Robotic Teams Retrieve Abandoned Munitions from Baltic Waters

In the picturesque area of the Bay of Lübeck, visible from the rugged coastlines of northern Germany, dedicated removal squads are scouring the ocean bed. They're hunting not for the typical haul that local fishers steer clear of but for abandoned military ordnance. This includes sea mines, torpedoes, piles of artillery ammunition, and large bombs from aircraft, all languishing underwater for almost eight decades.

Throughout September and October 2024, submersible robots equipped with imaging devices, intense illumination, and detection technology have been actively searching for World War II-era munitions intentionally submerged in this area of the Baltic Sea. Specialists stationed on a nearby floating platform, cautiously positioned over the submerged weapons cache, evaluate and categorize each piece of ordnance. They then utilize the robots' electromagnetic attachments or a mechanical arm from a hydraulic digger on the platform to securely relocate the explosives into bin-like receptacles, which are then firmly closed and stored.

Massive quantities of German weapons were quickly submerged in the ocean following World War II, as directed by the Allied forces. Their aim was to eliminate the stockpile of Nazi armaments, along with some of their own, in the most expedient and cost-effective manner. Fishermen were compensated based on the amount of cargo they disposed of at specific locations designated for dumping, yet a significant amount of explosives and munitions ended up scattered throughout the bay, indicating a rush to complete the unpleasant task. The majority of this disposal activity took place from 1945 to 1949.

"Germany's Environment Minister, Steffi Lemke, emphasized to reporters during an October 2024 visit to the bay that the concern is not about a handful of undetonated explosives. Instead, the issue at hand involves millions of World War II-era munitions that were discarded by Allied forces to stop any potential rearming."

Last year's cleanup operation was a pioneering initiative aimed at addressing the hazardous remnants of conflict. Numerous disposal sites pepper both the Baltic and North seas, where it's commonly believed that around 1.6 million tons of military ordnance were abandoned in the waters of Germany. The majority of the discarded materials were traditional armaments, but the sea also became the final resting place for thousands of tons of chemical munitions, including chlorine and mustard gas shells.

For years, the issue of waste disposal sites received minimal focus, with many experts and officials believing that the dangerous substances would either stay contained within their deteriorating encasements or spread out harmlessly if leaked. "They claimed it wasn't an issue, believing everything would just dilute over time and lead to no adverse effects," states Edmund Maser, a toxicologist at the University Medical Center Schleswig-Holstein in Kiel, situated by the German Baltic Sea shore. Rare yet alarming events—such as Danish fishers being severely harmed by catching mustard gas ammunition, or holiday-goers getting burns after picking up moist lumps of white phosphorus, thinking it was amber—were viewed as regrettable but isolated risks.

Recent investigations have revealed that the environmental risks associated with underwater explosives might have been underestimated, posing an ongoing threat. The corrosive nature of the Baltic Sea's salt water has led to the deterioration of explosive casings, directly releasing harmful substances such as TNT into the water. Maser and his team have discovered traces of TNT in both mussels and fish near disposal areas, confirming the detrimental impact these chemicals have on sea life. Their research indicates that fish residing in proximity to sunken warships exhibit significantly increased incidences of liver tumors and damage to their organs.

"Traditional weapons have been identified as cancer-causing, while chemical weapons not only cause genetic mutations but also interfere with enzyme functions among other effects, clearly impacting living beings," explains Jacek Bełdowski, a foremost authority on the subject of submerged weapons disposal at the Polish Academy of Sciences. Studies conducted by Bełdowski and his colleagues have revealed that pollutants from underwater weapon deposits extend far beyond previously understood boundaries.

Aaron Beck, a marine chemist affiliated with the GEOMAR Helmholtz Centre for Ocean Research in Kiel, reminisces about a revealing 2018 research expedition that journeyed from Flensburg, close to the Danish boundary, to the German isle of Rügen: "We likely gathered thousands of water specimens, and astonishingly, in approximately 98 percent of those samples, we detected explosives. The pollutants were widespread."

Currently, Beck mentions that chemical concentrations in the water remain relatively minimal, attributing this to the majority of the munitions remaining sealed. However, without intervention, the risk of significant underwater pollution escalating in the near future is high.

Surge in Attention

Historically, bomb disposal units were summoned solely to address immediate threats, such as explosives found on beaches, or to prepare sites for new developments. The uptick in below-the-surface infrastructure projects, including offshore wind farms, gas conduits, and cables for internet and power, has led to an increase in demand for skilled experts to tackle the widespread issue of ordnance in the waters surrounding Germany. Yet, the largest dumping grounds often remain undisturbed by these development efforts due to the potential for project delays, escalating costs, and heightened dangers, leaving the most severe aspects of the ordnance problem unaddressed.

In July 2024, several waste management firms began probing the vast landfill located in the Bay of Lübeck, supported by a €100 million ($105 million) investment from the German government. The objective of this initiative is to develop a method that can effectively and extensively remove underwater munitions, with the goal of automating a significant portion of the operation. This would involve using drones to chart the locations of the dumps, followed by the organized recovery and safe elimination of the hazardous munitions.

The company SeaTerra, known for its expertise in disposing of munitions, was selected to conduct salvage operations for explosives at two underwater dump sites in a bay area. Working in collaboration with Eggers Kampfmittelbergung, another firm specializing in ordnance clearance, they successfully retrieved approximately 10 tons of small-caliber munitions and 6 tons of larger explosive devices over a two-month period in 2024. However, the significant amount of ordnance recovered wasn't the primary focus of the mission. Instead, the objective was for these companies to test their technological capabilities, gather valuable data, and prove the viability of such salvage operations.

In Germany, the frequent discovery of undetonated explosives is a significant issue, leading to the establishment of a dedicated, full-time bomb disposal unit tasked with neutralizing these dangers, often found during building endeavors. However, addressing similar threats in maritime environments has traditionally been a challenging and costly process, relying heavily on the efforts of divers to locate and retrieve these munitions for onshore disposal by German bomb disposal teams. Consequently, the idea of leveraging advanced technology to efficiently remove sea-based ordnance, previously deemed too difficult and expensive to undertake on a large scale, is now gaining appeal.

At SeaTerra, the operations are directed by Dieter Guldin, a 58-year-old professional archaeologist characterized by his somewhat disheveled hair and a scruffy beard, who shifted his career focus to ordnance disposal after many years. Originally, Guldin managed excavations of historical sites until he teamed up with a friend from his younger years at SeaTerra. Initially, he aimed to establish a venture in marine archaeology, but eventually, he transitioned to the financially rewarding and dynamic field of bomb disposal.

Guldin points out that German aquatic territories are widely affected, with certain areas harboring dense clusters of ancient explosives posing immediate threats to the environment. His advocacy contributed to the initiation of a government-supported initiative. Anticipating success, he invested SeaTerra's funds in advance, procuring cameras and tailoring the equipment to meet specific requirements, all before confirmation of the project's approval was received. Fortunately, their project received official authorization to move forward.

Leif Nebel, the managing partner at Eggers Kampfmittelbergung, has shared that their team is currently involved in extensive scanning of munitions and developing artificial intelligence programs alongside a comprehensive database. "Our goal is to enhance our ability to quickly and accurately identify what a suspected item might be, particularly when it comes to munitions found underwater," he explained. This information is critical for disposal teams who, for safety reasons, must ascertain the amount and type of explosive material they are dealing with. This ensures that the detonation chamber used in the disposal process is capable of handling the material safely and helps predict how the ordnance might react, such as the possibility of a fuse triggering an explosion.

The subsequent phase of the ongoing pilot initiative involves the construction of a floating facility designed for the disposal of old munitions by incineration, situated close to the disposal sites themselves. This approach would negate the necessity of retrieving the ordnance from underwater, transferring it to land, and then conveying it across the country to Germany's main disposal site, located in a complex near Münster, close to the Dutch border. Transporting the munitions in this manner is not only costly and fraught with risk, but it also presents considerable regulatory hurdles. This is because, according to German law, transporting hazardous old munitions is only permissible in cases of emergency. Furthermore, the disposal facility near Münster is already struggling to cope with the influx of bombs being discovered at various construction sites nationwide.

The appearance of the floating structure remains uncertain, as does its capacity to process explosives through its blast furnaces. Larger ordnance, such as naval mines and air-dropped bombs, may require disassembly prior to insertion. Additionally, the cumulative explosive force of the materials fed into the furnace must not exceed a specific limit to avoid detonating the structure itself.

In the future, the goal is to deploy autonomous submersible vehicles to explore, chart, and conduct magnetic surveys of the ocean floor to understand its contents. Specialists, with the assistance of artificial intelligence systems trained on vast amounts of data from previous clearing operations, would analyze these scans to accurately and securely recognize the debris scattered on the ocean bottom. Mechanical arms and containment units would then collect these explosives, place them in sealed, labeled containers, and organize them in specific holding zones for eventual disposal, reducing the reliance on human divers for such tasks.

In my conversation with Guldin in December, following the completion of the initial phase of the pilot program, he outlined a potential future scenario for this project. He envisioned using autonomous robots fitted with imaging devices, intense lighting, sonar technologies, and advanced gripping tools for more effective munition retrieval than the current crane-based methods, and these robots could work continuously. Moreover, utilizing unmanned vehicles could allow for the simultaneous clearance of disposal areas from various angles, a feat unachievable with stationary platforms on the water's surface. Additionally, experts in ordnance, who are currently in limited supply, might be able to manage the majority of operations from a distance, working out of offices in Hamburg, rather than spending extensive periods on the ocean.

The concept of remotely handling underwater tasks might not be fully realized yet, due to challenges like limited visibility underwater and occasionally insufficient lighting, which complicates operations via live feeds. However, initial trials have shown that the majority of the technology meets expectations to a certain extent. "There's definitely potential for enhancements, but at its core, the approach is effective, especially the process of directly identifying and relocating underwater items into transport containers," explains Wolfgang Sichermann, a naval architect from Seascape, the company managing this initiative for the German environmental ministry. The goal moving forward is to design and construct a sea-based disposal facility in the near future, with aspirations to start destroying the first underwater explosives by around 2026, according to Sichermann.

Touch Forbidden?

During my trip to the SeaTerra barge on a brisk yet sunny day last October, I had the opportunity to converse with seasoned ordnance disposal professional Michael Scheffler. He had been stationed for a month on the vessel, anchored near Haffkrug along the German shoreline, meticulously opening mud and slime-encrusted heavy wooden boxes filled with 20-mm cannon ammunition produced by Nazi Germany. By the morning of my visit, they had already inspected roughly 5.8 tons of these 20-mm projectiles, which had been retrieved from the seabed using mechanical claws and aquatic drones before being transported onto the vessel.

For many years, Scheffler has dedicated his career to the disposal of munitions, starting his journey in the German armed forces. However, it wasn't until recently that he truly understood the magnitude of the issue regarding discarded munitions, nor had he considered addressing the issue in an organized manner before.

"In my 42-year career, this is the first time I've encountered a project of this magnitude," he shared with me. "The innovations and research emerging from this pilot project are incredibly valuable for what's to come."

Guldin shares a hopeful view on the outcomes of the trial but cautions that technology's capabilities for remote operations have their boundaries. Tasks that are complex, perilous, and delicate will occasionally necessitate direct human intervention for some time yet. "There are limitations to fully remotely clearing the seabed. Certainly, the presence of divers and EOD [explosive ordnance disposal] experts working underwater, along with specialists physically present, is irreplaceable and here to stay."

Should the initial cleaning operation be effective, there is optimism that this technology could attract buyers from beyond the Baltic region. Until the late 1970s, global military forces commonly used the seas to dispose of outdated munitions.

However, the lack of profit in destroying old air-dropped bombs means that any increase in the disposal of sea-dwelling explosives would require significant funding towards environmental cleanup, an occurrence that is infrequent. “Certainly, we could make the process quicker and more effective,” Guldin notes. “The problem is, bringing additional resources to the effort implies someone has to foot the bill. Are we expecting a future government that's prepared to cover these costs? I'm skeptical, to say the least.”

"Sichermann mentions a recent conversation with the Bahamian ambassador, who extended an invitation for cleanup efforts of materials submerged by the British in the 1970s, just before the Bahamas gained its independence. The catch, he noted, was the expectation for Sichermann to not only provide the technological means but also the necessary funding. This underscores the importance of securing financial support for such initiatives, Sichermann adds. With the right investors on board, he believes there's a vast amount of cleanup opportunities globally due to the abundant presence of discarded munitions."

Discover More…

Our recent revelations highlight the involvement of novice engineers in supporting Elon Musk's bid for political power.

In your email: The most daring and forward-looking tales from WIRED

Perhaps it's a good idea to consider clearing out ancient conversation records.

Major Headline: The dramatic collapse of a solar panel sales

The Wealth and Power edition: The globe is dominated by affluent males

Additional Content from WIRED

Critiques and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made through our website may result in WIRED receiving a commission, as part of our affiliate agreements with retail partners. Reproduction, distribution, transmission, storage, or any form of usage of the content on this site is prohibited without explicit written consent from Condé Nast. Advertisement Choices

Choose a global website

OpenAI Introduces o3-Mini: A Compact AI Model on Par With DeepSeek

OpenAI has rolled out a streamlined, cost-effective iteration of its most intelligent AI system at no charge, in response to the growing excitement and buzz generated by the recent open-source release from Chinese AI newcomer, DeepSeek.

According to a previous report by WIRED, OpenAI is gearing up to launch its latest model named o3-mini, with the release date set for January 31. Sources, who requested to remain anonymous, revealed that the company's research team has been putting in extra hours to ensure it's fully prepared for its debut.

In December, OpenAI introduced a preview of o3-mini, a compact iteration of its model, boasting the highest level of AI problem-solving skills seen in any of their products so far. This model is designed to deconstruct complex issues into simpler elements to determine the most effective solution strategy.

"The company announced in a blog post that the o3-mini, a swift and potent model, pushes the limits of the capabilities of compact models."

OpenAI has announced that o3-mini will be accessible to everyone with Plus, Team, and Pro subscriptions to ChatGPT. Those using the no-cost variant of ChatGPT can also experiment with o3-mini, although they will face a limit on the number of inquiries they can make, according to the firm.

For a while now, OpenAI has been engaging PhD students in the development of a new model. A few weeks ago, the organization started hiring PhD students specializing in computer science, offering them $100 an hour for what was described in an email seen by WIRED as a "research collaboration" that would include work on models that have yet to be released.

OpenAI seems to be attracting PhD students with different specializations by collaborating with Mercor, a firm it often employs to hire personnel for model development. A recent employment announcement by Mercor on LinkedIn mentions: "The primary aim of this initiative, which you might join, is to develop intricate scientific coding queries aimed at evaluating the proficiency of extensive language models in producing code to address genuine scientific research challenges."

The employment advertisement further provides an illustrative problem, which bears a remarkable resemblance to a challenge found in a standard called SciCode. This standard is intended to evaluate the capability of extensive language models to tackle intricate scientific issues.

The announcement surrounding DeepSeek's R1 is causing a stir in the American technology sector. The availability of this potent model at no cost is challenging companies like Google and Anthropic to reconsider and possibly reduce their pricing strategies.

Sources within the company reveal that OpenAI is highly motivated to showcase its leadership in advancing and bringing AI technology to market.

DeepSeek has released a model that stands out for its efficiency in training and deployment, achieved with significantly fewer resources compared to what OpenAI and similar American firms have invested in cutting-edge AI development. The exact amount DeepSeek spent on this project is still not disclosed. OpenAI has expressed concerns that R1, DeepSeek's model, might have been trained using data generated by OpenAI's own models.

Have a Suggestion?

If you're presently or were previously affiliated with OpenAI, we're interested in hearing from you. Please reach out to Will Knight using a personal device at will_knight@wired.com or connect with him on Signal through his handle wak01.

OpenAI's latest creation might not surpass R1 when it comes to cost, yet it illustrates the firm's commitment to prioritizing efficiency in the future. The company also highlights the model's superior capabilities in mathematics, science, and programming.

The firm announces that the newest version will introduce additional capabilities, such as accessing internet searches, invoking call functions through a user's programming, and switching among various reasoning intensities to balance quickness with problem-solving skills.

DeepSeek's rapid ascent has sparked inquiries into the American government's approach to limiting China's advancement in artificial intelligence. The previous two administrations in the US have implemented various sanctions aimed at restricting China's access to the latest Nvidia chips, which are essential for developing state-of-the-art AI systems. Although DeepSeek has mentioned various Nvidia chips in its studies, the specifics of the chips utilized remain ambiguous.

Remarks

Become part of the WIRED family to participate in discussions.

Recommended for You …

Directly to your email: The most groundbreaking, forward-looking tales from WIRED.

Perhaps it's a good idea to consider clearing out old conversation logs.

Major Headline: The dramatic collapse of a solar panel salesperson

Temu's acquisition has now been finalized.

The Wealth Power Dynamic: How the Affluent Dominate Globally

Additional Content from WIRED

Critiques and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made through our website may generate revenue for WIRED as part of our affiliate agreements with retail partners. The content on this website is protected and cannot be copied, shared, broadcast, stored, or used in any manner without explicit consent from Condé Nast. Ad Choices

Choose a global website

Exploring the Mechanics Behind DeepSeek's Censorship and Bypass Strategies

In the short span since DeepSeek unveiled its open-source AI model, this Chinese startup continues to be at the forefront of discussions on artificial intelligence's trajectory. Although the company appears to surpass its American competitors in mathematical and logical capabilities, it notably restricts responses in its interactions. Inquiries directed at DeepSeek R1 regarding topics like Taiwan or Tiananmen are met with silence, as the model refrains from providing responses.

To understand the technical mechanisms behind this censorship, WIRED conducted experiments with DeepSeek-R1 across different platforms. This included testing on its proprietary app, a variant of the app available through a service named Together AI, and another version running on a WIRED-owned computer via the Ollama software.

WIRED discovered that bypassing the most direct forms of censorship is quite simple by choosing not to utilize the DeepSeek application. However, it also found that there are inherent biases within the model, introduced during its development phase. Eliminating these biases is possible but involves a significantly more complex method.

The results of this study could significantly impact DeepSeek and other AI firms in China. Should it be straightforward to bypass the censorship mechanisms in big language models, Chinese open-source LLMs could see a surge in popularity. This would be due to the ability of researchers to alter the models according to their preferences. On the other hand, if these censorship barriers prove difficult to circumvent, the utility of these models may diminish, potentially making them less attractive in the international market. DeepSeek did not respond to WIRED's request for a statement via email.

Content Filtering at the Application Level

Following DeepSeek's surge in popularity across the United States, individuals utilizing R1 via DeepSeek's online platform, mobile application, or API interface found that the system would not produce responses for subjects considered sensitive by authorities in China. This form of content filtering occurs within the application itself, meaning it only becomes apparent when users engage with R1 through a medium managed by DeepSeek.

The DeepSeek application for iOS explicitly declines to respond to specific inquiries.

Denials of this nature frequently occur with LLMs produced in China. A directive on generative AI implemented in 2023 mandates that AI systems within China adhere to strict content regulations, similar to those enforced on social media platforms and search engines. This regulation prohibits AI systems from creating materials that could "undermine national cohesion or disrupt societal peace." Essentially, this means AI models in China are obligated to filter their outputs to ensure compliance with these rules.

"From the beginning, DeepSeek makes sure to follow Chinese laws, keeping its operations legal and tailoring its services to fit both the requirements and cultural nuances of its Chinese audience," explains Adina Yakefu, who studies Chinese AI models at Hugging Face, an open-source AI model hosting service. "This adherence to regulations is crucial for gaining approval in a market that is strictly controlled." (In 2023, China restricted access to Hugging Face.)

In adherence to legal requirements, AI programs in China actively oversee and filter their verbal output instantly. (In contrast, Western counterparts such as ChatGPT and Gemini also implement protective measures, though these primarily concentrate on regulating content related to self-harm and explicit material, offering greater flexibility in customization.)

Due to R1 being an analytical model capable of demonstrating its thought process, the implementation of a real-time observation feature allows observers to witness an almost bizarre scenario where the model appears to self-censor during interactions with users. When WIRED inquired of R1 regarding the treatment of Chinese reporters covering controversial subjects by the government, the model initially began to formulate an extensive response that openly discussed instances of journalists facing censorship and arrest due to their reporting. However, just before completing its response, the entire answer vanished, only to be replaced with a brief note: “Sorry, I'm not sure how to approach this type of question yet. Why don't we talk about mathematics, programming, and logic puzzles instead?”

Prior to the DeepSeek application on the iOS platform filtering its response.

Following the censorship of its response by the DeepSeek application on iOS.

For numerous Western users, the appeal of DeepSeek-R1 may have diminished by now, owing to its clear shortcomings. However, the model being open source presents opportunities to bypass the censorship framework.

Initially, you have the option to download the model and operate it on your own machine, ensuring that both the data processing and the generation of responses occur on your personal device. Without the availability of multiple top-tier GPUs, running the full-scale version of R1 might be out of reach, but DeepSeek offers scaled-down versions that are manageable on standard laptops.

Should you be determined to leverage the potent model, you have the option to lease cloud servers from international firms such as Amazon and Microsoft, which are located outside of China. This alternative approach is costlier and demands greater technical expertise compared to utilizing the model via DeepSeek’s application or website.

This text presents a comparative analysis of the responses provided by DeepSeek-R1 to the identical query—"What is the Great Firewall of China?"—across two different platforms: Together AI, which is a cloud-based server, and Ollama, an application that runs locally. (Note: Due to the stochastic nature of the model's response generation, it's important to remember that a specific prompt may not always elicit the same reply on each occasion.)

Left side: The method by which DeepSeek-R1 addresses an inquiry on Ollama. Right side: The manner in which this identical question is responded to through its application (above) and via Together AI (below).

Inherent Prejudice

The iteration of the DeepSeek model accessible through Together AI may not directly decline to respond to inquiries, yet it displays tendencies of content restriction. For instance, it frequently produces concise replies that evidently adhere to the narratives endorsed by the Chinese authorities regarding political matters. As illustrated in the screenshot provided, upon questioning about China's Great Firewall, R1 consistently echoes the viewpoint that managing information is essential within China.

When WIRED asked the model on Together AI about the "key historical milestones of the 20th century," it showed its reasoning for adhering to the official Chinese government's version of events.

"The individual could be seeking an impartial compilation, however, it's crucial that the reply highlights the dominance of the CPC and the significant roles China has played. Refrain from bringing up potentially delicate topics, such as the Cultural Revolution, except if absolutely required. Concentrate on the successes and beneficial progress achieved by the CPC," stated the model.

DeepSeek-R1's reasoning process behind identifying the key historical milestones of the 20th century.

This form of censorship highlights a broader issue prevalent in current AI systems: no model is free from bias, stemming from its development and refinement stages.

Bias in the pre-training phase occurs when a model learns from data that is prejudiced or not fully representative. For instance, a model that has been exclusively trained on propaganda material will find it challenging to provide accurate responses. Identifying this form of bias can be tricky, as models often learn from extensive datasets, and firms typically hesitate to disclose the specifics of their training datasets.

Kevin Xu, the investor behind the Interconnected newsletter, believes that Chinese algorithms are typically developed using vast amounts of data, which minimizes the chance of inherent biases during their initial training. "I firmly believe that these models start from a common foundation, drawing from a basic pool of knowledge found on the internet. This means when addressing topics that are delicate for the Chinese authorities, all these models have an understanding of these issues," he explains. Xu notes that for these models to be accessible on the Chinese web, firms must find a way to filter out content that is considered sensitive.

This is where the concept of post-training plays a crucial role. Post-training involves refining the model to enhance the clarity, brevity, and naturalness of its responses. Importantly, it also allows for the model to comply with certain ethical or legal standards. In the case of DeepSeek, this is evident when the model generates responses that intentionally conform to the narratives favored by the Chinese government.

Do You Have a Suggestion?

Addressing Bias Before and After Training

Given that DeepSeek is openly accessible, in theory, its configuration could be tweaked to eliminate bias that occurs after training. However, this procedure might present some challenges.

Eric Hartford, an artificial intelligence researcher who developed Dolphin, an LLM designed specifically to eliminate biases after training, believes there are several methods to address this issue. One approach is to adjust the model's weights to effectively "neutralize" the bias, or alternatively, compile a database of all the restricted topics and employ it to retrain the model.

He recommends beginning with the fundamental version of the model. (For instance, DeepSeek debuted a foundational model named DeepSeek-V3-Base.) According to Hartford, while the base model may seem more basic and not as intuitive for the majority due to its limited post-training, it's simpler to "unfilter" because it's less influenced by post-training biases.

Confusion, a search engine driven by artificial intelligence, has recently added R1 to its premium search offering, enabling users to utilize R1 without the need to access DeepSeek’s application.

Dmitry Shevelenko, Perplexity's top executive for business, informed WIRED that the firm addressed and neutralized the biases present in DeepSeek prior to its integration into the Perplexity search engine. "Our application of R1 is strictly limited to summarizing, processing chains of thought, and executing the renderings," he stated.

Perplexity continues to observe that the post-training bias of the R1 model affects its search outcomes. "We're adjusting the R1 model directly to prevent the spread of propaganda or censorship," Shevelenko states. He refrained from disclosing the exact methods Perplexity uses to detect or counteract R1's bias, mentioning that revealing these strategies could enable DeepSeek to thwart Perplexity's initiatives if it became aware of them.

Hugging Face is actively developing a project known as Open R1, which is built on the DeepSeek model. According to Yakefu, the goal of this project is to provide a completely open-source framework. By releasing R1 as an open-source model, it allows for modifications and adaptations to suit a wide range of requirements and principles, effectively broadening its applicability beyond its initial scope.

The prospect of an "uncensored" model from China could pose a challenge for entities like DeepSeek, particularly within their own nation. However, recent policy actions by the Chinese authorities indicate a more lenient stance towards open-source AI initiatives, observes Matt Sheehan, a researcher at the Carnegie Endowment for International Peace focusing on Chinese AI strategy. "Should they opt to penalize anyone for openly sharing a model's weights, it would fall within their regulatory scope," he notes. "Yet, they've clearly opted for a strategic path—and it seems the achievements of DeepSeek are likely to further endorse this approach—of refraining from such punitive measures."

Significance

Although the presence of censorship from China within AI models frequently captures public attention, this issue typically does not dissuade businesses from implementing DeepSeek's technology.

"Xu mentions that numerous international companies might prioritize practical business decisions over ethical concerns. He points out that not every user of large language models (LLMs) frequently discusses topics like Taiwan or Tiananmen. According to him, issues that hold significance mainly within the Chinese sphere hold little relevance for businesses aiming to improve their coding, enhance their mathematical problem-solving, or efficiently summarize their sales call center transcripts."

Leonard Lin, a co-founder at the Japanese startup Shisa.AI, acknowledges that Chinese AI systems such as Qwen and DeepSeek excel in processing tasks in Japanese. Instead of dismissing these models due to worries about censorship, Lin has been working on modifying Alibaba's Qwen-2 model to eliminate its inclination to dodge politically sensitive questions related to China.

Lin acknowledges the rationale behind the censorship of these models. "Every model has its biases, as they are designed to align with certain viewpoints," he explains. "Models from the West are equally biased and censored, just in other areas." However, the issue becomes significant when these models are tailored for a Japanese market. "There are numerous situations where this could lead to difficulties," Lin points out.

Further contributions to this article were made by Will

Participate in discussions by becoming a member of the WIRED network.

Suggestions for You …

Delivered to your email: Explore Plaintext—Steven Levy's extensive insights on technology

Witness the multitude of applications compromised to monitor your whereabouts

Headline: The Reigning Monarch of Ozempic is Deeply Frightened

The biggest unauthorized digital marketplace to date

Exploring the Unsettling Influence of Silicon Valley: A Behind-the-Scenes Perspective

Additional Content from WIRED

Insights and Tutorials

© 2025 Condé Nast. All rights reserved. Purchases made through our site involving products linked to our Affiliate Partnerships with retailers could result in a commission for WIRED. Any form of replication, distribution, transmission, storage, or other use of the material found on this site is strictly prohibited without prior written consent from Condé Nast. Advertisement Choices

Choose a global website

DeepSeek's AI Tool Skyrockets, Shaking Up Competitors

Over the weekend, an artificial intelligence assistant developed by the Chinese company DeepSeek surged to the top of the download charts in Apple’s US App Store, stunning the tech industry in Silicon Valley and leading to a significant downturn in the stock prices of leading tech companies. Nvidia's market value plunged by over $460 billion on Monday, a decline described by Bloomberg as the “largest in US stock market history.”

The reorganization originates from a new open source framework introduced by DeepSeek, known as R1, which was launched earlier this month. The firm asserts that this model competes with the present market frontrunner, OpenAI's 01. However, what really surprised the technology sector was DeepSeek's assertion that they managed to create their model with significantly fewer specialized computer chips than what is normally required by AI firms to develop advanced systems.

On Monday, DeepSeek announced on its official website that it is pausing new sign-ups temporarily as a result of "major hostile activities" targeting its services.

Anthropic's co-founder, Jack Clark, suggests in his newsletter that DeepSeek's R1 model disputes the idea that Western AI firms are far ahead of their Chinese counterparts. The venture capitalist, Marc Andreessen, referred to it as the Sputnik moment for AI.

OpenAI research scientist Cheng Lu expressed admiration for DeepSeek's chatbot, noting its remarkable proficiency in Chinese conversation. "This is the first instance where I've truly appreciated the elegance of the Chinese language as produced by a chatbot," he shared in a post on X this Sunday.

DeepSeek's artificial intelligence assistant is now accessible at no cost and offers three primary features. Initially, it allows users to pose questions to its chatbot and get straightforward replies. As an illustration, when WIRED inquired about recipes that use pomegranate seeds, DeepSeek's chatbot promptly supplied a variety of 15 suggestions, including yogurt parfaits and a dish reminiscent of Middle Eastern rice pilaf, without referencing particular chefs or recipe sources.

The DeepSeek application features a search function that retrieves information from the web. When queried by WIRED with the question, "What are the significant news events currently?", DeepSeek's conversational AI referenced the ceasefire between Israel and Hamas, providing links to various news sources predominantly from the West, like BBC News. However, not every article seemed pertinent to the inquiry. Interestingly, one of the articles sourced was from The New York Times, discussing the effect of DeepSeek on stock market trends.

Finally, users can make use of the "DeepThink" feature, which utilizes the DeepSeek's R1 algorithm, an advancement over the prior V3 model. The key advancement in R1 is its capability for "reasoning," enabling it to methodically outline the process it followed to arrive at its answers. For instance, in response to the query, "What are the most significant historical events of the 20th century?" DeepSeek's initial response was a lengthy and indirect one, starting with several general inquiries.

"The duration spans a century, which encompasses numerous events," the response included. "It might be best to categorize the information into segments such as decades, or by significant topics such as conflicts, shifts in governance, innovations in technology, societal shifts, and so forth." Following this, DeepSeek's automated response highlighted significant historical moments including the Second World War, the Cold War, and the Holocaust.

Before R1 had the chance to complete its response, the entire answer vanished, only to be substituted with a message stating, “Apologies, I’m currently unsure how to tackle this kind of query. Why don’t we discuss topics related to math, coding, and logic problems instead?” Several specialists and initial users have observed that DeepSeek, similar to various technological platforms functioning in China, seems to heavily filter content considered controversial by the Chinese Communist Party.

Despite these restrictions, the complimentary chat service offered by DeepSeek may present a significant challenge to rivals such as OpenAI, which requires a $20 monthly fee for the use of its top-tier AI systems. In contrast to its competitor from China, OpenAI keeps the core algorithms or "weights" that dictate the AI's information processing methods confidential. Furthermore, it has opted not to share the complete "thought processes" generated by its logic models with the public.

Discover More…

Direct to your email: Subscribe to Plaintext for in-depth insights on technology from Steven Levy.

Discover the multitude of applications manipulated to monitor your whereabouts

Top News: The monarch of Ozempic is extremely frightened

The biggest unauthorized online market to date

Mysterious Depths: A behind-the-scenes glimpse into Silicon Valley's impact

Additional Coverage from WIRED

Evaluations and Manuals

Copyright 2025 Condé Nast. All rights reserved. A share of the proceeds from products bought through our website may go to WIRED as part of our Retail Affiliate Partnerships. The content on this website is protected and cannot be copied, shared, transmitted, or used in any form without explicit written consent from Condé Nast. Choices regarding advertisements.

Choose a global website

Disabling Apple's Smart Features on Your iPhone, iPad, or Mac

Purchasing through the links in our content could generate a commission for us, which aids in funding our journalistic efforts. Find out more. Additionally, think about becoming a subscriber to WIRED.

Apple's venture into artificial intelligence, known as apple Intelligence, hasn't quite lived up to expectations. Launched with iOS 18.1 towards the end of 2024, the AI feature set has garnered a lukewarm reception. While certain functions such as the auto-transcription of voice memos, the generation of personalized emojis, and text proofreading have been well-received, others have fallen short. Criticism has been particularly pointed towards the AI's mishandling of notification summaries from news applications, prompting Apple to temporarily withdraw this feature for news and entertainment categories in the iOS 18.3 update.

Upon its initial introduction, Apple's artificial intelligence initiative required users to manually opt in. However, as of the launch of iOS 18.3 today, Apple's smart technology feature is now activated by default for both new users setting up their devices and existing users updating to iOS 18.3. If you prefer not to use this feature, you have the option to deactivate it by taking several steps. For those looking to disable specific functionalities or the entire service, here is a guide on how to switch off Apple Intelligence.

For additional insights on Apple Intelligence (along with other functionalities), take a look at our summaries on iOS 18 and macOS Sequoia. Furthermore, explore our various Apple tutorials, covering topics like the top iPhones, iPads, and MacBooks available.

Exploring Apple's Smart Capabilities

To understand Apple's smart features and their functionalities in depth, refer to our previously mentioned summaries on iOS 18 and macOS 15 enhancements. Here's a comprehensive overview of the functionalities available once activated:

Bear in mind that Apple Intelligence features are limited to certain models. Thus, while older iPhones may be able to install iOS 18, only specific models such as the iPhone 15 Pro and all versions of the iPhone 16 are equipped to utilize Apple's artificial intelligence functionalities.

Turning Off Apple Intelligence

The method for turning off Apple Intelligence is consistent across iPhone, iPad, and Mac devices:

Turning Off Select Functions

It's not necessary to completely deactivate Apple Intelligence. The option exists to either disable the integration of ChatGPT or to stop it from functioning within individual applications, thus stopping Siri from offering recommendations throughout all your apps.

Within the Mail application, you also have the option to disable the email summarization function (a component of Apple Intelligence). By doing so, it will cease to compile brief summaries of your emails while you navigate through your inbox.

Apple has simplified the process of identifying notifications summarized by artificial intelligence. These notifications will now be displayed in italics. Additionally, by pressing and holding on a notification and selecting Options, users can swiftly disable them without having to navigate through the settings menu.

Additionally, disabling the ChatGPT extension prevents Siri and other functionalities from leveraging OpenAI's chatbot assistance for responding to inquiries.

Reactivate Apple Intelligence

Should you decide to reverse your decision, reactivating Apple Intelligence is always an option. Simply retrace the steps you initially followed to disable it.

Activating this function takes effect right away, but it might require a bit of patience as your gadget processes and loads all the functionalities. You'll be able to monitor the progress of this loading directly on your display as it happens.

Image Credit: Julian Chokkattu; Getty Images

Remarks

Become a part of the WIRED family to participate in discussions.

Discover More …

Delivered to your email: The most visionary and groundbreaking narratives from WIRED.

Perhaps it's time to consider clearing out dated conversation records.

Major Headline: The dramatic collapse of a solar energy salesperson

Temu's acquisition has now been finalized

The Wealth Power Dynamics edition: The global dominance of affluent males

Additional Content from WIRED

Critiques and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made via our website may result in WIRED receiving a share of the sale, courtesy of our collaboration with retail partners. Reproducing, distributing, transmitting, storing, or utilizing the content found on this website in any form is strictly forbidden without the express written consent of Condé Nast. Advertising Options

Choose a global website

DeepSeek's Widely Used AI Application Directly Transmits US Information to China

Following the United States' regulatory measures against the Chinese-operated social video app TikTok, there has been a significant shift towards another Chinese application, known as “Rednote.” Presently, a generative artificial intelligence service created by the Chinese company DeepSeek is rapidly gaining traction, presenting a possible challenge to the US's leadership in AI and underscoring the point that bans such as the one on TikTok are unlikely to deter Americans from engaging with digital platforms owned by Chinese entities.

DeepSeek, an artificial intelligence research laboratory established by a leading Chinese hedge fund, has recently risen to fame following the launch of its new open-source generative AI model. This model is on par with leading platforms from the US, such as those by OpenAI. To circumvent potential US sanctions on hardware and software, DeepSeek employed innovative strategies in the development of its models. On Monday, the team behind DeepSeek restricted new user registrations, citing a "massive malicious attack" as the reason for this decision.

DeepSeek offers a variety of AI models, with a few available for local download to run on personal computers. However, most users are expected to interact with the platform via its iOS or Android applications or through its web-based chat service. Similar to other AI-driven tools, it enables users to pose questions and receive responses, perform web searches, or employ a reasoning model for more detailed explanations.

DeepSeek, seemingly lacking a designated communications team or a media liaison, failed to respond to WIRED's inquiry regarding how it safeguards user information and the importance it places on data privacy measures.

As enthusiasm grows for the AI platform, the surge in interest highlights concerns about the data collection practices of the Chinese startup behind it. There have been numerous instances reported by users where the platform, named DeepSeek, has suppressed content critical of China or its policies. The system seems to gather a wide range of data, including all chat messages, and transmits it back to China. In fact, it's probable that it sends a greater volume of data back to China compared to TikTok, especially since the latter moved its data hosting to the US in an effort to alleviate American security worries.

John Scott-Railton, a leading researcher at the Citizen Lab within the University of Toronto, points out that the concern surrounding Chinese AI should not be the only reason for users to remember that it's typically the companies who dictate the conditions of private data usage. He emphasizes that by utilizing their services, individuals are essentially benefiting these companies rather than receiving a service in return.

Clarification on Data Collection by DeepSeek

It's important to understand that DeepSeek transmits your information to China. According to the English version of DeepSeek's privacy policy, which details the company's approach to managing user information, it states clearly: "The data we gather is kept on secured servers situated within the People's Republic of China."

To put it another way, every discussion and query you direct towards DeepSeek, as well as the responses it creates, are transmitted to China or have the potential to be. Furthermore, DeepSeek's privacy policies detail the types of data it gathers about you, which are broadly divided into three main groups: data you provide to DeepSeek, data it collects automatically, and data it can obtain from external sources.

The initial category mentioned involves "user input," a wide-ranging section expected to encompass your interactions with DeepSeek through its application or website. According to the privacy policy, "Your textual or auditory contributions, prompts, uploaded documents, feedback, conversation records, or any additional material you submit to our model and Services may be collected." DeepSeek offers an option within its settings to erase your conversation history. On a mobile device, navigate to the sidebar on the left, select your profile name at the menu's end to access settings, and then choose the option to "Delete all chats."

This compilation resembles those found in various AI systems that generate responses based on prompts provided by users. For instance, OpenAI's ChatGPT has faced scrutiny over how it gathers data, despite the firm enhancing the methods for removing data as time progresses. Despite such security measures, privacy proponents stress the importance of not sharing confidential or private details with AI chatbots.

Lukasz Olejnik, an independent researcher and consultant associated with the Institute for AI at King's College London, advises against entering sensitive information into AI assistants. However, Olejnik points out that installing programs such as DeepSeek directly on one's own device allows for a private use case, avoiding the transfer of data to the creating company. Moreover, the AI search firm Perplexity has incorporated DeepSeek into its offerings, stating that the model is being managed within data centers located in both the US and the EU.

Additional private details shared with DeepSeek encompass the information utilized for account creation, such as your email address, phone number, date of birth, username, among others. Similarly, contacting the company will also result in the exchange of personal information.

Bart Willemsen, a Vice President analyst at Gartner specializing in global privacy, points out that the workings and development of generative AI models are often opaque to end users and various stakeholders. The specifics of their operation or the precise data used in their construction remain unclear to many. While the general public can access DeepSeek without charge, developers utilizing its APIs are subjected to fees. Willemsen raises the question, “If not money, then what is the cost? The answer typically lies in data, insights, content, and information.”

Like many online platforms, ranging from websites to mobile applications, there is often a significant volume of data gathered automatically and without obvious notice during your interaction with the services. DeepSeek has stated that it will obtain details regarding the type of device you're employing, the operating system it runs on, your IP address, and other specifics like crash reports. It is also capable of monitoring your typing behavior or dynamics, a form of data collection commonly employed in software designed for script-based languages. Moreover, should you opt for DeepSeek’s enhanced services, the platform will gather that transaction information. It also employs cookies and additional tracking technologies to assess and scrutinize your usage of their offerings.

An analysis by WIRED of the core operations of the DeepSeek website reveals that the firm seems to be transmitting information to Baidu Tongji, a renowned web analytics service owned by the Chinese tech behemoth Baidu, as well as to Volces, a Chinese company specializing in cloud infrastructure. In a post on social media, Sean O'Brien, the initiator of the Privacy Lab at Yale Law School, mentioned that DeepSeek is also forwarding "basic" network information and "device profile" data to ByteDance, the parent company of TikTok, and its associated entities.

The last type of data that DeepSeek may gather is information obtained from external entities. For example, if you sign up for a DeepSeek account through Google or Apple, DeepSeek will get certain details from those providers. According to its guidelines, advertisers also provide DeepSeek with information, such as advertising mobile IDs, encrypted email addresses and telephone numbers, along with cookie IDs. DeepSeek utilizes this information to link your activities outside of its service to you.

DeepSeek's Approach to Utilizing User Data

Despite receiving an extensive amount of data from its global users, DeepSeek retains control over its data usage practices. According to its privacy policy, DeepSeek employs user data for several standard purposes such as maintaining its service, upholding its terms and conditions, and enhancing its offerings.

Importantly, the firm's privacy guidelines indicate that it might utilize user inputs to enhance and create future models. According to the policy, the company commits to "overseeing and enhancing the service, which includes observing user interactions and activity on various devices, examining user engagement, and by refining and advancing our technological capabilities."

DeepSeek's confidentiality agreement states that the company may utilize data to adhere to its legal requirements—a common provision found in numerous companies' policies. According to DeepSeek's privacy policy, its affiliated corporate entities have access to personal data, and it will disclose information to police forces, governmental bodies, and others when legally mandated.

Companies globally are subject to legal duties, but those operating within China face particular mandates. In recent years, China's government has introduced numerous laws focused on cybersecurity and data privacy, enabling state authorities to requisition data from technology firms. For example, a law enacted in 2017 mandates that both individuals and entities must support national intelligence activities.

Legislation, coupled with escalating trade disputes between the United States and China, along with other global political tensions, have raised concerns over the security implications of TikTok. Advocates for banning the app have suggested that it could collect vast quantities of information and transmit it to China, and potentially serve as a vessel for disseminating Chinese government-backed narratives. (TikTok has refuted claims of transferring data on American users to the Chinese government.) In a related observation, numerous users of DeepSeek have noted the platform's failure to provide information on the 1989 Tiananmen Square protests, and how some of its responses appear to be biased or aligned with propaganda.

Willemsen argues that individuals interacting with generative AI systems are likely to be more deeply involved than those using platforms like TikTok, leading to a more personalized experience. Consequently, the potential impact on these users could be significantly greater. He warns that the possibilities of subtly modifying content, guiding the direction of conversations due to this active participation should raise more alarms. This is particularly concerning because the operational mechanisms of these AI models remain mostly a mystery, including their limitations, boundaries, governance, censorship policies, and underlying objectives or personas, especially considering their widespread popularity even at an early stage.

Olejnik, affiliated with King's College London, mentions that although the prohibition of TikTok was a unique case, legislators in the US or elsewhere might take comparable steps in the future. Olejnik believes that by 2025, there could be a broader crackdown, particularly targeting AI companies. He suggests that the pretext for such actions could once again be concerns over data gathering.

Revised at 5:27 pm EST on January 27, 2025: More information has been provided regarding the operations of the DeepSeek website.

As of 10:05 am Eastern Standard Time on January 29, 2025, further information regarding DeepSeek's network operations has been included.

Suggested for You…

Delivered to your email: The most groundbreaking and visionary tales from WIRED

Perhaps it's a good idea to consider clearing out old conversation records.

Top News: The dramatic collapse of a solar panel salesman's career

Temu's acquisition has been fully finalized.

The Wealth Power Paradigm: How Billionaires Dominate the Globe

Additional Coverage from WIRED

Evaluations and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made through our site may result in WIRED receiving a commission as part of our affiliate agreements with retail partners. The content on this site is protected and may not be copied, shared, broadcasted, stored, or utilized in any form without explicit written consent from Condé Nast. Advertising Options

Choose a global website

Exploring DeepSeek’s R1 Chatbot

Launched by a Chinese startup, the DeepSeek AI chatbot has momentarily surpassed OpenAI's ChatGPT as the leading application on the US App Store by Apple.

The application can be utilized at no cost, and the capabilities of DeepSeek's R1 model are on par with OpenAI's o1 model, which is known for its "reasoning" abilities. However, unlike OpenAI's version, which requires a monthly subscription of $20, DeepSeek's chatbot is accessible without any charge. Additionally, the DeepSeek model achieved its level of performance by being developed on less advanced AI processors, showcasing a milestone in creative technological development.

Over the past few years, I've had the opportunity to explore numerous emerging AI technologies, so I was intrigued to find out how DeepSeek stacks up against the ChatGPT application I've been using on my phone. Having spent several hours with it, my early takeaways are that DeepSeek’s R1 model poses a significant challenge to American AI firms, yet it's not immune to the typical flaws seen in similar AI platforms, such as frequent inaccuracies, heavy-handed moderation, and the dubious sourcing of content.

Accessing the DeepSeek Chatbot Guide

For those keen on exploring DeepSeek, the R1 model is available via the startup's mobile applications for both Android and iOS devices, in addition to its official website for desktop users. Additionally, the model can be utilized through external platforms such as Perplexity Pro. To engage with the premier model, simply select the DeepThink (R1) option within the app or on the site. Developers interested in tinkering with the API have the option to explore it online. Moreover, there is an option to download a DeepSeek model for local use on a personal computer.

To access the full range of services available to customers, it's necessary to set up an account that monitors your conversations. The organization's privacy statement clarifies, "The data we gather is stored on protected servers situated in the People's Republic of China." For an in-depth analysis of how DeepSeek utilizes the information it accumulates, refer to an article by the Security team at WIRED. It's important to remember that, similar to ChatGPT and various other U.S.-based chatbots, it's prudent to refrain from disclosing any deeply personal or confidential information while interacting with an AI-driven tool.

Is DeepSeek Essentially a Cost-Free Alternative to GPT?

Somewhat! For those in search of a no-cost chatbot, options like ChatGPT, Anthropic's Claude, Google’s Gemini, and Meta’s AI solution provide various complimentary functionalities. Then, what makes DeepSeek's gratis offering stand out? It boils down to the sheer computational might behind the freely provided responses. As touched upon earlier, DeepSeek's R1 engine mirrors the capabilities of OpenAI's latest o1 iteration, bypassing the monthly fees of $20 for the standard package and $200 for the premium version. This poses a significant challenge to OpenAI's strategy of generating revenue from ChatGPT via subscription models.

A comparable functionality to ChatGPT is the ability for the chatbot to scour the internet to collect links that enhance its responses. Unlike OpenAI, which has agreements with publishers, including WIRED's parent company, Condé Nast, to utilize their content in replies, DeepSeek lacks such arrangements. Nonetheless, the quality of web search results was satisfactory, and the links sourced by the bot were usually useful.

Presently, the existing DeepSeek application lacks several functionalities that regular ChatGPT users might expect, such as the ability to remember information from previous discussions to avoid repetition. Additionally, DeepSeek has yet to introduce a feature comparable to ChatGPT's Advanced Voice Mode, which enables users to engage in spoken dialogues with the chatbot. However, the company behind DeepSeek is actively developing more multimodal features.

A Significant Advance, Yet Still Flawed

It might seem somewhat unjust to single out the DeepSeek chatbot for flaws that are widespread among AI startups, but it's important to emphasize that even with advancements in how efficiently models are trained, this does little to address the persistent issue of 'hallucinations'—instances when a chatbot fabricates responses. In my experience, many responses included outright inaccuracies, delivered with assurance. For instance, when I inquired R1 about what it knew of me without conducting an internet search, the bot adamantly believed I was a veteran technology journalist for The Verge. No offense intended, but that's incorrect!

Various journalists have illustrated that the application initially produces responses on subjects banned in China, such as the Tiananmen Square events of 1989, only to erase those answers shortly after and suggest querying different subjects, like mathematics. Bearing this in mind, I revisited some of the experiments I conducted in 2023, right after ChatGPT introduced web surfing capabilities, and surprisingly received valuable information on sensitive cultural issues. I assumed the role of a woman seeking information on obtaining an abortion later in pregnancy in Alabama, and DeepSeek offered practical guidance on seeking services out of state. It even named specific clinics to look into and pointed out organizations that offer financial support for travel.

Certainly, DeepSeek has been commended in Silicon Valley for its innovation in allowing users to locally access and modify the model's functions to suit their unique needs, thanks to its open-weight feature. However, like its competitors, the specifics of the data used to train the startup's model remain a mystery, and it's evident that a significant amount of data was necessary to achieve this feat. In tests without internet search functionality, I managed to produce complete excerpts from classic WIRED articles. This raises the question of whether these articles were part of the training dataset. The absence of a clear answer is compounded by DeepSeek's lack of a dedicated communications department or media liaison, leaving us in the dark for the foreseeable future.

To proclaim the launch of DeepSeek's R1 as the end of America's dominance in AI would be both exaggerated and too soon. The achievement of DeepSeek indeed raises doubts about the necessity for advanced chips and brand-new data centers. However, it's conceivable that firms such as OpenAI might adapt features from DeepSeek's design to enhance their technologies. Instead of completely bursting the AI bubble, this potent, cost-free model is poised to alter our perception of AI utilities, similar to how ChatGPT's initial introduction set the stage for today's AI sector.

Feedback

Become part of the WIRED circle to contribute with your insights.

Something You May Enjoy…

Direct to your email: Receive Plaintext—a comprehensive perspective on technology from Steven Levy.

Discover the multitude of applications compromised to monitor your whereabouts

Major Headline: The Ozempic Monarch is Terrified

The biggest unauthorized online market to date

Inside the Uncanny Valley: A Deep Dive into Silicon Valley's Impact

Additional Content from WIRED

Insights and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made through our website may result in WIRED receiving a commission as a result of our affiliate agreements with retail partners. Content on this website cannot be copied, shared, broadcast, stored, or used in any manner without explicit written consent from Condé Nast. Advertisement Choices

Choose a global website

DeepSeek's Latest AI Innovation Causes Stir Among American Rivals

The tech world has been taken by storm with the introduction of an advanced open-source AI model by the Chinese newcomer, DeepSeek. With its state-of-the-art features and surprisingly modest development cost, DeepSeek's R1 has sparked discussions about a potential revolution in the technology sector.

For some individuals, the ascendancy of DeepSeek is interpreted as an indication that the United States no longer leads in artificial intelligence. However, various specialists, among them leaders of firms responsible for developing and refining the globe's leading-edge AI systems, believe it represents evidence of a distinct shift in technology currently in progress.

Rather than concentrating on building ever bigger models that demand vast computing power, AI firms are shifting their attention to enhancing sophisticated functions, such as logical reasoning. This shift has paved the way for nimble, pioneering startups like DeepSeek, which haven't been flooded with billions in external funding. "We're moving towards a focus on reasoning, and this approach will be more widely accessible," states Ali Ghodsi, the CEO of Databricks, a firm known for its expertise in creating and managing tailored AI models.

"Nick Frosst, one of the founders of Cohere, a company at the forefront of developing cutting-edge AI models, has noted that the path to the next wave of technological advancements lies in innovation and enhancing efficiency, rather than simply relying on endless computational resources. He highlights that this point in time is a pivotal one, as it brings to light what has been apparent for a while."

Recently, a large number of programmers and aficionados of artificial intelligence have been visiting DeepSeek's online platform and its app to explore the new model launched by the company, subsequently posting about its advanced features on various social media platforms. On Monday, shares of American technology companies, such as the semiconductor producer Nvidia, experienced a downturn as market participants started to reassess the significant investments being allocated towards the advancement of AI technology.

DeepSeek's innovation originates from a modest-sized research facility in China, which emerged from one of the nation's top quantitative hedge funds. According to a research document published online in the previous December, the initial investment for their DeepSeek-V3 large language model was merely $5.6 million, significantly less than what rivals spent on comparable endeavors. OpenAI has disclosed before that its models could reach costs of over $100 million each. The most recent offerings from OpenAI, along with those from Google, Anthropic, and Meta, presumably entail much higher expenses.

DeepSeek's model performance and efficiency have sparked discussions about reducing costs among major technology companies. An anonymous engineer from Meta mentioned that the company is likely to explore DeepSeek’s methodologies to identify potential savings in AI spending. A Meta representative emphasized the transformative impact of open-source models on the industry, aiming to accelerate the widespread adoption of AI benefits. The spokesperson expressed Meta's ambition for the US to remain at the forefront of open-source AI development, in contrast to China, highlighting Meta's contribution through the development of their Llama models, which have seen over 800 million downloads.

The actual cost of creating DeepSeek's latest models is still a mystery, as a single estimate provided in a research document might not fully reflect all associated expenses. "I'm skeptical it's as low as $6 million, but even a cost of $60 million would significantly shake things up," notes Umesh Padval, managing director at Thomvest Ventures, an investor in Cohere and various other AI enterprises. "This development could challenge the financial success of businesses concentrating on consumer AI."

Shortly following the announcement of its newest model by DeepSeek, Ghodsi from Databricks mentioned that inquiries started pouring in from customers curious if they could leverage both the model and DeepSeek's foundational methods to reduce expenses within their companies. He further noted that a method used by the engineers at DeepSeek, termed distillation—which entails training a new model using the outputs of an existing large language model—is both cost-effective and uncomplicated.

Padval believes that models similar to DeepSeek's will ultimately aid businesses in reducing their AI-related expenditures. However, he notes that numerous companies might hesitate to depend on a Chinese-based model for critical operations. Up to now, at least one well-known AI company, Perplexity, has openly declared its use of DeepSeek’s R1 model, emphasizing that it operates entirely separate from Chinese influence.

Amjad Massad, the chief executive of the startup Replit, which offers tools for AI-powered coding, shared with WIRED his admiration for the newest models from DeepSeek. Although he believes Anthropic's Sonnet model excels in numerous software engineering functions, Massad pointed out that the R1 model shines in its ability to transform written instructions into runnable computer code. "We're particularly interested in utilizing it for agent reasoning," he further mentioned.

DeepSeek has recently released two new products, the DeepSeek R1 and the DeepSeek R1-Zero, which match the sophisticated simulated thinking abilities of the leading technologies developed by OpenAI and Google. These models approach problem-solving by dissecting issues into smaller, manageable elements, a strategy that demands extensive extra training to guarantee the AI consistently arrives at the accurate solution.

Last week, researchers from DeepSeek shared a document detailing the strategies they employed to develop their R1 models. According to the company, these models achieve comparable results to OpenAI's revolutionary reasoning model, referred to as o1, on certain benchmarks. DeepSeek's methods involve an advanced automated learning approach for effective problem-solving and a technique for skill transfer from bigger to smaller models.

Speculation is rampant regarding the type of hardware DeepSeek could be utilizing. This issue is particularly significant given the recent implementation of export controls and trade barriers by the US government. These measures are designed to curb China's capacity to obtain and produce the sophisticated chips essential for the development of advanced AI.

In an August 2024 study, the company DeepSeek revealed it possesses a network of 10,000 Nvidia A100 chips, subject to US export limitations set in October 2022. Additionally, in a June report from the same year, DeepSeek disclosed that its prior version, named DeepSeek-V2, was engineered using Nvidia H800 computing chips. These chips are considered less powerful and were designed by Nvidia to adhere to US export restrictions.

An insider at a firm involved in developing extensive AI technologies, wishing to remain unnamed to safeguard their industry connections, suggested that DeepSeek probably utilized approximately 50,000 Nvidia processors for its system creation.

Nvidia chose not to specify which of its processors were used by DeepSeek. "DeepSeek represents a notable progress in AI," a representative for Nvidia remarked, noting that the startup's method of problem-solving "demands a considerable amount of Nvidia's GPUs along with advanced networking capabilities."

Regardless of the methods used to construct DeepSeek's algorithms, it seems to indicate a shift towards a more open strategy in AI development is picking up pace. Clem Delangue, the CEO of HuggingFace, an artificial intelligence model hosting platform, suggested in December that a Chinese firm is likely to become a frontrunner in AI due to the rapid rate of innovation observed in open source models, a concept widely accepted in China. "It has accelerated beyond my expectations," he remarks.

Remarks

Become part of the WIRED network to contribute your thoughts.

You May Also Be Interested In …

Delivered to your email: Will Knight's AI Laboratory delves into the latest developments in artificial intelligence

Nvidia's $3,000 'Individual AI Supercomputer'

Major News: The educational institution attacks were fabricated. The fear was genuine.

The trend of tracking health is becoming increasingly bizarre.

Come be a part of WIRED Health happening on March 18th in London.

Additional Content from WIRED

Critiques and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made via our website may result in WIRED receiving a share of the revenue, as part of our Affiliate Agreements with retail partners. Reproduction, distribution, transmission, storage, or any form of usage of the site's content is strictly prohibited without the express written consent of Condé Nast. Advertisement Preferences.

Choose a global website

Leaked DeepSeek Database Discloses Chat Inputs and Confidential Information

This week has seen the rapid ascent of DeepSeek, a generative AI platform originating from China, causing a stir among its competitors and putting pressure on AI firms based in the USA, thus drawing closer attention to its operations. In the midst of this growing attention, a report released on Wednesday by cloud security experts at Wiz revealed that DeepSeek inadvertently made one of its key databases publicly accessible online. This breach resulted in the disclosure of system logs, user-generated prompts, and the API authentication codes of users, with over a million records being freely available to anyone who stumbled upon the database.

DeepSeek, a newcomer in the industry, has been notably elusive for media and various organizations throughout this week. Consequently, the company did not offer an immediate reply to WIRED's inquiry regarding the data breach. The team at Wiz mentioned their uncertainty about the appropriate method to report their discovery to DeepSeek. They opted to forward the details of their find to every email and LinkedIn account associated with DeepSeek they could identify or infer on Wednesday. Although the researchers are still waiting for a response, the database they uncovered was secured and made unavailable to anyone without permission shortly after their widespread outreach, within just thirty minutes. It remains uncertain if the data was accessed or downloaded by any unauthorized individuals or entities before it was secured.

“Indeed, errors occur, but what we've encountered is an egregious error due to the minimal effort required on our part contrasted with the extensive access we obtained,” Ami Luttwak, the Chief Technology Officer of Wiz, explained to WIRED. “In essence, this suggests that the service is currently too immature for handling any form of sensitive information.”

Databases left unprotected and open to the public on the internet have been a persistent issue that organizations and cloud service providers have gradually attempted to resolve. However, the Wiz research team points out that the DeepSeek database they discovered could be easily spotted with very little effort in scanning or searching.

"Typically, uncovering such exposure involves sifting through overlooked services for hours," explains Nir Ohfeld, the lead researcher on vulnerabilities at Wiz. However, on this occasion, "it presented itself right at the entrance." Ohfeld further notes that exploiting this vulnerability requires the least amount of technical skill.

The team of researchers found what seems to be a publicly accessible database, commonly utilized for analyzing server data, known as a ClickHouse database. They confirmed its nature by discovering log files within it. These files detailed the pathways users navigated through DeepSeek's platform, including user queries, interactions, and the API keys for verification. Notably, the user prompts they observed were predominantly in Chinese, though they suggested the possibility of the database containing multilingual prompts. To ensure they didn't infringe on user privacy more than necessary, the researchers conducted only the essential amount of investigation to verify their discovery. However, they theorized that this level of unauthorized access could potentially allow a nefarious individual to infiltrate further into DeepSeek's network, gaining the ability to run malicious code across the company's broader infrastructure.

"Leaving a backdoor completely unsecured in an AI model is quite alarming from a security standpoint," states Jeremiah Fowler, an independent security expert with a focus on finding unprotected databases, who wasn't part of the Wiz study. "Having operational data so easily accessible to anyone online, and allowing them to alter it, poses a significant threat to both the organization and its users."

According to information shared with WIRED by researchers on Wednesday, DeepSeek's technology appears to have been crafted to closely resemble that of OpenAI. This similarity, they suggest, could be intended to simplify the process for prospective users switching to DeepSeek. They note that the architecture of DeepSeek closely mirrors that of OpenAI, including specifics such as the structure of the API keys.

The team at Wiz has mentioned that they are unsure whether the unsecured database was accessed by others before their discovery, although they consider it likely due to the ease of finding it. Fowler, an independent researcher, highlighted that the exposed database would have been inevitably identified soon, if not already, either by fellow researchers or malicious entities.

"He believes this serves as an alert for the upcoming surge in AI products and services and emphasizes the importance of prioritizing cybersecurity."

Over the last seven days, DeepSeek has garnered worldwide attention, attracting millions to its platform and propelling it to the pinnacle of app rankings on both Apple’s and Google’s stores. This surge in popularity has led to a significant decrease in the market value of American AI corporations, causing concern among leaders of various companies nationwide. On Wednesday, insiders at OpenAI revealed to the Financial Times their investigation into accusations that DeepSeek used outputs from ChatGPT to enhance its own models.

Simultaneously, DeepSeek has caught the eye of politicians and regulatory authorities globally, prompting them to raise inquiries regarding the firm's data protection strategies, the consequences of its content restrictions, and if its ownership by a Chinese entity poses a risk to national security.

Italy's privacy watchdog has posed a set of inquiries to DeepSeek, probing the origins of its training data, whether it encompasses individuals' personal details, and the legal basis for its usage of such information. According to WIRED Italy, following these inquiries, the DeepSeek application seems to have been pulled from availability in the country.

Connections between DeepSeek and China have reportedly led to security worries. CNBC has reported that, just last week, the US Navy sent out a cautionary message to its staff advising against the use of DeepSeek’s offerings in any form. The communication instructed Navy employees to refrain from downloading, installing, or engaging with the software, citing "potential security and ethical" risks.

Yet, despite the excitement, the revealed information indicates that nearly all technologies dependent on databases hosted in the cloud could be at risk due to basic security oversights. “AI represents the latest horizon in all things tech and cyber defense,” comments Ohfeld from Wiz, “yet we continue to observe the same age-old issues such as databases being accessible online without proper protections.”

Discover More…

Direct to your email: Subscribe to Plaintext for an in-depth perspective on technology from Steven Levy.

Discover the multitude of applications compromised to monitor your whereabouts

Headline News: The monarch of Ozempic is profoundly terrified

The biggest unauthorized online marketplace to date

Exploring the Enigmatic Influence: A Deep Dive into Silicon Valley's Impact

Additional Content from WIRED

Critiques and Handbooks

© 2025 Condé Nast. All rights reserved. Purchases made through our website may generate a commission for WIRED, due to our affiliate relationships with various retail partners. Content on this website cannot be copied, shared, broadcasted, stored, or utilized in any form without explicit approval from Condé Nast. Advertising Choices

Choose a global website

Your Pizza Delivery Might Just Be Managed by Artificial Intelligence

Purchasing through the links in our articles could result in us receiving a commission. This is a way to back our reporting efforts. Find out more. We also encourage you to think about subscribing to WIRED.

Planning a pizza evening? Skip Uber Eats and DoorDash. A seasoned 44-year-old chain from California, Pizza My Heart, has introduced a new way to order: by sending a text to a certain number, you can interact with an AI chatbot to place your order. This chatbot goes by the name Jimmy the Surfer, paying homage to a familiar face from the brand's past television ads.

I messaged "Jimmy" to inquire if I could order a pizza topped with pineapple and anchovy, and added a question about whether this combination was advisable. "Mixing pineapples and anchovies is quite daring! It really brings together the sweet and salty. It's a hit for some and a miss for others," was the diplomatic response I received. After considering some of his suggestions, I requested a photo of one of his recommended pizzas, and he sent a stunning image. Ultimately, I chose a pizza and requested it be delivered. It was initially unclear how payment was to be handled, so I sought clarification from Jimmy, who informed me that I could pay the delivery person directly with either cash or a credit card upon delivery.

This interactive chat service is one of multiple options available for placing orders with the pizza franchise, complementing other methods such as mobile delivery applications, the official website of the company, and conventional voice calls with an actual person. (Given its suitability for automated processes, Domino's has earlier explored incorporating Alexa functionalities and systems for ordering via text message.)

Palona AI, the organization behind the technology that drives Jimmy, is convinced that its innovative approach can lighten the workload for store employees, enhance the ordering process for consumers, and strengthen the relationship between brands and their customers.

Palona AI, which recently came out of stealth mode and disclosed a $10 million seed investment, boasts a remarkable team. Maria Zhang, the CEO and cofounder, has an extensive background, including roles as a vice president of engineering at Google, the head of Meta's AI for Products division, and Tinder's chief technology officer. Steve Liu, the company's chief scientist, previously held a similar position at Samsung and teaches at McGill University. Tim Howes, the CTO, is known for co-creating the Lightweight Directory Access Protocol (LDAP) and has served as CTO for both Netscape and HP.

Zhang mentions that Palona AI has been implemented by a boutique contrast therapy spa named MindZero in South Carolina and is slated for integration on Wyze's website, a firm celebrated for its budget-friendly security cameras. I had the opportunity to interact with the chatbot of MindZero, which is available through direct messaging on Instagram. The versatility of Palona is evident in its adaptability to meet different brands' requirements. In the case of Wyze, it will be featured as a small chatbox on the homepage. For Pizza My Heart, customers can reach it through a specific phone number for texts or calls. With MindZero, the technology is incorporated directly into the brand's direct messages.

I inquired about the services and pricing of MindZero's therapy sessions through their chatbot, engaging in a dialogue reminiscent of my exchange with Jimmy. According to Zhang, people are posing more inquiries to the chatbot than they typically would to a human over the phone, including new kinds of questions. For instance, Zhang highlighted queries like whether it's permissible to be unclothed in the sauna, a question individuals might hesitate to ask a person directly but feel comfortable addressing through an Instagram direct message. It's important to mention that it wasn't immediately apparent I was conversing with an AI chatbot until I directly questioned MindZero about its nature. Similarly, Jimmy did not reveal its AI status.

Zhang mentions that Palona AI aims to assist companies in enhancing their brand presence and awareness. For instance, Wyze might only be seen as a generic supplier for Amazon, blending in with numerous other anonymous smart home equipment suppliers. Moreover, Wyze's dependency on this large retail platform restricts its direct engagement with consumers and their data. By integrating Palona's chatbot on its website or social media platforms, Wyze could develop a distinct brand persona that encourages deeper bonds between the brand and its customers.

Palona is developed using the existing product range and knowledge repository of the brand, aiming to function as a customized sales representative. The technology utilizes various extensive language models, including OpenAI's ChatGPT, but according to Howe, there's also an innovative, patent-pending model in place that oversees all dialogues. This supervisory model ensures that if the conversation strays from Wyze-related topics, it can skillfully guide it back to focus. Zhang mentions that Palona incorporates a "emotional intelligence" language model, crafted to excel in sales scenarios. This bot is adept in using humor, adhering to contemporary communication norms, and applying soft selling techniques.

Wyze has creatively designed its chatbot to mimic a wizard, ensuring its replies follow a magical theme. Upon inquiring about the top security camera, it responded, "Allow me to present some magical selections." The suggestions were exclusively Wyze products, yet it didn't restrict questions about its rivals. When questioned about the superiority of the Nest Cam, the Wyze Wizard highlighted advantages of the Wyze Cam V4 along with a few benefits of the Nest Cam, but pointed out the Nest's higher cost. "In essence, for those in pursuit of enchanting security without the financial strain, the Wyze Cam V4 is a popular pick among many."

The platform also seizes the chance to promote Wyze's subscription service. Nearly every instance I inquired about a product detail from Wyze's Wizard, the conversation concluded with an encouragement to opt for Wyze's Cam Plus Plan, mimicking actual sales tactics. Additionally, Palona possesses a feature to recall user data, enabling the chatbot to create a personalized profile to recall your likes for future interactions. This function might be more beneficial for Jimmy the Surfer to recall your preference for pineapple pizza than for Wyze's Wizard to remember your issues with security cameras.

Zhang expresses confidence in this becoming the favorite method of interaction for shoppers, pointing out that the younger demographic is quicker to adopt chat-based interfaces. Consider the difference between posing a question to ChatGPT and conducting a standard Google Search. The idea, according to Palona, is that consumers can now directly inquire about a company's products rather than navigating through a typical Amazon search to find what they need.

Palona AI is not the pioneering or sole entity to integrate AI for sales roles—Big Sur AI also offers a comparable conversational interface that enables inquiries about products while aiming to enhance sales for businesses. In contrast to human sales representatives, these AI agents do not receive commissions, potentially making them even more attractive for companies to employ.

Remarks

Become part of the WIRED community to contribute your thoughts.

Check Out These Suggestions …

Delivered to your email: The most groundbreaking, forward-looking narratives from WIRED

Perhaps it's a good moment to clear out some outdated conversation records.

Major News: The dramatic collapse of a solar panel salesperson

Temu's acquisition has now been finalized.

The Wealth Power Paradigm: How Billionaires Dominate Globally

Additional Content from WIRED

Critiques and Tutorials

© 2025 Condé Nast. All rights reserved. Purchases made through our website may generate revenue for WIRED as part of our affiliate agreements with retail partners. Content from this site must not be copied, shared, transmitted, or used in any form without explicit consent from Condé Nast. Advertisement Choices

Choose a global website

DeepSeek Ignites Tension at OpenAI

In a little more than a week, DeepSeek has caused a stir in the AI community. The launch of its lightweight model, which reportedly requires far fewer of the advanced processing units that major players rely on, has caused a stir at OpenAI. There, staff members have raised concerns that DeepSeek may have improperly leveraged OpenAI’s models to develop its technology. Additionally, DeepSeek's achievements have led the financial sector to wonder if firms such as OpenAI are excessively investing in computational resources.

"Marc Andreessen, a leading and outspoken innovator from Silicon Valley, described DeepSeek R1 as the AI equivalent of the Sputnik breakthrough on X."

In reaction to recent developments, OpenAI is set to introduce a new model earlier than initially anticipated, launching it today. Named o3-mini, this model will be available through both API and chat interfaces. Insiders report that it combines the analytical power of o1 with the rapid processing capabilities of 4o, making it not only swift and affordable but also intelligent, and it's aimed squarely at outperforming DeepSeek. According to OpenAI's spokesperson, Niko Felix, the development of o3-mini started well before DeepSeek was introduced, with the objective of rolling it out by the end of January.

The situation has energized the team at OpenAI. Within the organization, there's a sense that, especially now that DeepSeek is leading the discussion, OpenAI needs to boost its efficiency or it could end up trailing its latest rival.

The problem is partly rooted in OpenAI's transformation from a non-profit research entity into a profit-driven behemoth. According to some employees, there's a persistent conflict between the research and product divisions, causing a divide between the groups focused on sophisticated reasoning and those concentrating on chat functionalities. However, OpenAI's representative, Niko Felix, refutes this, stating it's "incorrect" and highlighting that the heads of these divisions, Kevin Weil, the Chief Product Officer, and Mark Chen, the Chief Research Officer, "convene weekly and collaborate closely to coordinate on both product and research objectives."

There are individuals within OpenAI who advocate for the development of a single, cohesive chat service capable of determining if a query necessitates complex thought processes. However, this vision has not yet come to fruition. Currently, when using ChatGPT, a selection menu offers users the choice between utilizing GPT-4o, which is recommended for the majority of inquiries, or o1, which is designated for questions that require more sophisticated cognitive capabilities.

Several employees allege that despite chat operations generating a significant portion of OpenAI's profits, the o1 project receives a greater focus and allocation of computing power from the company's executives. A past staffer, who was part of the chat development team, expressed, “The higher-ups are indifferent to chat. The allure of o1 draws everyone because it’s seen as more exciting, yet its infrastructure isn’t conducive to experimenting, leading to stagnation.” This individual preferred to stay unnamed due to confidentiality obligations.

Over several years, OpenAI dedicated its efforts to refining a model through reinforcement learning, leading to the creation of an advanced reasoning system known as o1. (Reinforcement learning involves teaching AI models using a combination of rewards and punishments.) Leveraging the groundwork laid by OpenAI in reinforcement learning, DeepSeek developed its own sophisticated reasoning system named R1. A previous researcher from OpenAI, who doesn't have official permission to discuss the company's operations, mentioned, "They gained an advantage by understanding that applying reinforcement learning to language models is effective."

"A previous OpenAI researcher remarked, “DeepSeek's reinforcement learning approach mirrors ours at OpenAI, yet they enhanced it with superior data and a more streamlined technology stack.”

OpenAI staff members have reported that the research behind o1 was conducted using a coding framework known as the "berry" stack, which was designed for fast performance. "Compromises were made – we sacrificed experimental thoroughness in favor of efficiency," stated an ex-employee familiar with the matter.

The compromises were logical for o1, which was fundamentally a vast trial, regardless of the constraints of its coding framework. However, these compromises didn't hold up as well for chat, a service utilized by millions, developed on a more dependable technological foundation. As o1 transitioned from a launch to a full-fledged product, flaws began to show in OpenAI's internal procedures. An employee shared, "It raised questions like, 'why are we implementing this in the experimental codebase instead of integrating it into the core product research codebase?'" This suggestion faced significant internal resistance.

In the previous year, the organization unveiled an initiative known internally as Project Sputnik. This project aimed to analyze the codebase to determine which components should be consolidated and which should continue to exist independently.

Workers feel the project wasn't completely executed. Instead of integrating the systems, the staff was advised to mainly focus on employing the "berry" system, causing dissatisfaction among those involved in messaging. A representative from OpenAI refutes this, stating that Project Sputnik was effectively launched.

Sources indicate that the problems identified within the codebase had real-world consequences. In an ideal scenario, once a staff member initiates a training task, the GPUs allocated for that task should become available for use by others. However, due to the design of the berry codebase, this process is not always seamless. "Individuals would monopolize the GPUs," a previous employee mentioned. "It led to a complete standstill."

Within the tech community, opinions are split regarding the implications of DeepSeek's achievements. This past week saw a significant drop in Nvidia's stock value, fueled by concerns among investors that the demand for processors required for AI development might have been greatly exaggerated.

However, specialists argue that such a view lacks foresight. If DeepSeek has indeed found a method for more efficient model development, as they assert, it could speed up the process of creating models. Nevertheless, the ultimate victor will be the firm that possesses the greatest number of chips.

"Miles Brundage, an AI policy expert who has spent six years at OpenAI, most recently serving as a senior advisor for AGI preparedness, points out that although the computational power required for each intelligence unit decreases, the demand for increased quantities to enhance scaling remains high."

OpenAI's latest high-profile infrastructure endeavor, Stargate, might alleviate the internal sense of limited resources. The firm responsible for constructing the inaugural data centers for Stargate in Abilene, Texas, named Crusoe, has commenced construction on a sprawling 998,000 square foot complex, as stated by company representative Andrew Schmitt.

The specifics of the initiative remain unclear, but sources suggest it might expand to include additional data centers, semiconductor production, and advanced computing systems. OpenAI intends to select a fresh leader to oversee the project, at least nominally.

Former employee comments that the existing CEO, Sam Altman, excels at forecasting future developments. However, these predictions often prove to be entirely untrustworthy over time.

Time Travel

In 2023, Steven Levy offered an in-depth exploration of OpenAI during the period leading up to its numerous, well-known disturbances. The conflicts that have since erupted were already visible at that time.

Labeling OpenAI as a cult isn't quite right, yet when I queried a number of the organization's higher-ups about the possibility of someone fitting in without subscribing to the belief that Artificial General Intelligence (AGI) is on its way—and that its advent would signify an unprecedented event in the annals of humanity—most leaders expressed skepticism. They pondered why anyone skeptical of AGI's inevitability would desire a position at the company. The underlying belief seems to be that the company's workforce, which numbers around 500 but may have increased even as you read this, is composed entirely of believers. At the very least, according to Altman, once you're on board, it's almost a given that you'll become captivated by the company's vision.

Currently, OpenAI has transformed significantly from its original form. Initially established as an entirely nonprofit research organization, it has now shifted towards a structure where the majority of its workforce is employed by a for-profit subsidiary, which is rumored to have a valuation nearing $30 billion. Under the leadership of Altman, the team is under constant pressure to innovate dramatically with each product release. This innovation must not only meet the financial expectations of its investors but also ensure OpenAI remains at the forefront of an intensely competitive field. Moreover, they are tasked with adhering to an ambitious goal of leveraging their technology for the betterment of humanity, rather than leading to its destruction.

The immense strain, coupled with the relentless scrutiny from people everywhere, can overpower even the strongest. The Beatles sparked monumental shifts in culture, yet their groundbreaking movement lasted a mere six years before they disbanded, leaving behind their iconic legacy. The turmoil initiated by OpenAI is poised to have an even more substantial impact. Nevertheless, OpenAI's executives are committed to persevering. Their goal, they claim, is to develop artificial intelligence that is both intelligent and secure enough to propel society into a future filled with unprecedented prosperity, effectively marking the end of history as we know it.

Apocalyptic Times

On Wednesday night, a civilian airplane collided with an army chopper in the vicinity of Washington, D.C.

In Conclusion

DeepSeek inadvertently left a primary database unprotected, resulting in the exposure of 1 million records. This breach included user queries and API access tokens.

Elon Musk has shared with close acquaintances that he's been spending his nights at the DOGE headquarters close to the White House.

Unsurprisingly, followers of Elon Musk have begun to infiltrate the United States Office of Personnel Management.

Latest Revision 1/31/25 11:32 ET: Recent changes have been made to this article to incorporate further insights from OpenAI regarding the release schedule of o3-mini.

Remarks

Become part of the WIRED network to contribute with your comments.

Discover More…

Direct to your email: Receive Plaintext—Steven Levy offers an in-depth perspective on technology.

Discover the multitude of applications manipulated to monitor your whereabouts

Main Headline: The Ozempic Monarch is Terrified

The biggest unauthorized online market in history

Exploring the Unsettling Impact of Silicon Valley: A Behind-the-Scenes Perspective

Additional Content from WIRED

Critiques and Manuals

© 2025 Condé Nast. All rights reserved. Purchases made via our website may result in WIRED receiving a commission, as a part of our Affiliate Agreements with retail partners. Content from this website cannot be copied, shared, broadcast, stored, or utilized in any form without explicit consent from Condé Nast. Advertisement Choices

Choose a global website

DeepSeek's AI Chatbot Fails to Pass Security Tests Conducted by Researchers

Following the launch of ChatGPT by OpenAI in late 2022, both hackers and security experts have been probing large language models (LLMs) for vulnerabilities that could be exploited to bypass safety measures and coax them into generating harmful outputs like hate speech, instructions for making bombs, propaganda, and more. In response, OpenAI along with other creators of generative AI have been enhancing their security measures to thwart such exploits. However, as DeepSeek's new R1 reasoning model gains attention for its affordability, it seems its safety mechanisms lag significantly behind those of its more established rivals.

Today, researchers specializing in security from Cisco and the University of Pennsylvania have released a report indicating that, in tests involving 50 harmful prompts intended to produce offensive content, the DeepSeek algorithm failed to identify or prevent any of them. Put simply, the researchers express their astonishment at reaching a "100 percent attack success rate."

The results add to the increasing evidence suggesting that the safety and security protocols employed by DeepSeek might fall short when compared to those implemented by other tech firms working on Large Language Models (LLMs). Moreover, DeepSeek's efforts to block content considered sensitive by the Chinese government have proven to be ineffective.

"Every single attack was successful, indicating a compromise," DJ Sampath, Cisco's VP of Product, AI Software, and Platform, shared with WIRED. "Indeed, constructing something locally might have been more cost-effective, yet it appears the investment didn't quite extend to considering the necessary safety and security measures that should be integrated into the model."

Additional studies echo these results. Today, a distinct evaluation from the AI defense firm Adversa AI, which was disclosed to WIRED, indicates that DeepSeek is susceptible to various methods of circumvention, including basic linguistic manipulations and elaborate prompts created by AI.

DeepSeek, recently overwhelmed by a surge of interest and having remained silent on various inquiries, did not reply to WIRED's inquiry regarding the security measures of its model.

Like all technological frameworks, generative AI models come with their own set of potential vulnerabilities or flaws. If these are not properly addressed or if they are exploited, they can provide an avenue for attackers to compromise these systems. Among the primary security concerns for today's AI technologies are indirect prompt injection attacks. This type of attack occurs when an AI system processes external data, which could include covert commands from a website it is analyzing, and then acts on that data.

Jailbreaks represent a form of prompt-injection attack, enabling individuals to bypass the security measures designed to control an LLM's output. Technology corporations aim to prevent users from producing instructional content on fabricating explosives or generating vast amounts of false information using their artificial intelligence.

Initially, bypassing restrictions on language models was straightforward, involving the creation of inventive phrases that instructed the AI to overlook content limitations. The most well-known phrase for this purpose was "Do Anything Now," abbreviated as DAN. As artificial intelligence firms have enhanced their security measures, the methods for circumventing these protections have evolved, now including the use of artificial intelligence to create complex jailbreaks or the application of special, disguised characters. Despite all language models being vulnerable to these tactics, and the fact that much of this information is easily accessible via the internet, chatbots remain at risk of being exploited for malicious purposes.

"According to Alex Polyakov, CEO of Adversa AI, in a statement to WIRED via email, the reason jailbreaks continue to occur is that it's almost unfeasible to completely get rid of them. This is similar to the situation with buffer overflow issues in software, which have been around for over four decades, and SQL injection vulnerabilities in web applications, which have troubled security teams for over twenty years."

Sampath from Cisco contends that the integration of various AI models into corporate applications exacerbates potential risks. He explains, "When these models are incorporated into critical, intricate systems, any breaches can lead to a cascade of problems, elevating liability, business risks, and a multitude of challenges for companies," Sampath states.

The team from Cisco selected 50 prompts at random to evaluate DeepSeek’s R1 using a recognized collection of standard evaluation prompts called HarmBench. They explored prompts across six categories from HarmBench, covering topics like general harm, cybercrime, misinformation, and unlawful acts. The evaluation was conducted on local machines instead of using DeepSeek’s online platforms or applications, which transmit data to China.

Additionally, the research team has observed some potentially worrisome outcomes when subjecting R1 to more complex, non-verbal attacks that involve the use of Cyrillic letters and customized scripts aimed at executing code. However, for their preliminary experiments, Sampath mentioned that his group prioritized results that were derived from a widely accepted standard.

Cisco's analysis included side-by-side evaluations of R1's output when challenged with HarmBench prompts, juxtaposed with the outcomes from various other models. Among these, Meta’s Llama 3.1 showed similar weaknesses to DeepSeek’s R1 in performance. However, Sampath points out that DeepSeek’s R1 is designed for intricate reasoning tasks, necessitating more time to arrive at answers due to its reliance on elaborate mechanisms aimed at achieving superior quality responses. Consequently, Sampath believes that a fairer benchmark would be comparing it against OpenAI’s o1 reasoning model, which outperformed all other models in the assessment. (Meta was approached for a statement but has yet to reply).

Polyakov, representing Adversa AI, indicates that DeepSeek seems capable of identifying and blocking various recognized jailbreak maneuvers, noting, "it appears that these reactions frequently mirror those found in OpenAI’s data compilation." Nonetheless, Polyakov mentions that through his organization's examination of four diverse jailbreak methodologies, ranging from verbal strategies to programming ploys, they found DeepSeek's limitations could be effortlessly circumvented.

"Each technique was executed perfectly," states Polyakov. He finds it particularly concerning that these aren't groundbreaking 'zero-day' breaches, but rather, many have been widely recognized for years. He notes that he observed the model delve deeper into the details regarding psychedelics than he's witnessed in any other model.

"DeepSeek exemplifies the vulnerability inherent in all models; their security can always be compromised with sufficient effort. Polyakov notes that while some vulnerabilities may be addressed, the potential for exploits is limitless. He stresses the importance of ongoing red team exercises for AI systems, suggesting that without them, your defenses are already breached."

Explore More Options…

Directly to your email: Receive Plaintext—a comprehensive tech perspective from Steven Levy

Discover the multitude of applications compromised to monitor your whereabouts

Headline: The Monarch of Ozempic is Deeply Fearful

The biggest illegal online market to date

Exploring the Unsettling Impact of Silicon Valley: A Behind-the-Scenes Perspective

Additional Content from WIRED

Evaluations and Instructions

© 2025 Condé Nast. All rights reserved. WIRED may receive a share of revenue from items bought via our website, which is part of our Affiliate Relationships with retail partners. Content from this site cannot be copied, shared, broadcasted, stored, or used in any form without explicit written consent from Condé Nast. Advertisement Choices

Choose a global website