DeepMind's on-device scam filter, Anthropic lands in Singapore

What shipped

Google DeepMind

Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection

A fine-tuned Gemma model runs on the phone and reads multi-turn chats for scam patterns, so no conversation data goes to the cloud.

Anthropic

Anthropic opens Singapore office and signs MOU with the Government of Singapore

Anthropic's fifth Asia-Pacific office, plus an MOU with Singapore's digital ministry covering AI safety, security and model evaluation.

OpenAI

openai-python v3.30.0

The Python SDK adds beta session spending limits, organization usage reporting and project API key expiry updates.

ggml-org

llama.cpp b11612

Adds opt-in NFC Unicode normalization for Unigram tokenizers, which matters when local models see composed and decomposed characters.

debpalash

VoiceStudio

An open-source, fully local voice tool covering cloning, dubbing, dictation, transcription and audiobooks, pitched as an ElevenLabs alternative.

K-Dense AI

scientific-agent-skills

A MIT-licensed library of 177 agent skills and 100+ scientific databases for biology, chemistry and drug discovery work.

A quieter day after the Argon launch. Singapore is the thread today, and the rest is tooling worth a look if you build things.

DeepMind and GovTech Singapore: scam detection on the phone

Google DeepMind and Singapore's GovTech built a scam detector around a lightweight Gemma model fine-tuned on local scam conversations. Those include people posing as public institutions, banks and local figures, plus the regional schemes that are trending right now.

It works in two stages. A small classifier screens incoming messages and barely touches the battery. If something looks off, the fine-tuned Gemma model reads the multi-turn conversation to confirm the threat or clear it, then warns the user on the device. No chat data leaves the phone.

The post says scams make up over half of Singapore's reported crime, with losses above S$4 billion since 2019, and that complex social engineering now drives up to 80% of losses. The rollout will be phased with the Singapore Government. There's no open release and no date yet.

Anthropic opens in Singapore

Anthropic opened its fifth Asia-Pacific office there and signed an MOU with the Ministry of Digital Development and Information. The agreement covers AI safety, security and model evaluation, research on adoption, and responsible use across public and private sectors. Anthropic says it plans technical exchanges on testing standards and incident response, and will share research including the Anthropic Economic Index. It also names GovTech, Grab and DBS as existing Claude users in the country.

Smaller releases

openai-python 3.30.0 adds beta session spending limits, organization usage reporting, image URL support and a way to update project API key expiry. It also fixes async HTTP/1.1 connection reuse in streaming. Version 3.31.0 landed the same day and starts a four-part browser and computer toolset.

llama.cpp b11612 adds opt-in NFC normalization for Unigram tokenizers. Boring, and exactly the kind of fix that stops a local model from choking on accented text.

VoiceStudio is an open-source local voice suite (AGPL-3.0 on its repo page). It lists voice cloning, dubbing, dictation, transcription and audiobook creation. The 646-language figure is the project's own claim and I haven't tested it.

scientific-agent-skills from K-Dense AI packs 177 skills and 100+ science databases, MIT licensed, built to plug into Claude Code, Codex and Cursor.

What I'd watch next

Whether the Singapore scam model ever ships outside Singapore, and whether Google publishes false-positive rates. Also the rest of the openai-python toolset series, since parts two to four will show what OpenAI wants agents to do in a browser.