Guide
Read an AI launch without inheriting its hype
Turn a model announcement into a small decision record: what shipped, under which conditions, and what you still need to test.
Lab roadmapsPeople & claims
Source-based4 min ↗
CURIOUS MINDS, CLEARER ANSWERS
Your AI reference shelf.
Find an explanation, a practical guide, or a better question to ask. Every resource has a clear scope and original sources.
Explainer
Open weights, open code, open enough for what?
Inspect the actual artifacts and permissions behind an openness claim instead of treating open as a single product feature.
Open models & licensesData & training
Source-based4 min ↗
Guide
Your local AI assistant may still use the internet
Draw the model, retrieval, tool and logging boundaries before assuming that a local application is fully offline.
Local & edge AIData flow
Source-based4 min ↗
Explainer
Memory, context and training are different levers
Locate where information is stored and how it reaches the model before choosing a memory feature, retrieval system or training job.
Context & memoryTraining & adaptation
Source-based4 min ↗
Research note
Test the plan, not the confidence of the answer
Turn an impressive reasoning demonstration into a task with explicit states, constraints, perturbations and a checkable result.
Reasoning & planningBenchmarks & methods
Source-based4 min ↗
Guide
A real citation can still support the wrong claim
Audit a short AI-generated answer at the claim, passage and source-family levels. A working link is only the beginning.
Sources & evidenceResearch workflow
Source-based4 min ↗
Research note
Read the update before repeating the productivity number
Use METR’s early-2025 study and its 2026 follow-up to understand populations, selection effects and changing evidence.
Benchmarks & methodsEngineering practice
Source-based4 min ↗
Guide
An agent loop needs an exit
Choose between a fixed workflow and an agent by specifying feedback, budgets, terminal states and the point of human handoff.
State & orchestrationTool selection
Source-based4 min ↗
Guide
Approval is a gate. Recovery is a separate design.
Plan what happens before, during and after human approval, including rejection, changed payloads, restarts and duplicate actions.
Handoff & recoverySecurity & permissions
Source-based4 min ↗
Guide
Review the complete change, not the generated code
A compact review protocol for AI-assisted repository work, covering requirements, tests, configuration, dependency changes and rollback.
Repository contextDebugging & tests
Source-based4 min ↗
Explainer
A natural voice can still fail the task
Separate speech recognition, understanding, tool actions and audible delivery when designing a voice-workflow evaluation.
Speech & interactionTask evaluation
Source-based4 min ↗
Research note
From a useful demo to evidence of adoption
Keep technical feasibility, observed use, task outcomes and commercial demand on separate evidence lines.
Knowledge workOpportunity signals
Source-based4 min ↗
Checklist
Make a model-family manifest before choosing a checkpoint
A worksheet for separating model variants, release artifacts and access routes when a single launch name hides several different choices.
Model variantsRelease artifacts
Source-based4 min ↗
Workflow
Turn an expert’s model anecdote into a test pack
Preserve the question behind an early-access impression, then replace the implied ranking with a small, repeatable task record.
Model evaluationExpert claims
Source-based4 min ↗
Repository
Read a released coding agent as a specification
Use the code-simplifier source to identify the behavior a reusable agent promises, then design a small adoption review around it.
Open repositoriesCode simplification
Source-based4 min ↗
Explainer
Why KV caching turns context into a capacity question
Understand the memory side of token generation, then ask better questions about prompt length, concurrency and serving limits.
KV cacheServing capacity
Source-based4 min ↗
Repository
Read Local Deep Researcher before running it
A repository walkthrough for mapping model, search and report stages into an explicit configuration and data-flow record.
Local modelsConfiguration
Source-based4 min ↗
Checklist
Test API portability with a behavior contract
A reusable worksheet for comparing the responses your application depends on when two servers expose similarly named endpoints.
API portabilityMigration
Source-based4 min ↗
Prompt
Use a hypothetical document as a retrieval scratchpad
A prompt and comparison method for exploring HyDE while keeping generated search material separate from evidence used in an answer.
RAGHyDE
Source-based4 min ↗
Paper
Change the labels, keep the planning problem
A reading note on representation sensitivity, with a small exercise that separates familiar wording from valid state transitions.
PlanningRepresentation
Source-based4 min ↗
Paper
Read a vision model through its transfer task
Use the PaliGemma 2 paper to plan a vision-language experiment with explicit image conditions, labels and adaptation boundaries.
Vision-language modelsFine-tuning
Source-based4 min ↗
Research note
Read agent reliability beyond a single successful run
A short route into τ-bench’s final-state and repeated-trial ideas, with a practical way to record a tiny task suite.
Agent reliabilityRepeated trials
Source-based4 min ↗
Checklist
Evaluate the monitor as well as the agent
Use SHADE-Arena as a research example for asking what a transcript monitor can see, miss and falsely flag.
MonitoringEvaluation design
Source-based4 min ↗
Workflow
Give preference and correctness separate scorecards
A compact review workflow for comparing AI answers without letting warmth, verbosity or visual polish stand in for task success.
Human preferenceReview rubrics
Source-based4 min ↗
Explainer
A skill and an MCP server solve different handoff problems
Map instructions, reference material and callable operations before deciding what belongs in a reusable agent package.
MCPAgent Skills
Source-based4 min ↗
Workflow
Write a delegation brief that can be merged
An original worksheet for dividing a research question into independent assignments and defining what a usable return must contain.
DelegationResearch workflow
Source-based4 min ↗
Checklist
Write an agent bug report that names the failing layer
A recovery checklist for separating the model, harness, tools and application state when an agent behaves differently than expected.
Failure analysisRecovery
Source-based4 min ↗
Prompt
Give a coding assistant a behavior-preserving refactor brief
An original prompt for stating invariants, allowed files and review evidence before an assistant reorganizes an existing codebase.
RefactoringRepository context
Source-based4 min ↗
Workflow
Make deployment dependencies visible before the final build
Use a first-person app-building account to create a handoff ledger for configuration, identities and service boundaries.
DeploymentConfiguration
Source-based4 min ↗
Guide
A faster cache still needs a freshness check
Read a performance anecdote through cache mechanism, update behavior and rollback evidence before repeating a speed claim.
CachingMaintenance
Source-based4 min ↗
Research note
Inspect a video capability across time
An original evaluation note for checking identity, object state and requested changes rather than judging a generated clip from one attractive frame.
Video understandingTemporal control
Source-based4 min ↗
Checklist
Check speech controls with a listening rubric
Separate a text-to-speech system’s advertised controls from the pronunciation, meaning and timing a listener actually receives.
SpeechEvaluation design
Source-based4 min ↗
Reading list
Ask what a world model actually outputs
A reading guide for distinguishing rendered observations, structured state and predictions before comparing systems under the same label.
World modelsOutput contracts
Source-based4 min ↗
Workflow
Build a support-routing sandbox before connecting a channel
Turn a fast customer-support demo into a small table of fictional tickets, expected destinations and escalation conditions.
Support routingSandbox evaluation
Source-based4 min ↗
Checklist
Check what you can do after the tutor closes
An original self-study worksheet for separating assisted progress from an independent explanation and a new problem.
LearningIndependent practice
Source-based4 min ↗
Research note
Check who decides that a robot succeeded
A technical review note for separating task execution, control architecture and the labels used to judge a recorded attempt.
RoboticsSuccess labels
Source-based4 min ↗
No matching resources.
Try a broader phrase or remove a filter. Full-text search is also available in the header.