America's AI leaders speak about DeepSeek.
Did DeepSeek steal from OpenAI?
OpenAI releases its o3-mini models.
The US Copyright Office publishes part 2 of its report on AI.
After DeepSeek's entry into the market wiped out over $600 billion in market cap from Nvidia and another several hundred billion from other AI companies, AI leadership from America's tech industry are making their voices heard.
On social media this week, venture capital giant Marc Andreessen said: "Deepseek R1 is AI's Sputnik moment"
He's probably right. America felt confident that it had the edge in the space race until the USSR launched Sputnik and showed that there was competition.
In a blog post, Anthropic CEO Dario Amodei discussed DeepSeek and export controls.
To summarize his point on DeepSeek's progress compared to prior models: this is one data point on a curve of increasing efficiency and power, and it just so happens to be the leading data point right now. But other AI labs are still hard at work, and he believes that they will all will continue to leap frog one another.
As for export controls, Amodei said it best:
I don't see DeepSeek themselves as adversaries and the point isn't to target them in particular. In interviews they've done, they seem like smart, curious researchers who just want to make useful technology. But they're beholden to an authoritarian government that has committed human rights violations, has behaved aggressively on the world stage, and will be far more unfettered in these actions if they're able to match the US in AI. Export controls are one of our most powerful tools for preventing this, and the idea that the technology getting more powerful, having more bang for the buck, is a reason to lift our export controls makes no sense at all.
Meta’s Chief AI Scientist, Yann LeCun, gave kudos to the DeepSeek team.
In a statement on LinkedIn, LeCun emphasized that DeepSeek's success is more a testament to the power of open-source models than a sign of China surpassing the US in AI development. He said: "Open source models are surpassing proprietary ones," and pointed out that DeepSeek has benefited from open research and open-source projects, including Meta's own contributions like PyTorch and Llama.
Before the following story came to the surface, OpenAI CEO Sam Altman also gave his kudos to the DeepSeek team.
In a post on X (formerly Twitter), Altman expressed his excitement to have a new competitor in the market and how it will motivate his team to “pull up some releases.”
OpenAI has claimed it has evidence that DeepSeek distilled its proprietary models in training the r1 reasoning model, a clear breach of OpenAI's terms of use if true.
According to a story that first circulated on Reddit and has since made mainstream news, r1's chain-of-thought will sometimes refer to OpenAI policies as applying to it. Unless this is a byproduct of training data that taught DeepSeek that that's just what AIs are supposed to say in situations like this, it seems likely that DeepSeek distilled OpenAI's models in training DeepSeek. The irony of OpenAI, famous for respecting IP rights, having its IP stolen...
What's distillation? AI distillation is the process of transferring the learned knowledge from a large, complex AI model to a smaller, more efficient one, enabling similar performance with lower computational costs. This is extremely common in AI training.
It comes as no surprise that DeepSeek's training involved distillation of other models. After all, their page on HuggingFace (think GitHub but for AI models) is clear that there are versions of their model based on Meta's LLaMA and Alibaba's Qwen. But those models are "open weight"; not quite open source in the conventional sense, but the licensing allows distillation, among other perks.
Open weight AI makes a model’s parameters (its “weights”) publicly available for use and modification, but without necessarily sharing the full training code or data, whereas open source AI goes further by releasing all the underlying code, data, and training processes under a permissive license to ensure complete transparency and collaborative innovation.
In completely unrelated news, OpenAI released o3-mini and o3-mini-high, its latest small reasoning models, today. Definitely unrelated to DeepSeek being free, o3-mini is available to ChatGPT’s free users.
The U.S. Copyright Office has finally released the second part of its report on AI, addressing copyrightability of AI-generated works.
While the report is welcome and long overdue, it does not overhaul the standing policy as many had hoped.
The key takeaway is that works fully generated by AI without additional contribution from humans will still not be protectable. Prompts (no matter how expressive on their own, or how many it takes to land on a final work) are considered mere instructions to the AI which are at best subject to the final approval of the human user.
A few notable quotes:
... there is an important distinction between using AI as a tool to assist in the creation of works and using AI as a stand-in for human creativity. While assistive uses that enhance human expression do not limit copyright protection, uses where an AI system makes expressive choices require further analysis. This distinction depends on how the system is being used, not on its inherent characteristics.
... given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output. Prompts essentially function as instructions that convey unprotectible ideas. While highly detailed prompts could contain the user’s desired expressive elements, at present they do not control how the AI system processes them in generating the output.
Repeatedly revising prompts does not change this analysis or provide a sufficient basis for claiming copyright in the output. First, the time, expense, or effort involved in creating awork by revising prompts is irrelevant, as copyright protects original authorship, not hard work or “sweat of the brow.” Second, inputting a revised prompt does not appear to be materially different in operation from inputting a single prompt. By revising and submitting prompts multiple times, the user is “re-rolling” the dice, causing the system to generate more outputs from which to select, but not altering the degree of control over the process. No matter how many times a prompt is revised and resubmitted, the final output reflects the user’s acceptance of the AI system’s interpretation, rather than authorship of the expression it contains.
Some commenters drew analogies to a Jackson Pollock painting or to nature photography taken with a stationary camera, which may be eligible for copyright protection even if the author does not control where paint may hit the canvas or when a wild animal may step into the frame. However, these works differ from AI-generated materials in that the human author is principally responsible for the execution of the idea and the determination of the expressive elements in the resulting work.
Whether the Copyright Office is right or wrong with this policy has been the subject of debate among copyright nerds like me for a while now.
Clearly the debate is nowhere near done.