For the past two years, the AI revolution has mostly worked like a walkie-talkie. You ask something. The AI stops, listens, thinks, and eventually replies. Then it waits again. Whether you are using ChatGPT, Gemini, Claude, or Grok, almost every mainstream tool today follows this “turn-based” model. Humans speak first, machines respond after.
But some of Silicon Valley’s biggest AI players believe that approach is already beginning to feel outdated.
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati has offered a glimpse into what could come next: systems designed not just to answer prompts, but to remain continuously engaged with users, more like a colleague in an ongoing conversation than a chatbot waiting for instructions.
On Monday, Murati’s startup published details about what it calls “interaction models.” These are AI systems capable of listening, seeing and processing information in real time while a conversation is still unfolding.
In practical terms, the ambition is to make AI feel less machine and more human.
The Shift Beyond Turn-Based AI
The idea may sound incremental, but researchers say it changes the nature of human-machine interaction.
Today’s AI tools operate sequentially. A user speaks or types, the model processes the request, and only then responds. Thinking Machines is experimenting with systems that process conversations continuously in tiny 200-millisecond chunks, allowing the AI to react while someone is still speaking.
That opens the door to more natural interactions including interruptions, contextual reactions and conversational cues that humans use instinctively.
AI educator and founder of The Cutting Edge Group Ansh Mehra says the real breakthrough is not merely lower latency. Low latency is the ability of a computing system or network to provide responses with minimal delay.
“Their biggest moat is the new real-time interaction system, which is not just a harness of individual components glued together; it’s a foundational model built for two-way exchanges,” he told NDTV.
Mehra argues that current AI systems fundamentally behave differently from humans during conversation. “Most people in conversation are not really listening, they are simply waiting for their turn to reply. When we speak to a human, they’re not waiting for my sentence to end to form an opinion. The current AI tools we use work on a turn-by-turn basis. They wait for my entire request to end, process it and then think about it.”
The interaction models attempt to mimic that more fluid style of judgment, he says. “This model can think and see as you’re talking, which is incredibly powerful if you think about it. Once you see the demo, you’ll ask yourself why this feature was not mandatory from Day 1.”
AI That Acknowledges You Mid-Conversation
One of the more subtle changes involves something researchers call “backchanneling” which are the small acknowledgements humans constantly make during conversation, such as nodding, saying “mm-hmm,” or briefly reacting while another person is speaking.
The Thinking Machines preview appears to demonstrate early signs of this behaviour. “The model also shows glimpses of backchanneling, which is its ability to make human nods and agreements,” said Mehra. “These incremental changes don’t add to the function but significantly shape the form of AI.”
The broader implication is that AI systems may increasingly optimise not only for accuracy, but for presence thus making conversations feel smoother, more responsive and emotionally intuitive.
The ‘Dual-Brain’ Architecture
To make this possible, the company describes a two-layer architecture. One layer – the interaction model, acts as a fast conversational engine handling real-time exchanges, interruptions and immediate responses. A second background reasoning model performs heavier analytical work behind the scenes.
The structure resembles a split between reflexes and deeper reasoning where one system keeps the conversation flowing while another handles more computationally intensive thinking.
Why It Could Matter In India
The implications may be especially significant in countries like India, where voice-first internet usage is widespread and millions of users are still entering the AI ecosystem for the first time.
Current AI systems still depend heavily on effective prompting which means it’s important to know how to structure requests clearly. Real-time conversational systems could lower that barrier substantially, especially for users more comfortable speaking than typing.
An AI assistant capable of naturally handling Hindi, Bengali, Tamil or other Indian languages in continuous conversation could eventually make digital systems more accessible for students, small business owners and first-time internet users.
The larger shift is philosophical as much as technical. For decades, humans adapted themselves to machines by learning how to type commands, search efficiently and craft prompts. Interaction models suggest the industry is now trying to reverse that equation by making machines adapt more naturally to human behaviour.
The Caveats Remain
Despite the excitement, Mehra cautions that many claims surrounding the technology remain difficult to independently verify because the system is still only a research preview. “Nobody can really validate these amazing claims via API access,” he said.
He also pointed to one of AI’s persistent technical limitations: long conversations. “Long sessions are still unsolved because as you speak more, the context bloats and AI starts to get confused.”
Still, he believes the trajectory is clear. “When a baby is introduced to the world, it is first taught to speak and see, and then to build skills. Our path to building AI is trying to imitate the exact same pattern.”
And despite the unresolved challenges, Mehra says the pace of improvement remains staggering. “This is the latest worst version of AI we will ever see.” In other words, today’s AI systems – powerful as they already seem – may soon look primitive compared to what is coming next.
