Thinking Machines unveiled interaction models that process voice and video simultaneously rather than waiting for turn-based responses, while Perceptron launched its Mk1 video analysis model at 80-90% lower cost than rivals from Anthropic, OpenAI, and Google. The developments signal a shift from sequential AI interactions toward fluid, real-time multimodal capabilities across enterprise applications.
Real-Time Interaction Models Challenge Sequential AI
Thinking Machines, the startup founded by former OpenAI CTO Mira Murati and researcher John Schulman, introduced what it calls “interaction models” that treat real-time responsiveness as core architecture rather than external software layering. According to the company’s announcement, these models can process and respond to human inputs while simultaneously handling new incoming data across text, audio, and video formats.
The approach represents a departure from current AI systems that operate on turn-based interactions — where users provide input, wait for processing, and receive output before continuing. Traditional models require anywhere from milliseconds to hours for complex queries, creating bottlenecks for natural human-AI collaboration.
Thinking Machines reported “impressive gains on third-party benchmarks and reduced latency” from treating interactivity as a first-class architectural component. The company plans to open a limited research preview in coming months before broader availability.
Cost-Effective Video Analysis Enters Enterprise Market
Perceptron launched its Mk1 video analysis model at $0.15 per million input tokens and $1.50 per million output tokens — pricing that undercuts Anthropic’s Claude Sonnet 4.5, OpenAI’s GPT-5, and Google’s Gemini 3.1 Pro by 80-90%. The company spent 16 months developing what CEO Armen Aghajanyan, formerly of Meta FAIR and Microsoft, calls a “multi-modal recipe” for physical world understanding.
The model targets enterprise applications including security monitoring, marketing video analysis, and candidate assessment. Perceptron designed Mk1 to understand cause-and-effect relationships, object dynamics, and physics principles alongside traditional pattern recognition capabilities.
A public demo allows potential customers to test the model’s video analysis capabilities. The pricing structure aims to make advanced video AI accessible to organizations previously priced out by existing solutions.
Structured Reward Systems Improve Multimodal Training
Researchers introduced Auto-Rubric as Reward (ARR), a framework that converts implicit AI preferences into explicit, criteria-based evaluation systems for multimodal models. The arXiv paper describes how ARR externalizes vision-language models’ internalized preference knowledge as prompt-specific rubrics before any pairwise comparison.
Traditional reinforcement learning from human feedback (RLHF) reduces complex human judgments to scalar or pairwise labels, creating what researchers call “opaque parametric proxies” vulnerable to reward hacking. ARR addresses this by translating holistic intent into independently verifiable quality dimensions.
The framework includes Rubric Policy Optimization (RPO), which distills structured multi-dimensional evaluation into binary rewards for training stability. On text-to-image generation and image editing benchmarks, ARR-RPO outperformed standard pairwise reward models and vision-language model judges.
Enterprise Voice Agents Scale Beyond Rule-Based Systems
Parloa evolved from rule-based voice automation to an AI Agent Management Platform built on GPT-5.4, enabling enterprises to design customer service interactions through natural language rather than rigid intent mapping. The OpenAI case study details how the Berlin-based company handles everything from simple routing to complex multi-step requests.
The platform allows business users and subject matter experts to build AI agents without coding, focusing on production consistency where performance, latency, and edge cases matter for real-time conversations. Parloa continuously tests models against real customer scenarios before deployment.
“The models only matter if they work in production. We work closely with OpenAI on how to make the models fast and reliable enough for real-time conversations,” Engineering Manager Ciaran O’Reilly Ibañez told OpenAI.
Model Convergence Reveals Universal Reality Representation
Research from MIT and other institutions suggests major AI models converge toward similar internal representations as they improve at reasoning tasks, regardless of training data differences. Analysis published in Towards Data Science indicates models trained separately on images versus text develop comparable “thinking cores” when they reach sufficient capability levels.
The convergence phenomenon becomes more evident as models improve at reasoning tasks. Researchers theorize that accurate models must create similar representations of reality since there is fundamentally one reality to model correctly.
This finding challenges assumptions that different architectures and data types would produce entirely different AI “brains.” The research draws parallels to Plato’s Allegory of the Cave, suggesting successful AI systems discover similar underlying structures of reality.
What This Means
The multimodal AI landscape is shifting from sequential, turn-based interactions toward real-time, fluid communication systems that better match human conversational patterns. Cost barriers are falling as specialized models like Perceptron’s Mk1 deliver enterprise-grade video analysis at fraction of previous pricing.
Structured reward systems like ARR represent progress toward more reliable AI training that avoids the opacity and gaming vulnerabilities of current approaches. Meanwhile, the convergence of successful models toward similar reality representations suggests there may be optimal ways to structure AI reasoning that transcend specific architectures or training approaches.
These developments collectively point toward multimodal AI systems that can engage more naturally with humans while processing multiple input types simultaneously — a significant step toward AI that feels less like software interaction and more like natural communication.
FAQ
What makes interaction models different from current AI systems?
Interaction models process and respond to inputs while simultaneously handling new incoming data, rather than the turn-based approach where AI waits for complete input before processing. This enables more natural, fluid conversations across voice, video, and text.
How much cheaper is Perceptron’s video analysis compared to major providers?
Perceptron’s Mk1 costs $0.15 per million input tokens and $1.50 per million output tokens, representing 80-90% savings compared to Anthropic’s Claude Sonnet 4.5, OpenAI’s GPT-5, and Google’s Gemini 3.1 Pro.
Why do different AI models converge to similar internal representations?
Researchers suggest that as models become more accurate at reasoning tasks, they must develop similar representations of reality since there is fundamentally one correct way to model the physical world and its relationships.
