Technology · 2026-11-04 · Afternoon briefing
Startups 'realign' Chinese AI models for Western use
A new industry is emerging to adapt powerful Chinese artificial-intelligence models for Western businesses. Open-weight models from firms like DeepSeek and Alibaba are gaining popularity, with top performers lagging leading American systems by just 3% on benchmark scores. However, concerns over censorship and security flaws are driving startups like Hirundo and Lazarus AI to "realign" them, stripping out state propaganda and biases.
The risks are tangible. A cybersecurity analysis by CrowdStrike found that one Chinese model was up to 50% more likely to generate code with severe security vulnerabilities when prompted with politically sensitive topics. The realignment process, described as "neurosurgery," involves techniques like machine unlearning to remove censorship. After one such intervention on Alibaba's Qwen model, it was able to name pro-democracy activists, suggesting the knowledge had been suppressed rather than absent.
Despite the commercial appeal—with companies like Airbnb and Uber already building on Chinese models—some experts are sceptical of the goal. Oriane Peter, a researcher at King’s College London, argues that creating a universally "safe" model aligned to "Western values" is an "impossible standard." She suggests any model, regardless of origin, must be adapted to the specific social context in which it operates.
Summary written by Cup of News with AI from the sources below. Check the originals for the full story.
Read the sources
- Startups recalibrate Chinese AI for the West
- Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
- Huge companies have dominated AI. Two startups with an entirely different approach think they can do better
- Chinese AI Models Challenge OpenAI and Anthropic on Cost and Enterprise Risk
- Lawmakers probe growing use of Chinese AI models in U.S. companies
Ten stories, twice a day. Calm, no notifications.
Get the appMore stories in this edition
- Delhi air quality hits 'very poor' as Diwali nears
- Hebron counts the cost of 11-day Israeli army raid
- European Parliament warns MEPs off AI for drafting laws
- Indonesia chief of staff fired over presidential wiretap claims
- Revolut valuation surges to $115bn ahead of US bank launch
- Police warn of sharp objects in Halloween sweets across central Europe
- Yemen: Government forces claim strategic gains from Houthis
- Nations and sponsors revolt over Russia's ice hockey return
- Denmark: 40,000 students strike over grant cuts