Intelligence
The benchmark ends where the workflow begins
A high score is useful evidence about a defined test. An operational evaluation also has to count refusals, handovers, repeated runs and the state left behind.
9 min read
The real test of an AI agent starts after the demo
A fluent plan is the easy part to demonstrate. Authority, retries and recovery determine whether an agent can be trusted with ordinary work.
Read the storyThe latest
Explore the archive
Machines
Robots at work: the case for the ordinary cart
A machine that moves supplies through a building can reveal more about useful automation than a spectacular humanoid demonstration. Its hardest problems may belong to the building and the service around it.
8 min readInfrastructure
AI's power problem has more than one clock
Buying accelerators, connecting a building and running a useful AI service happen on different schedules. Confusing power, energy and timing makes the infrastructure argument harder to solve.
9 min readSignals & systems
Explore the archiveIntelligence
The operating costs behind downloadable AI
An imagined purchasing memo promises control, privacy and zero cost from one model download. Reading its claims separately produces a more realistic deployment decision.
9 min readIntelligence
Your documents can make an AI answer worse
A search result can be relevant, readable and obsolete. Follow a fictional support answer from the document library to the final citation to see where grounding can fail.
9 min readSociety
What the AI coding studies actually measured
The evidence has changed, but not into a universal replacement percentage. Reading the chronology reveals why tool version, task selection and the definition of output matter.
9 min readInfrastructure
The privacy promise inside your phone still needs a boundary
Local inference can keep a request off a server. A useful privacy claim also explains cloud fallback, surrounding app behavior and what happens to the answer afterwards.
9 min readSociety
Three jobs for captions, three costs of error
A transcript's average error rate can conceal the one word that changes an instruction. Three ordinary viewing situations show why accessibility needs more than a language-model score.
10 min readIntelligence
When the exam becomes part of the textbook
A model can know the subject, recognize the question or have encountered the answer. A credible evaluation needs to distinguish those possibilities without treating every high score as evidence of cheating.
9 min readInfrastructure
The dangerous instruction is already in the document
The document is supposed to provide evidence. A connected assistant can mistakenly let it provide authority as well. Follow one fictional research task through that boundary.
9 min readSociety
A signed photograph can still tell a misleading story
Content Credentials can help establish a file's history. Three versions of an imaginary street photograph show why that history and the truth of a caption remain different questions.
9 min readIntelligence
The model teaching its own replacement needs an outside examiner
Generated training examples can improve a model or gradually narrow what it knows. The difference is in the data workflow, the checks and the evidence kept outside the loop.
9 min readInfrastructure
An agent handshake, a token and a question of authority
A common connector can make a tool discoverable and callable. The service still has to establish whose request it is, which records are in scope and what the action is allowed to change.
9 min readMachines
The small model earns its place by knowing when to stop
A narrowly assigned model can be useful without winning every benchmark. Its deployment needs a tested stopping rule and a destination for the work it cannot reliably finish.
9 min read