Unpacks how the word "AI" ends up governing us, and shows practitioners how to reclaim their agency over designing technology.
Agency in Language, Alane Suhr | Compile 26 · Alane Suhr
15 min total·Actually worth watching closely: ~6 min·2 must-watch clips
- 0:00 – 1:45Listen
Language shapes us, and saying what you mean is a search problem
Opens with the frame for the whole talk: the language we use and create shapes us and the world around us, and since we can never fully say exactly what we mean, we create new concepts or reach for vague ones to gesture broadly at what we're after.
The meaning of a concept can be thought of as the things it compels us to do.
Purely spoken, nothing depends on the visuals - fine as a podcast; the point is establishing the framework used throughout.▶ Jump to 0:00Speaker · Alane Suhr - 1:45 – 3:31Listen
Three kinds of meaning: intention, extension, connotation
Lays out the tool used throughout: intentional meaning is the criteria that comprise a concept, extensional meaning is whatever it refers to in this context, and connotation is what the concept does to the context it's placed in (we infer a lecturer has the authority to lecture).
Hold on to these three - the entire dissection of "AI" later is built on this split.
The definitions unfold through spoken examples, with no frames worth grabbing; just follow the three-way distinction.▶ Jump to 1:45Speaker · Alane Suhr - 3:34 – 6:20Listen
"AI" is a concept that promises the impossible
Concepts like learning, consciousness and reasoning have never had a satisfying intentional meaning, and vagueness is a normal tool of language; what's special about "AI" is that it makes us believe there are criteria not yet revealed which, once revealed, everyone would agree on - criteria that cannot exist.
When a concept promises what it fundamentally cannot deliver and we don't notice, we hand our agency to something we can only believe in.
The densest stretch of argument in the talk, but still pure spoken reasoning - don't speed it up, follow it line by line.▶ Jump to 3:34Speaker · Alane Suhr - 6:20 – 8:20Listen
Benchmark chasers and denialists are driven by the same concept
The empiricist who builds criteria for AI out of benchmarks and the denialist who asserts that probabilistic models can never be AI (Gary Marcus) look opposed, but share one pattern: both still believe in the concept and let themselves be driven by it.
People who argue against the AI narrative are often as governed by the concept as the people chasing it.
The contrast is made entirely in speech, with no demo or diagram; just catch the premise the two sides share.▶ Jump to 6:20Speaker · Alane Suhr - 8:20 – 10:34Listen
Extension, reduced: an LLM is a compressed representation of human language
The reference of "AI" today is LLMs: a neural model is a compression of its training data, and the most economical compression is to represent the structures that generated it - syntax, analogy, common sense - so even rudimentary models encode the country-capital analogy consistently. Instruction tuning only adjusts the model ever so slightly, so that what comes back appears like a response.
Model output is hard to distinguish from what a person would say, which exploits our drive to read intentionality into language - and once the relationship gets personal, there are already deaths linked to chatbots.
Both the technical reduction and the risk argument are spoken, with no visual moments flagged; dense, so listen at normal speed.▶ Jump to 8:20Speaker · Alane Suhr - 10:35 – 12:40Listen
The concept justifies itself through stories and prophecies
The concept of AI keeps justifying itself through fiction (the Terminator and HAL 9000 are brought to life by what it connotes), benchmarks and predictions: it really does exist if we believe it does and let ourselves be governed by it.
Taking "AI is imminent" as a prophecy is how you make a fictional concept real with your own hands.
The cultural references are only mentioned in passing, so there's nothing to watch; focus on the self-fulfilling-prophecy mechanism.▶ Jump to 10:35Speaker · Alane Suhr - 12:40 – 14:33Listen
Closing: people are in the room, and people have agency
Pulls the talk together: concepts don't have agency, we do. The people who build technology are in the rooms where these decisions are made, and can refuse the framing of technological inevitability and design technology around human values.
You can refuse the inevitability narrative - agency over how technology gets designed has always been in human hands.
A call-to-action ending where tone and phrasing carry the weight; hearing it through matters more than watching.▶ Jump to 12:40Speaker · Alane Suhr