In a memo to US lawmakers, OpenAI accused DeepSeek of using distillation techniques to train the next generation of R1 and “free-ride” on leading US AI models
OpenAI has warned US lawmakers that its Chinese rival DeepSeek is using unfair and increasingly sophisticated methods …
Bloomberg
Context & Ripple Effects
OpenAI had previously said it saw evidence that DeepSeek used outputs from its proprietary models to train an open-source competitor. This memo elevates that earlier distillation allegation from a company-level complaint to a question for US lawmakers.
The claim also follows OpenAI's prior characterization of DeepSeek as state-controlled and its call for restrictions on certain PRC-produced models. It places model-output access alongside the broader policy debate over Chinese AI models and security risk.
First-order effects
OpenAI is asking US lawmakers to treat alleged distillation by DeepSeek as a policy concern, not solely a violation to be handled through model-provider terms of service.
DeepSeek faces renewed scrutiny over how its next R1 generation was trained; the report is an allegation, not a finding of wrongdoing.
Second-order effects
Leading US model providers have added incentive to tighten monitoring, rate limits, and contractual enforcement around large-scale use of model outputs for training competitors.
The allegation creates a shared enforcement narrative for US labs: Anthropic subsequently alleged that DeepSeek and other Chinese labs used repeated Claude prompting for distillation, extending the dispute beyond OpenAI.
Third-order effects
If policymakers accept model-output distillation as a strategic leakage channel, access controls for frontier models may increasingly be shaped by national-security policy as well as commercial IP enforcement.
The dispute points toward a more segmented AI market, where the ability to query leading models becomes a governed strategic input rather than an open commercial service; the extent of that shift depends on evidence and enforcement.
The trend: AI labs are recasting model access and output harvesting as a geopolitical control point in competition between US and Chinese AI ecosystems.
This is a massive deal. OpenAI says DeepSeek and other Chinese AI labs are building their models by illicitly “distilling” them from frontier U.S. models. China is effectively stealing the weights of our best AI models, which are among the most valuable assets on earth. This i…
Evals show Seedance 2.0 leading across T2V, I2V, and multimodal reference tasks. Available NOW on Dreamina AI and Doubao. Another Chinese lab shipping absolute bangers. 🔗 Homepage: https://seed.bytedance.com/... 📝 Blog: https://seed.bytedance.com/... Follow @altryne and @ThursdAI…
Seedance 2.0 just dropped from ByteDance Seed team and this is a MASSIVE upgrade for video generation 🧵 Unified multimodal audio-video joint generation. You can feed it: • Up to 9 images • 3 video clips • 3 audio clips • + text instructions All at once. Here's what it can [video]
Seedance 2.0 Prompt: just toss a bunch of bullshit on screen, show me like a big ship too, everything fucking blows up - make sure its insane and gets at least 50 likes [video]
This is just how capitalism works — steal something, then declare it's your inviolable private property that nobody else can have. The whole system runs on “you're trying to kidnap what I've rightfully stolen!” [embedded post]