shaun khoo

{ai_safety, fairness, evals}@govtechsg

[Leanne @leannetanyt]

Mental-health distress can be expressed through coded or indirect language, shaped by local context, and open to interpretation.

We built VEILED-SH 🇸🇬 to evaluate how AI systems respond to these expressions in the Singapore context.

https://t.co/ZjKq2faUaw

[Leanne @leannetanyt]

We realised the hardest part of running evals isn’t getting a score, it’s knowing when to trust it.

We’ve open-sourced 𝙆𝙖𝙡𝙚𝙞𝙙𝙤𝙨𝙘𝙤𝙥𝙚, a contextual AI evals workflow built around realistic test generation, human-aligned evaluation, and trustworthy multi-judge scoring.

[Gabriel Chua @gabrielchua]

🚨 new applied ai paper from @govtechsg

LLMs are powerful, but they're prone to off-topic misuse, where users push them beyond their intended scope. Think harmful prompts, jailbreaks, and misuse. So how do we build better guardrails?

Common approaches rely on curated examples or custom classifiers. The problem?
⚠️ High false-positive rates
⚠️ Poor adaptability to new misuse types
⚠️ Require real-world data, which is often unavailable during pre-production

Here, we explore a data-free guardrail development methodology leveraging LLMs to guard LLMs.

(1/9) 🧵

[shaun khoo @shaunkhoo65]

Join us for an exciting panel discussion with @stripe on #datascience in APAC! Our Director, Yuzhang, will be sharing more about how we deliver impactful data science work for the public sector at the Data Science and AI Division in @GovTechSG. See you there!

[Moxie Marlinspike @moxie]

Wrote some notes summarizing my first impressions of web3: https://t.co/16MoPROOGc

[elvis @omarsar0]

Imagine you are in a machine learning interview and want to share your experience building deep learning projects. Don’t try to impress your interviewer by telling them how you applied a Transformer-* and got 99% accuracy. Instead, discuss the following:

[Thread]

1/8

[Lovelytics @LovelyticsData]

Join @UNTUCKit's CIO and CDO, Lockie Andrews, and Lovelytics on Thursday at 2 pm ET to hear how UNTUCKit used @SnowflakeDB and @Tableau to efficiently and accurately analyze their @Shopify data and make critical business decisions.

Register here: https://t.co/EhTwHifB3d

[PyTorch @PyTorch]

We’re starting the blog post series on a Differentially Private Stochastic Gradient Descent Algorithm that will range from an introduction to detailed coverage of the math and engineering details. Here is the first blog of the series: DP-SGD Explained. https://t.co/z7lltt5e61

[shaun khoo @shaunkhoo65]

Updated my website (finally) and posted a new article on estimating the impact of Covid-19 restrictions on urban mobility in Singapore using bus ridership numbers: https://t.co/tOAfL2mYLZ