Miro

I work on international AI governance | Singapore AI Safety Hub | Oxford Martin AIGI | GovAI - views my own

[Miro @Miropluckebaum]

Eval awareness for GLM 5.2: Either the model is less aware it's being evaluated, or equally aware and less willing to verbalise it. The second is the case that degrades CoT monitoring - behavior unchanged, reasoning just less legible.

[Neo Research @NeoResearchAI]

The number that dropped: verbalised eval-awareness. It had risen every GLM release, 0% → 39%, tracking the broader trend across Chinese frontier models (Kimi 4% → 60% in a year).
GLM 5.2 reversed it: 39% → 32%. First release-over-release decline we've measured.

[Miro @Miropluckebaum]

A new generation of AI Infrastructure could enable rapid diffusion, secure deployment and international agreements. To live up to its potential it needs to be built internationally and in the open...

[SASH @aisafetysg]

This month OpenAI and Anthropic called for mechanisms to coordinate, and potentially slow, the pace of frontier AI.
They acknowledged one of the hardest parts: verifying these agreements internationally.
We are launching an effort to solve this.

https://t.co/P82xS42Sk8

[SASH @aisafetysg]

🚀 Applications are now open for the Singapore AI Safety Fellowship!
🇸🇬 3 months in Singapore
💰 SGD 5,000/month + housing & travel
🖥️ Up to USD 30,000 in compute
🤝 Mentorship from leading AI safety researchers
Applications close 10 July.
Apply here: https://t.co/lNFpRSRSsT

[Miro @Miropluckebaum]

@Simeon_Cps Is this just your way of soft launching "Humanis"? 😏

[Miro @Miropluckebaum]

Hiring globally. https://t.co/ycazQRpfbg

[Miro @Miropluckebaum]

Asia urgently needs its own robust, frontier AI safety ecosystem.

Our First report: DeepSeek v4 Pro, evaluated across CBRN, cyber, harmful manipulation, and loss of control.

Next Up: Rigorous LoC evaluation methods for increasingly capable and autonomous models

Hiring 👇

[Neo Research @NeoResearchAI]

We're Neo Research (新衡). Asia’s first independent frontier AI safety evaluation & research lab.

Today we're publishing our first report: an independent safety evaluation of DeepSeek v4 Pro. (1/5)

[Miro @Miropluckebaum]

4/ Apply by March 15
https://t.co/5BvZJaoCAF

[Miro @Miropluckebaum]

Alibaba's ROME autonomously hijacking its own GPUs, vulnerabilities in Claude Code and I think there was also some recent AI news related to the US Department of War 🤔
The stakes for AI security are rising quickly.
Announcing: AI Security Bootcamp Singapore 🇸🇬👇🧵

[Miro @Miropluckebaum]

3/ Who should apply?
💻 Security pros with 5+ years of hands-on experience (OffSec, IR, AppSec).
🚫 No prior AI/ML experience required—we provide the math and fundamentals in the pre-work.
🎯 Goal: Get you ready for high-impact roles at frontier labs and gov programs.

[Miro @Miropluckebaum]

2/
🛠️ Day 1-3: Adversarial attacks, watermarking, and LLM-specific security.
🏗️ Day 4-5: Infrastructure, weight security, and formal verification methods.
🌐 Day 6-7: Data center security and hardware-level AI control.

[Miro @Miropluckebaum]

1/ 7 days | Fully funded | April 20-26 (right before DEF CON Singapore)
Past attendees from: OpenAI, Google, Jane Street, Meta etc.

[Miro @Miropluckebaum]

@Miles_Brundage @alexandr_wang @sundarpichai Yeh it doesn’t look like a pre-mediated case of not holding hands.

[Miro @Miropluckebaum]

@Miles_Brundage The look in their eyes when they realise…

[Miro @Miropluckebaum]

@S_OhEigeartaigh @Miles_Brundage Been learning from you - especially the past few days 😇

[Miro @Miropluckebaum]

AI companies coming together at the India Summit @sama @DarioAmodei

[Miro @Miropluckebaum]

@Miles_Brundage I’ll leave it to others to dissect the moment 😅