Machine-Learning Capability Stack
Separate recognition, intent parsing, and task execution before judging AI
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 98%
Analyze an AI product as a stack rather than a single intelligence. First, the system converts raw input—such as audio or pixels—into a usable representation. Second, it interprets that representation and produces a structured query or classification. Third, it must route the result to an execution capability that can actually complete the task. Machine learning has sharply improved recognition and natural-language processing by learning complex statistical rules from labeled examples, but it does not automatically create every downstream workflow. A voice assistant may understand a request perfectly while lacking the hand-built module needed to act on it. Evaluate each layer independently, distinguish learned components from manually authored ones, and treat competence in one bounded task as narrow automation rather than evidence of general intelligence.
Origin
Benedict Evans presented this stack on Masters in Business while explaining why better speech recognition does not automatically produce an assistant capable of answering or executing every possible request.
Core principles
- 01Reliable recognition is not general intelligence
- 02Understanding a request and executing it are separate capabilities
- 03Machine learning can replace hand-written recognition rules with learned patterns
- 04Narrow systems may be transformative without becoming human-like
- 05The execution layer often remains manually engineered
How to run it
- 1
Map the raw input
Identify the waveform, image, text, sensor stream, or other input the system receives. Specify what usable representation must be extracted.
Pro tip Measure recognition accuracy separately from the rest of the product.
Watch out A failure in input recognition can look like a failure of reasoning.
- 2
Map the interpretation
Determine how the representation becomes an intent, classification, or structured query. Define the entities, verbs, and parameters required.
Pro tip Test paraphrases that express the same intent in different language.
Watch out Correct transcription does not guarantee correct interpretation.
- 3
Inspect the execution layer
Identify the service, workflow, database, or manually authored module that receives the structured query. Verify what actions it can actually complete.
Pro tip Inventory supported tasks rather than relying on an open-ended conversational interface.
Watch out Do not assume that understanding a request means the system knows how to fulfill it.
- 4
Separate learned from hand-built
Mark which capabilities emerge from trained models and which depend on rules, integrations, or modules written by people.
Pro tip Use this map to estimate the marginal cost of supporting a new task.
Watch out A product may face severe scaling limits if every new capability requires custom implementation.
- 5
Bound the claim
State the narrow task, data conditions, and measured accuracy within which the system works. Compare the claim with the actual capability stack.
Pro tip Describe the system as an appliance for a task rather than as a synthetic person.
Watch out Avoid general-intelligence language when the system cannot transfer competence beyond its trained function.
In the wild
Speech recognition may accurately transcribe the request, and natural-language processing may identify the meeting and desired change. Execution still requires access to calendars, conflict resolution, participant preferences, permissions, notifications, and fallback behavior—all of which must exist as product logic.
→ The team discovers that the bottleneck is workflow execution rather than transcription accuracy.
A trained model can recognize behavior that differs from labeled normal patterns. The product should be evaluated on that bounded detection task and on the response workflow connected to it, not presented as a generally intelligent security analyst.
→ Capabilities and limitations become testable, reducing exaggerated AI claims.
Common mistakes
Equating transcription with understanding
Turning speech into accurate text is distinct from determining what the user means and what action is required.
Ignoring the missing endpoint
A system may route a request correctly but still lack an implemented service capable of fulfilling it.
Calling narrow automation general AI
A model trained for one classification or recognition task cannot necessarily reason about unrelated tasks.
Is it for you?
Best for
It is best for product managers, investors, and executives evaluating voice assistants, machine-learning products, or automation claims.
Not ideal for
It is not ideal as a detailed guide to training, deploying, or auditing a specific machine-learning model.
From the transcript
“when you talk to a computer, the three things that are going on.”
“you can fill in a dialog box by talking to the computer, and the computer can work out which dialog box you were asking for.…”
“It can't do anything else, and it can't clean. You can't wash your dishes”
From the episode
Benedict Evans Talks About the Tech Bubble
Benedict Evans