Interaction Is Intelligence
The Battle Mira Murati Won't Fight

If you've used ChatGPT's voice mode, you've likely felt that subtle discomfort—you finish a sentence, wait about a second, and then the other side "reacts." This latency isn't "slow," but it just doesn't feel human.

The reason is that nearly all "real-time voice" AI today has a turn-based underlying structure: either listening or speaking, never both at once. The model listens to an entire turn, then speaks an entire turn, relying on an external module called VAD (Voice Activity Detection) to guess—"Has the user finished speaking?" This judgment is far cruder than the model's own intelligence.

Thinking Machines' core argument is sharp: All of today's major labs' "real-time models" are just turn-based models wrapped in hardcoded scaffolding and then called "real-time." This violates Rich Sutton's "Bitter Lesson"—any handcrafted scaffolding will ultimately be replaced by end-to-end learned large models. With this conviction, TML-Interaction-Small didn't build VAD, ASR, TTS, and dialogue manager as separate modules to stitch together; instead, it trained a model from scratch that natively consumes continuous audio-video streams. Thinking Machines Lab Official Blog

The model is named TML-Interaction-Small (the "Small" implies larger ones to come), with a Mixture-of-Experts architecture totaling 276B parameters, activating 12B per token, natively processing audio, video, and text streams. All components (audio frontend, video frontend, Transformer backbone, audio decoding flow head) are jointly trained from scratch. It is currently a research preview with limited internal testing available.

The most underrated yet critical design in the paper is the dual-model collaboration architecture: the Interaction Model stays online continuously, handling conversational rhythm, immediate responses, and visual perception; the Background Model runs asynchronously, tackling deep reasoning, web browsing, and complex tool calls. Both share full context—when the Interaction Model judges the current task as "manageable," it responds itself; when it judges it as "too complex," it packages the context and sends it to the Background Model, continuing to chat with the user, and naturally weaving in the answer once the Background Model delivers results. This "fast-slow layered" paradigm is similar to the reasoning + agent split Anthropic has done with Claude, but Thinking Machines has made it an explicit dual-model real-time collaboration.

None of today's OpenAI / Google real-time APIs can handle these entirely new dimensions—not because their models aren't strong enough, but because these tasks fundamentally cannot exist under a "turn-based + VAD scaffolding" architecture.

The first key concept to understand this release is time-aligned micro-turn. Traditional models flatten input and output into a single token sequence with no concept of time—only token order. User pauses, silences, and interruptions, these "meanings carried by time," are entirely imperceptible. The Interaction Model sees a world where it processes a slice of input and generates a slice of output every 200 milliseconds: if the user is speaking in this 200ms slice, the model outputs "silence"; when it should interject, it outputs audio tokens. Silence is explicitly modeled as a type of model output—this is the key, enabling proactive interjection, simultaneous speech (live translation), and visually proactive speaking to emerge naturally.

Most "full-modal" models follow an encoder + LLM route: audio passes through a Whisper-like encoder, images through a ViT-like encoder, becoming tokens fed to the LLM, and output goes through TTS—each component is a pre-trained "part," but the parts lack gradient connectivity. Thinking Machines chose the opposite path: audio is represented with dMel, images are sliced into 40×40 patches processed by hMLP, audio output is decoded via flow head, and all frontends and the Transformer backbone are trained together from scratch. The cost is much higher training expense; the benefit is that the model truly learns to "listen" and "see."

The 200ms hard latency constraint means the inference stack requires extensive low-level optimization: Streaming Sessions (open-sourced into SGLang) maintain persistent session sequences in GPU memory, only appending new 200ms chunks each time; MoE under small-batch scenarios uses gather+gemv instead of grouped gemm; deterministic kernels (continuing the earlier Defeating Nondeterminism in LLM Inference work) achieve trainer-sampler alignment with end-to-end overhead <5%; NVIDIA Blackwell's NVLS is used for all-reduce communication optimization. These engineering details explain why no one previously managed to make a 276B model full-duplex real-time.

A score of 77.8 vs. 54.3 for second place on FD-bench V1.5 is not a 10% improvement—it's nearly a 50% jump. In an era where AI benchmarks are contested over 1–2 point margins, this gap indicates evaluation of a dimension systematically overlooked by existing architectures. However, engineer Sean Goedecke raised a criticism worth taking seriously: scores vary dramatically depending on whether the Background Model is included—on BigBench Audio, the Interaction Model alone scores 75.7%, but with the Background Model it jumps to 96.5%; FD-bench V3's Pass@1 of 68% was also measured with the Background Model. This isn't cheating, but it does blur the question of "how intelligent the Interaction Model itself actually is." Sean Goedecke

Thinking Machines itself acknowledges in its blog: full-duplex conversation is not their invention—Moshi (Kyutai Labs, 2024) and PersonaPlex (NVIDIA) have already validated the micro-turn and interrupted streams technology. Goedecke ran the numbers: TML-Interaction-Small is roughly 2x Moshi's activated parameters and 40x its total parameters—the real technological breakthrough of Interaction Models lies in scale + video + industrial-grade deployment, not in the invention of micro-turn itself.

Murati secured $2 billion in seed funding when she left OpenAI, and is now in talks for a new $5 billion round, with valuation reportedly approaching $50 billion. Bloomberg The market originally expected another GPT-5-level frontier model; Interaction Models provided a different answer: don't fight this battle. On the frontier model track, OpenAI has the most compute, Anthropic has the sharpest training methodology, and Google has the most complete data stack—a lab founded in 2025 fighting frontier models head-on is almost guaranteed to lose, especially when its greatest advantage is talent (John Schulman, Barret Zoph, and over 30 former OpenAI researchers) rather than compute. So Murati found a dimension where others aren't pushing hard: treating "how AI collaborates with humans" as a first-order problem.

The blog quotes Anthropic's own model card: "When our models are used in a 'hands-on-keyboard' synchronous interaction mode, their value is significantly discounted... placed into autonomous long-task agents, they actually better unlock coding capabilities"; and Hayek's 1945 paper The Use of Knowledge in Society—distributed knowledge about particular times and places cannot be captured by any central dispatch system. Reading these two statements together: Thinking Machines is saying that the current direction AI labs are pursuing "autonomy" is actively abandoning the unique knowledge humans possess.

The research preview demos all look smooth, but full-duplex systems have a host of devilish details in real scenarios: the 200ms hard constraint means extreme sensitivity to network jitter; in noisy environments like restaurants or cars, whether the end-to-end model outperforms VAD is not directly addressed in the blog; Thinking Machines itself admits that continuous audio-video streams cause context to accumulate rapidly, and long sessions still require "careful context management."

Goedecke's question also bears repeating: when the Background Model slowly produces results, will the Interaction Model exhibit "pretend I didn't say that" self-corrections? Having a fast model delegate tasks to a slow model at the right time is a classic problem in AI system design, and neither OpenAI nor Anthropic has fully solved it yet. Also, "the model proactively points out a bug while you're coding" sounds cool, but the same capability applied elsewhere could be unsettling—where the boundary of proactive speaking lies remains an open question.

Reshuffling the voice AI startup landscape—many "AI customer service / AI sales" companies have tech stacks assembled from STT + LLM + TTS + VAD; if Thinking Machines opens its API, voice infra companies like Vapi, Retell, and Bland face sudden pressure. AI hardware design philosophy—after the Humane AI Pin failed and Rabbit R1 faltered, the dual-model architecture (running Interaction on-device, Background in the cloud) provides a technical foundation for next-gen AI glasses/earbuds. Reflection on the "Agent" narrative—2024–2025 was the peak of the agent autonomy narrative; Thinking Machines uses Anthropic's own model card to argue that models actually perform better with human-in-the-loop, and the industry re-embracing "collaborative AI" is not impossible. Updating evaluation science—today's benchmarks are almost entirely turn-based paradigms; FD-bench, TimeSpeak, and CueSpeak are just the beginning.