that also needs scenarios, experts, and monitoring.
Explainer
Why do AI models hallucinate?
Fluent generation is not the same operation as checking a claim against the world.
Explainer
What is inference—and why does it cost so much?
Training creates a model; inference pays the bill each time someone uses it.
Explainer
How does retrieval-augmented generation work?
RAG gives a model selected documents at answer time, but it does not guarantee that the answer is correct.
Explainer
What changes when software can use tools?
The moment a model can send, buy, edit, or deploy, interface design becomes authorization design.
Explainer
What is a world model?
The term covers systems that learn useful representations of how environments behave over time.
Research
Where people use Claude at work
A reading of Anthropic’s Economic Index, occupation by occupation.
Research
The public evidence for longer-running agents
What task duration reveals about reliability, supervision, and the changing unit of AI work.
Research
A field guide to frontier model documentation
We compared the public artifacts used to explain capability, safety, and release decisions.
Research
The geography of the current AI company
Company headquarters and event calendars show a dense Bay Area core with important European and global nodes.
Research
The agent toolchain, mapped
Editors, runtimes, model providers, observability systems, databases, and deployment platforms now form a recognizable stack.
Research
How AI companies describe safety
A taxonomy of public commitments across labs, infrastructure providers, and application companies.
Research
The hidden work in a generated video
A workflow analysis of prompting, selection, editing, sound, rights, and delivery.
Research
What startup homepages promise in the agent era
A language study of autonomy, speed, teammates, and the disappearance of implementation.
Research
The compute market between NVIDIA and the model lab
Specialized clouds are creating a distinct layer for capacity, scheduling, and systems integration.
Research
An index of AI conference disclosure rules
NeurIPS and ICML are turning tool use, reproducibility, and research ethics into explicit publication operations.