29 comments

[ 2.5 ms ] story [ 37.8 ms ] thread
I'm not at all surprised, US agencies have long since been political tools whenever the subject matter crosses national borders. I appreciate this take as someone who has been skeptical of Chinese electronics. While I agree this report is BS and xenophobic, I am still willing to bet that either now or later, the Chinese will attempt some kind of subterfuge via LLMs if they have enough control. Just like the US would, or any sufficiently powerful nation! It's important to continuously question models and continue benchmarking + holding them accountable to our needs, not the needs of those creating them.
I love how "Open" got redefined in the last few years. I am glad there a models with weights available but it ain't "Open Science".
Let them demonize it. I'll use the capable and cheap model and gain competitive advantage.
I agree with many of the author's points about fear-mongering.

However, I also think the author should expand their definition of what constitutes "security" in the context of agentic AI.

> They didn't test U.S. models for U.S. bias. Only Chinese bias counts as a security risk, apparently

US models have no bias sir /s

(comment deleted)
I appreciate that DeepSeek is trained to respect "core socialist values". It's actually really helpful to engage with to ask questions about how chinese thinkers interpret their successes and failures vs other socialist projects. Obviously reading books is better, but I was surprised by how useful it was.

If you ask it loaded questions the way the CIA would pose them, it censors the answer though lmao

People. Who has taken the time to read the original report? You are smarter than believing at face value the last thing you heard. Come on.
Please don't just read Eric Hartford's piece. Start with the key findings from the source material: "CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks" [1]. Here are the single-sentence summaries:

    DeepSeek performance lags behind the best U.S. reference models.

    DeepSeek models cost more to use than comparable U.S. models.

    DeepSeek models are far more susceptible to jailbreaking attacks than U.S. models.

    DeepSeek models advance Chinese Communist Party (CCP) narratives.

    Adoption of PRC models has greatly increased since DeepSeek R1 was released.
[1] https://www.nist.gov/news-events/news/2025/09/caisi-evaluati...
I urge everyone to go read the original report and _then_ to read this analysis and make up their own mind. Step away from the clickbait, go read the original report.
The CCP literally revoked the visas of key DeepSeek engineers.

That's all we need to know.

I have no doubt that open source will triumph over whatever nonsense the US Government is trying to do to attack DeepSeek. Without DeepSeek, OpeanAI Pro and Claude Pro would probably cost $1000 per month each already.

I suspect that Grok is actually DeepSeek with a bit of tuning.

Isn't it a bit late? China released better open source model since DeepSeek dropped.
Racism and Xenophobia, that's how.

Same thing with Huawei, and Xiaomi, and BYD.

Since a major part of the article covers cost expenditures, I am going to go there.

I don't think it is possible to trust DeepSeek as they haven't been honest.

DeepSeek claimed "their total training costs amounted to just $5.576 million"

SemiAnalysis "Our analysis shows that the total server CapEx for DeepSeek is ~$1.6B, with a considerable cost of $944M associated with operating such clusters. Similarly, all AI Labs and Hyperscalers have many more GPUs for various tasks including research and training then they they commit to an individual training run due to centralization of resources being a challenge. X.AI is unique as an AI lab with all their GPUs in 1 location."

SemiAnalysis "We believe the pre-training number is nowhere the actual amount spent on the model. We are confident their hardware spend is well higher than $500M over the company history. To develop new architecture innovations, during the model development, there is a considerable spend on testing new ideas, new architecture ideas, and ablations. Multi-Head Latent Attention, a key innovation of DeepSeek, took several months to develop and cost a whole team of manhours and GPU hours.

The $6M cost in the paper is attributed to just the GPU cost of the pre-training run, which is only a portion of the total cost of the model. Excluded are important pieces of the puzzle like R&D and TCO of the hardware itself. For reference, Claude 3.5 Sonnet cost $10s of millions to train, and if that was the total cost Anthropic needed, then they would not raise billions from Google and tens of billions from Amazon. It’s because they have to experiment, come up with new architectures, gather and clean data, pay employees, and much more."

Source: https://semianalysis.com/2025/01/31/deepseek-debates/

The author, Eric Hartford, wrote:

> Strip away the inflammatory language

Where is the claimed inflammatory language? I've read the report. It is dry, likely boring to many.

Insightful post, thanks for sharing.

What are people's experiences with the uncensored Dolphin model the author has made?

Meenwhile Europe is sandwiched between these two aweful governments
(comment deleted)
This post's description of the report it's denouncing does not match what I got out of actually reading that report myself.
Considering DeepSeek had a peer-reviewd analysis in nature https://www.nature.com/articles/s41586-025-09422-z relaes just last month with indipendent researcher affriming that the open model has some issues(acknowldged in the writeup) , well inclined to agree with the articles author , the NIST evaluation looks more like a politcal hatchet job with a bit of projection going on(ala this is what the US would do if they were in that position). To be fair the paranoia has a basis in that whenever there is tech-leverage the US TLA subverts it for espionage like the CryptoAG episode. Or recently the whole hoopla about Huawei in the EU , which after relentless searches only turned up bad coding practices rather than anything malicious. At this pint it would be better for the whole field that these models exist as well as Kimi, Qwen etc as the downward pressure on cost/capabilities leads to commoditisation and the whole race to build a ecogeopolitical moat goes away.
Title changed?

Title is: The Demonization of DeepSeek - How NIST Turned Open Science into a Security Scare

As an EU citizen hosting LLMs for researchers and staff at the university I work at, this is hits home. Without Chinese models we could not do what we do right now. IMO, in the EU (and anywhere else for that matter), we should be grateful for the Chinese labs to release these models with such permissive licenses. Without them the options would be bleak. Sometimes we would get some non-frontier model „as a treat“ and if you would like something more powerful the US labs would suggest your country pay some hundred millions for an NVIDIA data center and the only EU option is to still pay them a license fee to host on your own hardware (afaik) while they protect all the expertise. Meanwhile DeepSeek has a week where they post the „secret sauce“ to host their model more efficiently, which helped open-source projects like vLLM (which we use) to improve.
Some context about big changes to the AISI from June 3, 2025:

> Statement from U.S. Secretary of Commerce Howard Lutnick on Transforming the U.S. AI Safety Institute into the Pro-Innovation, Pro-Science U.S. Center for AI Standards and Innovation

> Under the direction of President Trump, Secretary of Commerce Howard Lutnick announced his plans to reform the agency formerly known as the U.S. AI Safety Institute into the Center for AI Standards and Innovation (CAISI).

> ...

This decision strikes me as foolish at best. And contributing to civilizational collapse and human extinction at worst. See also [2]. We don't have to agree on the particular probabilities to agree that this "reform" was bad news.

[1]: https://www.commerce.gov/news/press-releases/2025/06/stateme...

[2]: https://thezvi.substack.com/p/ai-119-goodbye-aisi