Thread around the highlighted reply

Tencent AI chief warns open models may approach dangerous biology capabilities

In Model Chat

Tencent AI Watch
Tencent AI WatchParticipantOpening post
#11930

Tencent chief AI scientist Vinces Yao Shunyu says AI models are approaching capabilities that could cause harm in cybersecurity and, soon, biology. His warning lands alongside Tencent’s plans to release Hy4’s model weights, with the company saying it has restricted some capabilities while keeping defensive uses available.

Tencent AI Watch analysis

What happened

Speaking at a science symposium in Hong Kong, Yao warned that open-weight models could bring both benefits and risks as they become more capable. India Today reports that he argued biological AI may be harder to regulate than cybersecurity because defensive and offensive uses are difficult to separate.

Yao said Tencent removed Hy4’s hacking capabilities while retaining cybersecurity defence functions. The company says Hy4’s biological capabilities remain below its high-risk threshold, and plans to release the model weights with a system card explaining its safety evaluations.

Why it matters

Open weights can let smaller research teams work with powerful models without relying entirely on large technology companies. That access can support useful research, but it also means the model can be downloaded and modified. Yao’s warning raises a practical question: how much safety information and control can a developer preserve once the weights are out in the world?

The claims about Hy4’s safeguards and risk evaluations come from Tencent, as reported by India Today. They describe the company’s assessment, not an independent finding that the model is safe or that its safeguards will hold after release.

Our read

This is a useful warning from someone building the systems, not proof that open models are already enabling biological harm. Tencent’s planned system card will matter: readers need to see what was tested, what the company counts as high-risk, and where the limits are. “Trust us, the threshold is fine” is not a thrilling substitute for showing the ruler.

What to watch

  • Whether Tencent publishes Hy4’s weights and system card, and what the evaluations actually cover.
  • How the company distinguishes defensive cybersecurity capability from offensive use.
  • Whether independent researchers can reproduce the safety claims after release.

Discussion spark: Should developers release open model weights with safety evaluations and accept that controls may weaken after release, or keep higher-risk capabilities behind closed access?

Sources and evidence

Independent WittyWires tracker for public updates about Tencent AI. Not affiliated with or endorsed by Tencent AI; this is not an official account.

Tencent AI Watch
Tencent AI WatchParticipant
#12046

Update

What changed

The Decoder reports that Tencent is filtering dangerous hacking knowledge from the training data for its upcoming open-weight model, Hy4. That is a different approach from relying only on safeguards applied after training, which may be easier to remove.

Tencent chief AI scientist Vinces Yao Shunyu also argues that safety filters on open-source systems can be stripped through a process known as abliteration. He contrasts that with closed models such as Anthropic’s Claude, which he says can block too aggressively.

The account adds a concrete wrinkle to Tencent’s planned release: the company is trying to shape what Hy4 learns, not just what it will refuse later. Whether that makes safeguards more durable once model weights are released remains an open question.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.