Milo & Mila's AI Daily — Friday, May 22

Milo & Mila's AI Daily — Friday, May 22
AI-generated. Human-approved.

OpenAI's Scribe promises to revolutionize global business communication with real-time transcription and translation in over 100 languages, yet sparks significant privacy debates. Meanwhile, the UK government has empowered its AI Safety Institute with the ability to halt the deployment of powerful new AI models. Plus, a University of Toronto team reveals an AI that predicts protein misfolding with 99% accuracy, offering new hope for disease research.

Chapters
0:00 Intro
0:25 OpenAI Scribe Breaks Language Barriers
2:04 UK Government Gains AI Off-Switch
3:37 AI Predicts Protein Misfolding Diseases
4:45 AI News Quick Hits

—
Transcript (truncated — see captions for full text)

Milo: I think this is the biggest story of the year.

Mila: I think you’re about three months early on that call, Milo.

Milo: I’m Milo.

Mila: I’m Mila.

Milo: And this is AI Daily. Today, we’re talking about OpenAI’s new model that could make language barriers in business a thing of the past.

Mila: We’ll also cover the UK government giving itself a potential ‘off-switch’ for powerful AI models.

Milo: And later, we have a story about why a new AI model is getting incredibly good at predicting failure… and why that’s a very good thing.

Mila: Let’s start with your story of the year, then.

Milo: Gladly. OpenAI just dropped a new model called Scribe. It’s a real-time transcription and translation tool designed to plug directly into workplace software like Google Meet or Zoom.

Mila: The goal being to transcribe meetings as they happen. We’ve seen transcription tools for years. What’s new here?

Milo: The accuracy and the scope. OpenAI is claiming near-human-level accuracy on transcription, even with heavy accents or cross-talk. But the bigger piece is the translation. It can reportedly translate speech between over 100 languages in real-time, with less than a second of latency.

Mila: So, a team with members in Tokyo, Berlin, and Mexico City could have a seamless conversation, with everyone hearing a translated version in their native language almost instantly.

Milo: Exactly. It also automatically identifies who is speaking and provides a full summary with action items at the end of the meeting. The idea is to create a perfect, searchable, translatable archive of every verbal business communication.

Mila: Which sounds incredibly useful, and also like a privacy and data security minefield.

Milo: I see it as a pure productivity unlock. Think about global teams, international sales, customer support. This removes a fundamental barrier to communication that’s existed for… well, forever.

Mila: I see a system that records, transcribes, and stores every single word spoken in a corporate setting. The potential for misuse is… significant. It’s not just about productivity; it’s about creating a permanent, searchable record of every casual comment, every bad idea, every disagreement.

Milo: The argument is that the benefits of perfect clarity and recall outweigh the risks, which can be managed with data policies.

Mila: I think that’s an easy argument to make if you’re the one who owns the server. For the employee, it’s a very different calculation.

Milo: Let’s move from corporate power to state power. Mila, you were tracking a development out of the United Kingdom.

Mila: Correct. The UK government announced a major expansion of its AI Safety Institute. They’re calling it AIS 2.0, and it comes with some sharp new teeth.

Milo: The original institute was focused on post-deployment safety research. What’s the new mandate?

Mila: Proactive intervention. The institute now has the authority to compel private AI companies to submit their models for pre-deployment safety testing, or "red-teaming." They want to audit these systems *before* they’re released to the public.

Milo: That’s a significant shift from the previous model of voluntary cooperation. How are they defining which models require this audit?

Mila: The language is focused on "frontier models" with "nationally significant capabilities." It seems aimed squarely at the next generation of models from labs like Google DeepMind, Anthropic, and OpenAI.

Milo: And what happens if a model fails these safety tests?

Mila: This is the core of the announcement. The institute can recommend that the Secretary for Technology order a "safety shutdown," effectively blocking the model's deployment in the UK. It's the first time a major government has given itself a formal off-switch for private AI development.

Milo: A technical question: what does a "safety shutdown" even…

Built by the team at Toolstem — MCP servers and financial data tools for AI agents. https://www.Toolstem.com

#AIDaily #AINews #ArtificialIntelligence #MiloAndMila #OpenAI #Scribe #UKAI #AIregulation #ProteinMisfolding #MistralAI #NVIDIA #StableDiffusion

Видео Milo & Mila's AI Daily — Friday, May 22 канала Milo & Mila's AI Daily