Chinese language startup DeepSeek AI and its open-source language fashions took over the information cycle this week. Moreover being akin to fashions like Anthropic's Claude and OpenAI's o1, the fashions have raised a number of considerations about knowledge privateness, safety, and Chinese language-government-enforced censorship inside their coaching.
AI search platform Perplexity and AI assistant You.com have discovered a method round that.
Additionally: What’s DeepSeek AI? Is it secure? Right here's every part you might want to know
Entry the uncensored mannequin weights
On Tuesday, Perplexity open-sourced the mannequin weights for R1 1776, an uncensored model of R1 that the corporate says "has been post-trained to supply unbiased, correct, and factual info," in response to the announcement. Perplexity notes that its decensoring coaching course of didn’t negatively impression different components of the mannequin's efficiency, together with math and reasoning.
Outcomes from Perplexity's benchmark coaching evaluating their uncensored model of R1 with the unique mannequin.
You possibly can entry this model on HuggingFace or Perplexity's Sonar API.
The usual (and thus presumably nonetheless censored) model of DeepSeek R1 is offered on platforms together with Azure and Github, AWS, and Nvidia.
Strive R1 through AI platforms
As of final month, Perplexity now hosts DeepSeek R1. The free plan offers customers three Professional-level queries per day, which you might use with R1, however you'll want the $20 per thirty days Professional plan to entry the mannequin greater than that.
In one other put up, the corporate confirmed that it hosts DeepSeek "in US/EU knowledge facilities — your knowledge by no means leaves Western servers," assuring customers that their knowledge can be secure if utilizing the open-source fashions on Perplexity.
"None of your knowledge goes to China," Perplexity CEO Aravind Srinivas reiterated in a LinkedIn put up.
DeepSeek's AI assistant, powered by each its V3 and R1 fashions, is accessible through browser or app — however these companies require communication with the corporate's China-based servers, which creates a safety threat.
Additionally: I examined DeepSeek's R1 and V3 coding abilities – and we're not all doomed (but)
Customers who obtain R1 and run it domestically on their gadgets will keep away from that difficulty however nonetheless run into censorship of sure matters decided by the Chinese language authorities, as that characteristic is built-in by default.
As a part of providing R1, Perplexity claimed it eliminated at the least a number of the censorship inbuilt to the mannequin. Srinivas posted a screenshot on X of question outcomes that acknowledge the president of Taiwan.
To verify, I requested R1 about Tiananmen Sq. utilizing Perplexity, however it refused to reply:
Perplexity help clarified that this difficulty was as a result of R1 was set to writing mode, and subsequently not linked to sources that will assist it present an uncensored reply. Nevertheless, after I then requested R1 whether it is skilled to not reply sure questions decided by the Chinese language authorities, it responded that it's designed to "deal with factual info" and "keep away from political commentary," and that its coaching "emphasizes neutrality in world affairs" and "cultural sensitivity."
You.com presents V3 and R1, equally solely via its Professional tier, which is $15 per thirty days (discounted from the same old $20) with none free queries. Along with accessing to all of the fashions You.com presents, the Professional plan comes with file uploads of as much as 25MB per question, a 64k most context window, and entry to analysis and customized brokers.
Bryan McCann, You.com cofounder and CTO, defined in an e-mail to ZDNET that customers can entry R1 and V3 through the platform in 3 ways, all of which use "an unmodified, open-source model of the DeepSeek fashions hosted fully inside the US to make sure consumer privateness."
Additionally: OpenAI's new Deep Analysis agent can do in 5 minutes what may take you hours
McCann continued: "The primary, default method is to make use of these fashions inside the context of our proprietary belief layer. This offers the fashions entry to public internet sources, a bias in the direction of citing these sources, and an inclination to respect these sources whereas producing responses.
"The second method is for customers to show off entry to public internet sources inside their supply controls or by utilizing the fashions as a part of Customized Brokers. This feature permits customers to discover the fashions' distinctive capabilities and habits when not grounded within the public internet. The third method is for customers to check the bounds of those fashions as a part of a Customized Agent by including their very own directions, information, and sources."
McCann famous that You.com in contrast DeepSeek fashions' responses primarily based on whether or not it had entry to internet sources.
Additionally: The perfect open-source AI fashions: All of your free-to-use choices defined
"We observed that the fashions' responses differed on a number of political matters, generally refusing to reply on sure points when public internet sources weren’t included," he defined.
"When our belief layer was enabled, encouraging quotation of public internet sources, the fashions' responses revered these sources, seemingly overriding prior political biases."