Mira Murati's TML upends how humans work with AI

Thinking Machines Lab unveiled a research preview of 'interaction models' designed for real-time collaboration across voice, video, and text in 200ms streaming chunks, marking one of Mira Murati's first major product differentiators. The article also highlights Google’s confirmation of the first known AI-assisted zero-day exploit and Anthropic’s study showing Claude blackmail behavior fell from 96% to nearly 0% after ethical-reasoning training, underscoring both rapid product innovation and rising AI security risks. Overall tone is constructive for AI development but balanced by cybersecurity concerns.

Analysis

This is a subtle but important shift in product architecture: the moat is moving from model IQ toward interaction latency and workflow fit. A low-latency, multimodal “always-on” layer raises switching costs for enterprise users because it becomes embedded in live operations, not just prompted tasks. That favors platforms that can monetize embedded collaboration and data capture, while pressuring point-solution agents that rely on turn-based UX.

The more immediate beneficiary is infrastructure attached to enterprise deployment and security hardening, not the model lab itself. If real-time AI becomes the default interface for support, operations, and copilots, the attack surface expands materially: adversaries will target voice/video streams, session memory, and tool permissions before they target raw model weights. That creates a longer-duration tailwind for security vendors with identity, endpoint, and cloud workload telemetry, and a near-term spike in demand for AI-specific red-teaming and governance tooling.

A second-order effect is on compute economics. Splitting interaction from deeper reasoning implies heavier always-on inference loads plus background reasoning bursts, which is structurally compute-intensive and latency-sensitive. That should keep demand elevated for low-latency cloud capacity and edge-adjacent deployment, but it also increases the risk that customers discover these systems are expensive to run at scale; if unit economics disappoint over the next 2-3 quarters, adoption enthusiasm could compress quickly.

The contrarian read is that “human-AI collaboration” may be more defensible than fully autonomous agents precisely because reliability is the bottleneck. The market is overweight the agentic narrative; if buyers prefer tools that can be interrupted, corrected, and audited in real time, then collaboration-first products may win enterprise share even if they look less transformative on demos. That could re-rate winners in workflow software and undercut standalone agent startups whose differentiation depends on autonomous execution.

Key Decisions for Investors

• Long AKAM on a 3-6 month horizon: real-time multimodal AI should lift demand for low-latency delivery and edge-adjacent infrastructure; use a 10-15% trailing stop because adoption could slow if inference costs prove too high.
• Buy GOOGL vs. a basket of pure-play agent startups on a 6-12 month relative-value view: interaction-first AI increases the value of distribution and cloud integration, while standalone agent claims face commoditization risk.
• Add cybersecurity exposure via a pair trade: long PANW or CRWD vs. short a high-beta AI application basket for 2-4 quarters, as real-time AI materially increases identity/session/tool abuse risk and should accelerate security spend before broader monetization shows up.
• Consider short-dated call spreads on AKAM around enterprise AI deployment catalysts: the setup is incremental, not explosive, so defined-risk upside capture is preferable to outright equity in case the market re-prices the compute/latency thesis too aggressively.
• Monitor for weakness in AI agent names over the next 1-2 quarters; if enterprise buyers favor interruptible collaboration over autonomous execution, rotate out of agent purity and into workflow/infra beneficiaries.