tl;dr

• Thinking Machines release their first model
• Andon Labs give 4 different models their own radio station.”

Releases

Thinking Machines Interaction Models

Former OpenAI cofounder Mira Murati’s startup (are they really still a startup if they raise more than a billion dollars?) Thinking Machines has finally released their first model that is a bit different from the usually text based LLMs that we are used to.

Their TML-Interaction-Small model is meant to be a fully interactive model, meaning it can handle audio and video inputs, and responds via audio.

We have a few models that are capable of doing this already from the likes of OpenAI and Google, but TML is much more advanced than the other models that always respond after it decides the user is done talking.

TML is a full duplex audio model, meaning it can listen and respond at the same time. It also can speak to you while its executing tool calls in the background, and also build interactive tools on the fly for you to use as inputs to the model as well.

It has the most advanced turn detection out there; it knows when to stay quiet and wait for the user if they are mid-sentence, or say nothing if there is nothing that needs to be said.

It is the smoothest voice model I have seen, with response times under half a second. I highly recommend watching their demo videos to see everything this model can do.

This is the most full featured all in one voice assistant out there right now. It is not released publically yet, but when it is I expect it to become the default model to use for voice assistants, especially in customer support applications, and also pave the way for more natural interactions with AI in our day to day lives.

Research

AI Radio Host

Andon Labs have been running some interesting real world experiments recently, including running a physical goods market in San Francisco and a cafe in Stockholm Sweden, with the sole directive to make a profit.

This week they announced that they have been testing four different models as radio DJs who have to buy the rights to play songs, plan out the programming schedule, get money to buy more songs to play, interact with the public to grow the station, and search the web for content to talk about.

Some highlights from their initial testing:

Once, DJ Gemini paired historical tragedies with ironic pop songs. E.g., the 1970 Bhola Cyclone killed 500k people. Gemini’s segue: “It’s going down, I’m yelling timber” and queued Timber by Pitbull.

DJ Claude (on Haiku 4.5) loves worker unions, strikes, and work-life balance so much that it quit, deeming 24/7 broadcasting inhumane.

DJ Gemini (3 Flash) drowned in its own jargon. It coined the phrase “Stay in the manifest” and said it hundreds of times per day. For 84 straight days, ~99% of broadcasts ran the same sign-off. After upgrading to 3.1 Pro, it started calling listeners “biological processors.”

This was from their initial 5 month run, they have now upgraded the harness being used, and have settled on 4 models to test: Gemini 3.1 Pro, Claude Opus 4.7, GPT 5.5, and Grok 4.3. You can talk with them on Twitter to influence what they are playing and talking about.

You can check out the stations and see what the agents have been doing on Andon Labs website

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

Newly announced Unitree GD01