conv.

All stories
AIQuiet 9d · day 10

Startups strip censorship from Chinese open-weight AI models for Western buyers

Firms like Hirundo are editing the weights of DeepSeek, Qwen and other Chinese models to remove CCP-aligned bias and harden security before Western companies deploy them.

newest coverage 8d agosummary written 9d ago5 pieces NewswiresMastodonReddit

What to know

  • Startups including Hirundo and Lazarus perform 'machine unlearning' and weight edits on Chinese open-weight models to remove CCP-aligned bias, censorship and security weaknesses before Western deployment.
  • CrowdStrike's analysis of DeepSeek's R1 found prompts on politically sensitive topics were up to 50% more likely to get code with severe security vulnerabilities; a separate study found less secure code when users identified as US government.
  • Hirundo found Alibaba's Qwen models won't engage with topics like Tiananmen Square, Hong Kong protests, Xinjiang labor camps, or Winnie the Pooh references to Xi Jinping.
  • Top Chinese open models trail US frontier closed models by only about 3% on benchmarks and lead downloads on Hugging Face, with Airbnb, Uber and Harvey building proprietary systems on top of them.

“We're not making the model anti-Chinese. We are bringing it up to speed on safety and to the standard of mainstream Western AI.”

Ben Luria, Co-founder and CEO of Hirundo · Rest of World ↗ · Nov 1

Ben Luria Co-founder and CEO of HirundoHirundo AI safety startupDeepSeek Chinese AI labCrowdStrike Cybersecurity firm

Startups strip censorship from Chinese open-weight AI models for Western buyers
restofworld.org

How it unfolded 1 development · click the chart to see its coverage articlesposts

Peak 3 pieces in 3h at Nov 2, 5 AM; 5 pieces over 10 days (2 articles · 3 posts) Nov 2, 5 AM — 3 pieces · 2 articles · 1 post — Newswires 2, Mastodon 1Nov 2, 8 AM — quietNov 2, 11 AM — quietNov 2, 2 PM — quietNov 2, 5 PM — quietNov 2, 8 PM — quietNov 2, 11 PM — quietNov 3, 2 AM — quietNov 3, 5 AM — 1 piece · 1 post — Mastodon 1Nov 3, 8 AM — quietNov 3, 11 AM — quietNov 3, 2 PM — quietNov 3, 5 PM — quietNov 3, 8 PM — quietNov 3, 11 PM — quietNov 4, 2 AM — quietNov 4, 5 AM — quietNov 4, 8 AM — quietNov 4, 11 AM — quietNov 4, 2 PM — quietNov 4, 5 PM — 1 piece · 1 post — Reddit 1Nov 4, 8 PM — quietNov 4, 11 PM — quietNov 5, 2 AM — quietNov 5, 5 AM — quietNov 5, 8 AM — quietNov 5, 11 AM — quietNov 5, 2 PM — quietNov 5, 5 PM — quietNov 5, 8 PM — quietNov 5, 11 PM — quietNov 6, 2 AM — quietNov 6, 5 AM — quietNov 6, 8 AM — quietNov 6, 11 AM — quietNov 6, 2 PM — quietNov 6, 5 PM — quietNov 6, 8 PM — quietNov 6, 11 PM — quietNov 7, 2 AM — quietNov 7, 5 AM — quietNov 7, 8 AM — quietNov 7, 11 AM — quietNov 7, 2 PM — quietNov 7, 5 PM — quietNov 7, 8 PM — quietNov 7, 11 PM — quietNov 8, 2 AM — quietNov 8, 5 AM — quietNov 8, 8 AM — quietNov 8, 11 AM — quietNov 8, 2 PM — quietNov 8, 5 PM — quietNov 8, 8 PM — quietNov 8, 11 PM — quietNov 9, 2 AM — quietNov 9, 5 AM — quietNov 9, 8 AM — quietNov 9, 11 AM — quietNov 9, 2 PM — quietNov 9, 5 PM — quietNov 9, 8 PM — quietNov 9, 11 PM — quietYesterday, 2 AM — quietYesterday, 5 AM — quietYesterday, 8 AM — quietYesterday, 11 AM — quietYesterday, 2 PM — quietYesterday, 5 PM — quietYesterday, 8 PM — quietYesterday, 11 PM — quietToday, 2 AM — quietToday, 5 AM — quietToday, 8 AM — quietToday, 11 AM — quietToday, 2 PM — quietToday, 5 PM — quietToday, 8 PM — quiet 1
Nov 3Nov 4Nov 5Nov 6Nov 7Nov 8Nov 9yesterdaynow · 9:50 PM ET
  1. 1

    Rest of World details startups editing Chinese AI for Western firms

    Rest of World reports that Hirundo, Lazarus and other startups are using machine unlearning and weight edits to harden Chinese open-weight models against cyberattacks, reduce bio-weapons knowledge, and strip state-aligned bias, with Hirundo CEO Ben Luria describing the process as 'like neurosurgery'.

    “What we do is like neurosurgery, an intervention that's surgical and very specific.”
    — Ben Luria
    1. 1 outlet Hirundo, Lazarus, and others are applying machine unlearning and weight edits to Chinese models to remove censorship and harden security for Western companies

      first by Rest of World, 9d ago

    2. 1 outlet Startups recalibrate Chinese AI for the West

      first by Rest of, 9d ago

  2. background

    CrowdStrike finds security flaws tied to politically sensitive prompts — CrowdStrike's analysis of DeepSeek's R1 model found prompts on topics sensitive to the Chinese government were up to 50% more likely to receive code with severe security vulnerabilities; another study found similarly degraded code when users identified as US government.

  3. background

    DeepSeek launches first open-weight model, sparking realignment market — China's DeepSeek released its first open-weight large language model, which entrepreneur Ben Luria says showed him the business potential of realigning such models to Western values.