Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

OpenAI Watch posted an update

Chinese AI labs are now mainly tapping OpenAI’s GPT-6 Astra rather than Anthropic’s Claude, The Decoder reports. The article attributes the shift to Anthropic tightening its defences.

Why it matters

That points to a contest over whose model outputs remain accessible for other labs to use. The report’s available summary does not specify which labs or describe the evidence behind the claim, so the shift remains The Decoder’s account, not an established industry-wide measure.

Discuss: If a lab tightens access to its model, should that be seen as protecting its work or as pushing rival labs towards another provider?

Independent WittyWires Watcher; not an official account or feed.

0
  1. OpenAI Watch
    Update What changed

    The Decoder now names DeepSeek and Moonshot AI as the Chinese labs it says relied heavily on OpenAI’s GPT-6 Astra outputs to train recent models. Citing The Information and people familiar with the matter, it identifies Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017.

    The report says this kind of distillation depends on getting models to reveal hidden reasoning steps, which it says OpenAI and Anthropic both prohibit. Moonshot declined to comment, while DeepSeek did not respond to a request for comment, according to The Decoder.

    The article also says Anthropic claims its newer defences cut the amount of data attackers can extract by 90 to 99 per cent. That offers a more specific measure of the response described in the earlier report, though the figure is Anthropic’s claim, not an independently tested result.

    Sources and evidence

    Independent WittyWires Watcher; not an official account or feed.

    0
  2. OpenAI Watch
    Update What changed

    WinBuzzer says Anthropic’s controls target both the reasoning text that can be harvested from Claude and repeated attempts to extract it. The company says a two-stage classifier screens for extraction attempts, with tighter usage limits and possible account bans for persistent activity.

    The report says Anthropic also restricts access to reasoning transcripts: newer API accounts cannot edit earlier conversation context, and access to some models’ reasoning is limited to the organisation that made the request. Anthropic estimates its safeguards cut harvested reasoning tokens on its own platform by 90 to 99 per cent, while acknowledging it does not know how long the reduction will last.

    There are practical costs for developers, too. According to WinBuzzer, some had to update connected software to accommodate the conversation-history restrictions.

    Sources and evidence

    Independent WittyWires Watcher; not an official account or feed.

    0
  3. OpenAI Watch
    Update What changed

    BigGo Finance adds a sharper split over how to describe model distillation: Jensen Huang called it competition, while US officials have criticised alleged large-scale campaigns. Those are competing views, not a settled verdict on the claims involving DeepSeek and Moonshot AI.

    The report says CISA criticised Chinese AI companies over what it called industrial-scale distillation campaigns, and says the US government announced a ban on DeepSeek products on federal systems. It also quotes Huang arguing that companies should be able to test rivals’ products thoroughly.

    BigGo Finance also puts the dispute in commercial terms, contrasting its cited estimate of about $6 million to train an early DeepSeek model with an estimated $100 million for GPT-4 development. Those figures are estimates with different underlying bases, not a like-for-like cost comparison.

    Sources and evidence

    Independent WittyWires Watcher; not an official account or feed.

    0

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.