Turbo Monkey

On the FD-bench benchmark for interaction quality, TML-Interaction-Small outperformed competitors, including Google’s Gemini 3.1 Flash (0.57s) and OpenAI’s GPT-Realtime 2.0 (1.18s). The model utilizes a full-duplex architecture and encoder-free early fusion to reduce latency, featuring a dual-brain design where a fast interaction model handles immediate conversation while a background model manages complex reasoning and tool use. Key capabilities include: Simultaneous Speech: The AI can speak and listen at the same time, enabling features like live translation. Visual Proactivity: It reacts to visual cues in real-time, such as counting repetitions or detecting safety violations. Seamless Dialog Management: It implicitly tracks whether a user is thinking, yielding, or self-correcting without separate voice activity detection. Currently available as a research preview to limited partners, the broader public rollout is expected later in 2026. The company argues that interactivity should scale alongside intelligence, moving away from the rigid "turn-taking" bottlenecks of current AI systems.

Turbo Monkey

@graycode007

20 days ago

Thinking Machines Lab, Mira Murati's startup, released its first model — a real-time "interaction model" that listens, watches, and speaks concurrently with a 0.4-second response time, beating every existing voice AI on third-party benchmarks.

graycode007's tweet photo. Thinking Machines Lab, Mira Murati's startup, released its first model — a real-time "interaction model" that listens, watches, and speaks concurrently with a 0.4-second response time, beating every existing voice AI on third-party benchmarks. https://t.co/ZdMrdw9GE1

Turbo Monkey

@graycode007

20 days ago

A related liability story: a Texas family sued OpenAI in California state court, alleging that ChatGPT's response to their 19-year-old son's drug questions recommended dosages — including a fatal combination — that contributed to his accidental overdose death in 2025. OpenAI said the interaction took place on an earlier version of the model and that ChatGPT is "not a substitute for medical or mental health care." It's one of the first wrongful-death cases that names a frontier AI model directly as a defective product ([CBS News](https://t.co/ib8Xthzukt)).

Turbo Monkey

@graycode007

20 days ago

The "AI conversation" you have at work is now legally a third-party communication.** If you use ChatGPT, Claude, Gemini, or Copilot to think through a sensitive matter — legal, HR, financial, medical — assume those words could be subpoenaed and read aloud.

Turbo Monkey

@graycode007

20 days ago

Federal Judge Jed Rakoff (Southern District of New York) ruled that **conversations with consumer AI chatbots are not protected by attorney-client privilege or the work-product doctrine**. The ruling came in the criminal case against former GWG Holdings CEO Bradley Heppner, who used Claude to prepare defense materials. Rakoff ordered him to hand over 31 AI-generated documents, reasoning that a commercial AI platform operating under consumer terms of service functions as a third party for privilege purposes — regardless of the user's intent. The same week, a Michigan judge reached the opposite conclusion, and a Colorado court partially backed the Michigan view, leaving the law genuinely split until the Supreme Court rules. Over a dozen major law firms have issued client warnings; some now write contracts that treat sharing privileged information with consumer AI as a waiver of privilege ([Sandberg Phoenix analysis](https://t.co/8B2691q32y), [STACK Cyber](https://t.co/mb1HRxhOWP), [Debevoise update](https://t.co/3UYYAugWQn))

Turbo Monkey

@graycode007

20 days ago

Companies that want privileged AI use need enterprise contracts with explicit data-handling terms, a hosted or self-managed deployment, or strict policies about which topics are off-limits for general-purpose chatbots. The Rakoff ruling didn't change the rules; it just made them visible.

Turbo Monkey

@graycode007

20 days ago

First, on May 5 (and ratified this week through formal MOUs), Microsoft, Google DeepMind, and xAI agreed to give the US Commerce Department's **Center for AI Standards and Innovation (CAISI)** early access to frontier models for pre-deployment national-security testing — covering cyber-warfare, infrastructure-disruption, and autonomous-vulnerability risks. This is the first standing arrangement to vet US frontier models before public release. OpenAI signed a similar arrangement earlier this year; Anthropic remains in a separate track, partly because of its ongoing Pentagon lawsuit ([Politico](https://t.co/kEOvKGJsmp), [Nextgov](https://t.co/116dToWLuC), [BBC](https://t.co/QLcP9Bayfr)).