Turn-Based Interface Bottleneck

Sources#

Summary#

Thinking Machines Lab's framing of why current AI interfaces limit collaboration: the turn-based interface is a bandwidth bottleneck between human and model. It is the problem Interaction Models are built to dissolve.

The two-claim argument#

AI labs over-optimize for autonomy. Labs treat autonomous capability as the model's most important property; as a result, today's models and interfaces "aren't optimized for humans to remain in the loop." But in most real work users can't fully specify requirements upfront and walk away — good results come from a collaborative loop of clarification and feedback.
Humans get pushed out by the interface, not the work. "Humans increasingly get pushed out not because the work doesn't need them, but because the interface has no room for them." The fix is to let people collaborate with AI the way they collaborate with other people: messaging, talking, listening, seeing, showing, interjecting — and the model doing the same.

The mechanism: a single thread#

Today's models "experience reality in a single thread":

Until the user finishes typing/speaking, the model waits with no perception of what the user is doing or how.
Until the model finishes generating, its perception is frozen — no new information arrives until it finishes or is interrupted.

This narrow channel limits how much of a person's knowledge, intent, and judgement can reach the model, and how much of the model's work is legible to the human. Analogy: "trying to resolve a crucial disagreement over email rather than in person."

Why harnesses don't fix it#

Existing real-time systems bolt interactivity on with a harness — VAD (voice-activity detection), turn-boundary prediction, dialog state machines — components "meaningfully less intelligent than the model itself." That harness precludes whole interaction modes:

proactive interjection ("interrupt when I say something wrong")
reaction to visual cues ("tell me when I've written a bug in my code")
speak-while-listening ("translate Spanish→English live")
speak-while-watching ("live-commentate this sports game")

The Bitter Lesson says these hand-crafted systems get outpaced by general capability growth → the resolution is to make interactivity model-native (see Time-Aligned Micro-Turns).

The harness dissolves in production: GPT-Live (July 2026)#

Two months after TML's argument, OpenAI shipped its conclusion: GPT-Live "removes the turn detector from the audio path" of production ChatGPT Voice. OpenAI's own retrospective states the bottleneck exactly as this page does — the turn detector "faced an unenviable task: guess too soon, and the user gets cut off; guess too late, and the response feels sluggish. Only after the detector made its decision could the much larger LLM get to work." Even the intermediate speech-to-speech generation, which absorbed transcription into the model, kept the gate: "the model handled more of the interaction, but the interaction remained turn-based." The less-intelligent harness component named by this page's thesis is now gone from a deployed system — Harness Shrinkage as Models Improve landing at the interaction layer, from a second lab, as production engineering rather than research argument.

The nuance the production system adds: turns don't disappear — they move. ChatGPT's conversation UI, analytics, and safety systems still consume discrete user/assistant messages, so GPT-Live's application server derives turns from the continuous stream after the fact — speculative and authoritative views, a freshness-for-certainty trade (details in Live-Path Minimalism). The turn-based structure was a bottleneck as a gate on inference; as a data model for the surrounding product it survives, maintained by inference over the stream rather than imposed on it.

Connections#

Interaction Models — the proposed resolution
The Bitter Lesson — why the harness-based status quo loses
Time-Aligned Micro-Turns — the architectural move that removes turn boundaries
Full-Duplex Interaction — the interaction modes the bottleneck currently blocks
Harness Shrinkage as Models Improve — the general version of "the less-intelligent harness should dissolve into the model"
AI Employee Framing / Human-AI Accountability Redesign — the org-side mirror: both warn against treating autonomy as the goal and pushing the human to the margin; this page is the interface-side version of the same critique
Design Concept Grilling — argues the value is in collaborative iteration; this page argues the interface is what blocks it
Context Window Smart Zone — orthogonal limitation that also makes "fully autonomous, walk away" brittle
Configurable Human Participation — HAS-Framework's typed human/agent-initiated channels are what an interface with room for the human would route; the measured variation in agent-initiated clarification quality (and in whether the agent asks at all) is this bottleneck scored as an outcome
GPT-Live — the production system that removed the turn detector
Live-Path Minimalism — where turns re-enter as a derived application-layer view after leaving the audio path

Sources#

Interaction Models: A Scalable Approach to Human-AI Collaboration
How we built a realtime system for responsive voice AI in six months — OpenAI, 2026-07-29 (case-study): the turn detector removed in production; turns re-derived at the application boundary