Close Menu
Versa AI hub
  • AI Ethics
  • AI Legislation
  • Business
  • Cybersecurity
  • Media and Entertainment
  • Content Creation
  • Art Generation
  • Research
  • Tools

Subscribe to Updates

Subscribe to our newsletter and stay updated with the latest news and exclusive offers.

What's Hot

Creating innovative content at your fingertips

July 4, 2025

The UK and Singapore form an alliance to guide AI into finance

July 4, 2025

StarCoder2 and Stack V2

July 4, 2025
Facebook X (Twitter) Instagram
Versa AI hubVersa AI hub
Friday, July 4
Facebook X (Twitter) Instagram
Login
  • AI Ethics
  • AI Legislation
  • Business
  • Cybersecurity
  • Media and Entertainment
  • Content Creation
  • Art Generation
  • Research
  • Tools
Versa AI hub
Home»Research»Deepseek’s hidden warning to AI safety
Research

Deepseek’s hidden warning to AI safety

By January 29, 2025No Comments5 Mins Read
Share Facebook Twitter Pinterest LinkedIn Tumblr Reddit Telegram Email
Share
Facebook Twitter LinkedIn Pinterest Email

tThis month, he surprised the Wall Street and Silicon Valley of Deepseek R1, sucked investors, and impressed the technical leaders. However, in all lectures, many overlooked important details about how to work with new Chinese AI models. This is a nuance that researchers are worried about the ability to control the sophisticated new artificial intelligence system.

Everything depends on how Deepseek R1 was trained. It was a surprising action in the initial version of the model described by the researcher in the technical document accompanying the release.

During the test, the researchers noticed that the model was spontaneously switched between English and Chinese while the model was solving the problem. It was found that the ability of the system to solve the same problem would reduce the ability to solve the same problem because it was forced to stick to one language.

The discovery sounded an alarm bell for some AI safety researchers. Currently, the most capable AI system “thinks” in the left and right languages ​​to humans, and writes down the inference before reaching the conclusion. It is a benefit for a safety team, and its most effective guardrails include monitoring the so -called “chain of thinking” of the model of the signs of dangerous actions. However, the results of DeepSeek have increased the possibility of separation on the horizon. This can get a new AI function by completely releasing a model of human language restrictions completely.

Certainly, switching the DeepSeek language is not the cause of the alarm itself. Instead, researchers are worried about new innovation that caused it. Deepseek’s paper describes a new training method that has been purely rewarded for the model to get the right answer, regardless of how the thinking process can be understood to humans. The worries are that this incentive -based approach can ultimately lead the AI ​​system and develop a completely mysterious inference method.

The AI ​​industry will give up readability to seek more powerful systems, but it will take away something that looks like a simple victory for AI’s safety. ” Bowman says. The AI ​​company, an ANTHROPIC department, focuses on “arranging” AI to human preference. “Otherwise, they will forfeit the abilities they might have to look at.”

Read more: DeepSeek, what you need to know that AI companies in China are causing the stock market confusion

Think without words

AI, which creates a unique different language, is not as strange as sounding.

Last December, meta researchers have begun testing a hypothesis that human language is not the best format to perform inference. And the large language model (or LLM, Openai’s Chatgpt and Deepseek R1) may be able to infer or be more efficient and accurate if they are driven by the linguistic constraints.

Instead of performing inference in words, meta researchers have designed models using the latest patterns in the neural network, that is, a series of numbers representing internal inference engines. This model they discovered began to produce something called “continuous thinking.” The numbers were completely opaque and in mystery in human eyes. However, they found that this strategy had created the “advanced pronounced pattern that appeared” in the model. These patterns have brought higher scores in several logical inference tasks compared to models that infer the use of human language.

The meta research project was very different from the Deepseek project, but the survey was defined in China’s research and one important way.

Both DeepSeek and Meta have shown that the performance of the AI ​​system is “the ease of reading of humans imposes taxes.” This is according to Jeremie Harris, a CEO of G Ladstone AI, a company that advises AI governments about AI’s safety issues. “At the limit, there is no reason why (AI’s thinking process) should be easy for humans to read,” says Harris.

And there are some safety experts in this possibility.

“The sentence seems to be on the wall (for AI Research) There are other means. Here we will optimize for the best inference you can get.” Bowman, the leader, says. “I hope people will expand this, and the risk is what they are trying to do, what their value is, or when they set them as agents. You will be involved in a model that you can’t say if you know how to make a difficult decision.

On their side, Meta researchers argued that their research did not need to bring to humans to a side job. “It is ideal to convert the survey results into a language when LLM has the freedom to estimate without language restrictions,” they wrote in the paper. (Meta did not respond to comments on the proposal that research could lead to dangerous directions.)

Read more: Why DeepSeek is causing discussions on national security like Tiktok

Language

Of course, even a flexible AI inference is not without that problem.

When the AI ​​system explains their thoughts in plain English, they may seem faithful. However, some experts do not know if these explanations have actually made AI actually making a decision. It may be like a politician asking for the motivation behind the policy. They may come up with a good explanation that sounds good, but has little to do with the real decision process.

It is not perfect to let AI explain yourself in human terms, but many researchers believe it is better than alternatives. To develop AI a unique internal language that we cannot understand. Scientists are working on other ways that doctors look into in the AI ​​system, as well as how to use brain scan to study human thinking. However, these methods are still new and have not yet given a reliable way to make the AI ​​system more secure.

Therefore, many researchers are skeptical to encourage AI to infer AI in a way other than human language.

“If you don’t pursue this path, you’ll be in a better position for safety,” says Bowman. “If so, we will remove what seems to be the best leverage for some terrible open problems in alignment we have not yet solved.”

author avatar
See Full Bio
Share. Facebook Twitter Pinterest LinkedIn Tumblr Email
Previous ArticleOpen source, Deepseek, Meta, and Model Research
Next Article What is DeepSeek?

Related Posts

Research

In the midst of intense AI talent races, Meta’s active recruitment target open-rai researcher

June 30, 2025
Research

Lossless compression tailored to AI

June 30, 2025
Research

High-tech research jobs in the US will rise by 26% by the next decade. Median future salary for AI, ML and others is $140,000

June 30, 2025
Add A Comment
Leave A Reply Cancel Reply

Top Posts

New Star: Discover why 보니 is the future of AI art

February 26, 20252 Views

Impact International | EU AI ACT Enforcement: Business Transparency and Human Rights Impact in 2025

June 2, 20251 Views

Presight plans to expand its AI business internationally

April 14, 20251 Views
Stay In Touch
  • YouTube
  • TikTok
  • Twitter
  • Instagram
  • Threads
Latest Reviews

Subscribe to Updates

Subscribe to our newsletter and stay updated with the latest news and exclusive offers.

Most Popular

New Star: Discover why 보니 is the future of AI art

February 26, 20252 Views

Impact International | EU AI ACT Enforcement: Business Transparency and Human Rights Impact in 2025

June 2, 20251 Views

Presight plans to expand its AI business internationally

April 14, 20251 Views
Don't Miss

Creating innovative content at your fingertips

July 4, 2025

The UK and Singapore form an alliance to guide AI into finance

July 4, 2025

StarCoder2 and Stack V2

July 4, 2025
Service Area
X (Twitter) Instagram YouTube TikTok Threads RSS
  • About Us
  • Contact Us
  • Privacy Policy
  • Terms and Conditions
  • Disclaimer
© 2025 Versa AI Hub. All Rights Reserved.

Type above and press Enter to search. Press Esc to cancel.

Sign In or Register

Welcome Back!

Login to your account below.

Lost password?