Mira Murati Revolutionizes Conversational AI with Thinking Machines

Mira Murati, a Visionary in Conversational AI

Fourteen months after her departure from OpenAI, Mira Murati makes a notable return with the launch of TML-Interaction-Small, the first major project from her new company, Thinking Machines. Contrary to what one might expect, this is not just a simple emulator of GPT. Murati has designed an artificial intelligence capable of conversing fluently, listening and reacting in real-time, closely mimicking an authentic human conversation.

Today, interactions with voice assistants are often comparable to an email exchange, where one must wait for the end of each response before being able to speak again. TML-Interaction-Small aims to transform this dynamic. With its ability to listen and analyze continuously while speaking, the AI can adjust its responses live, evoking the atmosphere of a lively family dinner where conversations naturally overlap.

The Technology Behind TML-Interaction-Small

The key term used by Thinking Machines to describe this innovation is "full-duplex." This concept allows two interlocutors to speak simultaneously, much like in a traditional phone call. To achieve this, the company has developed two distinct models that work in tandem.

The first model is dedicated to managing real-time interactions, handling voice, interruptions, and the pace of the discussion. The second, more complex model operates in the background to process tasks requiring advanced reasoning, such as web searches or the use of external tools, before relaying the results to the main conversation.

The system is designed to break exchanges into micro-turns of 200 milliseconds, allowing the AI to continue listening even while responding. This avoids the typical interruptions of traditional voice assistants that can lose track of the conversation, offering a more natural and continuous experience.

Performance and Comparison with Competitors

TML-Interaction-Small stands out in performance tests, particularly on the FD-bench v1.5, a benchmark specialized in evaluating the fluidity of voice interactions. The model achieves an impressive score of 77.8 points, far surpassing GPT-realtime-2.0, which only reaches 46.8 points.

In terms of latency, TML-Interaction-Small shows a response time of 0.40 seconds, compared to 0.57 seconds for Google Gemini 3.1 Flash Live and 1.18 seconds for GPT-realtime-2.0 minimal. These figures bring the AI closer to the fluidity of human conversations, which typically range between 200 and 250 milliseconds between two interventions.

Beyond the benchmarks, the architecture of TML-Interaction-Small overcomes a well-known limitation of current language models: time management. Unlike other AIs, this model can understand and process complex temporal instructions, such as "Remind me to check the temperature every four minutes," which is crucial for industrial, medical, or scientific applications. According to VentureBeat, this ability to manage time accurately is a major asset that distinguishes TML-Interaction-Small from its competitors.