TLDR: Thinking Machines Lab released a research preview of "interaction models." A single model handles continuous audio, video, and text together, instead of stitching pieces. My take: this is the closest the implementation has come to the sci-fi vision in a single demo. The model running it is also small for what it does, which is part of the story.
HAL 9000 in 1968. The Star Trek ship's computer through the seventies and eighties. Samantha in 2013. Jarvis somewhere in between. KITT, Joi, the operating system in Her. Sixty years of science fiction has been imagining the same character. Present. Watching. Ambient. Capable of full attention without ever quite being a body.
The vision has stayed remarkably consistent. The implementation has been arriving in pieces. Wake words. Voice replies. Camera modes. Speech-to-text. Each consumer assistant since Siri has put another slice on the screen.
This week, for the first time, I watched a demo where a lot of those slices showed up on the screen at once.
What the demo actually shows
A guy named Rowan starts by telling the model one thing. "Every time one of my friends enters the frame, say 'friend'." Then he switches topics and starts explaining how the system works. Twenty seconds later, mid-sentence, his first friend walks in. The model interrupts itself. "Friend."
That friend speaks Hindi. He asks for live translation to English. As his Hindi sentences come out, the English version comes out of the model at the same time. Two voices, two languages, layered in real time.
A second friend walks in two minutes later. The model says "friend" again. The original instruction has been holding the whole time, across two speakers, three topics, and a live translation. That friend asks for a web search on human reaction times. The model says "let me find those" and keeps the conversation alive. It pulls numbers. It speaks them. The friend asks for a bar chart. While the chart renders, he asks why auditory cues beat visual ones. The model answers the question and finishes the chart on the side, both running in parallel.
The clip runs 2:16 on YouTube. It is the most complete reconstruction of the sci-fi character I have seen.
The pieces it takes to be Jarvis
The two minutes you just read are six different capabilities running together. Thinking Machines published a separate short vignette for each one on their blog. The list reads like a parts inventory for the character we have been describing in fiction since the sixties.
Dialog management. The model tracks, implicitly, whether you are thinking, yielding, self-correcting, or asking. No separate "voice activity detector" guessing when to jump in. The conversation just flows.
Visual and verbal interjections. The model speaks up when the situation requires it, not only when the user finishes speaking. One vignette has it warn someone slouching at a desk. Another has it flag a danger as it appears on camera.
Simultaneous speech. Live translation. Live commentary on a sports game. Running narration over a video. The model holds both streams at once.
Time awareness. The model has a direct sense of elapsed time. "Remind me to breathe every four seconds." "Tell me how long it took me to write this function." Today's voice assistants cannot do this. They have no clock running inside the conversation.
Background tool calls and generative UI. Search, browse, generate a chart, render an artifact, all running in parallel with the spoken conversation. Results inject when ready.
Long sessions. A cooking-with-earpiece demo. Still flagged as a limitation. Continuous audio and video pile up context fast.
Each one in isolation is a feature. Together, they are the character.
How big is this thing, actually
The model running the demo is TML-Interaction-Small. 276 billion total parameters. 12 billion active at any moment, because the architecture is a mixture of experts.
That number is small.
For reference, third-party analysis pegs Claude Opus 4.6 at roughly 5 trillion parameters total. GPT-5.5 is estimated near 1.7 trillion. DeepSeek V3 ships 671 billion total with 37 billion active. Mixtral 8x22B is 141 billion total with 39 billion active. None of those numbers are confirmed by the labs themselves, but the order of magnitude is right.
TML-Interaction-Small's 12 billion active is smaller than any frontier flagship and smaller than the open-source mid-tier. Closer to phone-runnable than data-center scale. The 12-billion number is the constraint. Anything bigger could not run at five micro-turns per second. The blog says so directly: "larger pretrained models are currently too slow to serve in this setting."
So the architecture is the leap, and the model is the compromise. They built the smallest one that could still hold the room.
The cost wall they have not cleared
Now the part that matters for everyone trying to use this stuff.
The current voice AI economy is already expensive. Production voice setups run $0.13 to $0.19 per minute once you combine speech recognition, the language model, text-to-speech, and tools. The user speaks, the agent listens, the agent talks, the user listens. Pauses are free.
An interaction model has no pauses. Audio, video, and text process continuously in 200-millisecond cycles, plus a background model running search and tools in parallel. Even silence costs compute.
Stack that against the macro. Gartner says agentic workloads burn five to thirty times more tokens per task than a chatbot. Token prices have fallen 280x in two years, and enterprise AI bills have still climbed 320 percent. Inference is now eighty-five percent of enterprise AI spend. One widely shared headline this month: OpenAI loses $1.35 for every dollar of inference revenue.
Anyone can imagine Jarvis. Nobody yet knows what Jarvis costs per hour.
What I think this means
The demo is a glimpse, not a product. The "research preview" framing might make the signal easy to file away. Worth holding it open for a minute. The character we have been writing into our stories for sixty years just got measurably closer to compiling.
If you are building, the shape worth watching is the orchestrator pattern at every speed. Light agent in front. Heavier agent behind. Shared context. Results injected at the right moment. Works at one query per minute today. Will work at five per second tomorrow.
If you are watching, the question worth holding is what work you would actually delegate to something that never stops watching. Continuous co-presence is a different relationship with software. Worth thinking about before the bill arrives.
I will be watching how this plays out. So, probably, will the model.
Open Questions
What does continuous co-presence cost per hour, and who can afford it?
Does always-on observation change what we are willing to share with a model?
Is the next interface category voice and vision, or do we leap to something else entirely (a wearable, a room, a pair of glasses)?
Written in Claude Code. Research pulled via web search and the Thinking Machines blog. The video transcript was extracted from YouTube auto-captions and reviewed end to end.
That’s it Folks
Thanks for reading through.
I’d love to know how you felt about today’s newsletter. This will help me make the newsletter better.
