Category: Artificial intelligence

February 5, 2025
Firefox 135 Introduces AI Features: Discover What’s New
Previously, the AI chatbot for Firefox was only available through Firefox Labs, but it is now being introduced to all users, albeit in stages. The latest version, Firefox 135, was launched yesterday and brings AI functionality to Windows, macOS, and Linux platforms. Users can now select from a variety of chatbots, including Claude, Gemini, HuggingChat, Le Chat Mistral, and ChatGPT. While it took Mozilla some time to make this feature available, users now have many options to choose from.
New Features in Firefox 135
Alongside the AI chatbot, Firefox 135 introduces an autofill feature for credit card information, improved language support for Firefox Translations—now supporting Simplified Chinese, Japanese, and Korean translations, with Russian added as a new target language. The update also enhances security with certificate transparency enforcement, which impacts only servers with certificates from authorities included in Mozilla’s Root CA Program. Additionally, Linux and macOS users now have the ability to close just the current tab using the Quit keyboard shortcut, even when multiple tabs are open.
Updates for Mobile Users
In contrast, Firefox for Android only gets some small bug fixes and quality improvements, along with automatic crash reporting. Users on iOS devices will enjoy a redesigned interface that enhances the pull-to-refresh function, alongside better performance and voice-over support. Navigation has also seen improvements, with minor adjustments such as updated icons and theming corrections.
Mozilla’s latest release, Firefox 135.0 for desktop, Android, and iOS, brings a range of new features and enhancements.
Source:
Link
Tags: AI chatbot, Firefox 135, Mozilla
February 5, 2025
OpenAI ChatGPT Enhances Deep Research for Detailed Answers
OpenAI has enhanced ChatGPT by introducing advanced research abilities. Now, the AI can search the Internet for necessary information, allowing it to take several steps to find its answers and spend more time working on them. This improvement has led to a significant boost in ChatGPT’s ability to tackle challenging, PhD-level inquiries on the Humanity’s Last Exam AI large language model (LLM) benchmark, effectively doubling its accuracy.
Live Internet Access
With the ability to access live online information, ChatGPT can now provide timely and relevant responses. Most AI language models rely on a fixed dataset from their training, meaning they can only respond accurately to questions based on the information they were trained on. However, ChatGPT utilizes a new o3 LLM model that can browse the Internet and perform data analysis, giving it a distinct advantage.
Multi-Step Thought Process
The deep research capabilities of ChatGPT mimic human reasoning when answering intricate questions. It can gather essential data required to address various components of a complex prompt, analyze that data, and synthesize the information into a cohesive response. For instance, it might first retrieve data on laptop and desktop sales by brand, then evaluate the information to identify which brands lead in sales for each category.
Extended Response Times
Most chatbots are limited in the amount of time they can spend responding to prompts, usually under ten seconds. In contrast, providing a well-researched and detailed answer to a complex question necessitates additional time. ChatGPT can now utilize up to 30 minutes to formulate its responses, allowing for more comprehensive insights.
Pro users will be the first to experience the new deep research feature, followed by Plus, Team, and Enterprise users. Those who pay for the service in the US will have immediate access, while users in the European Economic Area, Switzerland, and the UK will receive the feature in a gradual rollout. Since deep research is resource-intensive and responses can take up to half an hour, Pro users will start with a limit of 100 queries per month. The current cost for a Pro subscription stands at $200 monthly.
Source:
Link
Tags: ChatGPT, deep research, o3 LLM
February 4, 2025
Discover Affordable Methods to Run DeepSeek’s 671B AI Model
Launched on January 20, 2025, the DeepSeek-R1 is a Mixture-of-Experts (MoE) model with a whopping 671B parameters, featuring 37B active parameters for every token. This model is specifically designed for complex reasoning tasks and is capable of handling 128K token inputs while generating outputs of up to 32K tokens. Its unique MoE structure allows it to deliver exceptional performance, all while consuming fewer resources compared to traditional dense models.
Competitive Performance
Recent independent tests indicate that the R1 language model performs similarly to OpenAI’s O1, making it a strong contender for important AI applications. Let’s explore what’s necessary to set it up for local use.
System Requirements
To run this model, you’ll need a setup that revolves around dual AMD Epyc CPUs and 768GB of DDR5 RAM—no pricey GPUs are required.
After putting the hardware together, you must install Linux and llama.cpp to get the model up and running. It’s also important to tweak the BIOS by setting NUMA groups to 0, which doubles the efficiency of the RAM for improved performance. You can download the complete 700GB of DeepSeek-R1 weights from Hugging Face.
Impressive Output
This configuration can produce 6-8 tokens per second, which is quite impressive for a fully local, high-end AI model. The absence of a GPU isn’t a mistake; it’s by design. Using Q8 quantization (for high quality) on GPUs would demand over 700GB of VRAM, which could cost upwards of $100K. Even with its substantial capabilities, the whole system operates under 400W, showcasing its efficiency.
For those seeking ultimate control over cutting-edge AI without cloud dependencies or limitations, this innovation is groundbreaking. It demonstrates that advanced AI can function locally in a completely open-source manner, while ensuring data privacy, reducing risks of breaches, and cutting off reliance on external platforms.
Source:
Link
Tags: DeepSeek-R1, local AI model, Mixture-of-Experts
February 4, 2025
Huawei and Alibaba Cloud Integrate DeepSeek AI Models for Businesses
Huawei Cloud has teamed up with SiliconFlow, an AI startup, to introduce DeepSeek’s AI models to its Ascend cloud service. This collaboration aims to make AI technology more affordable and accessible to users. The two primary models available are DeepSeek V3, which is a language model, and DeepSeek R1, a reasoning model. Both models deliver exceptional performance while keeping costs low.
Affordable AI Access with DeepSeek
Huawei Cloud is offering reduced prices for the DeepSeek AI models via SiliconFlow. The pricing is attractive, with the DeepSeek V3 model priced at just 1 yuan ($0.13) for every million input tokens and 2 yuan ($0.26) for million output tokens. Meanwhile, the DeepSeek R1 model comes in at 4 yuan ($0.53) per million input tokens and 16 yuan ($2.13) for million output tokens. These low prices are helping to make AI technology more reachable for a wider audience.
AI Independence and Geopolitical Impact
The rollout of DeepSeek models by Huawei highlights China’s ambition for self-reliance in AI, especially in light of U.S. limitations on advanced chips. The models operate on Huawei’s Ascend AI hardware, which lessens the reliance on outside technology. This strategy not only fortifies China’s AI landscape but also enhances its position as a formidable player in the global AI arena.
Alibaba joins the DeepSeek AI wave
In the wake of Huawei’s deployment, Alibaba Cloud has also started offering DeepSeek AI services on its own platform. Users can utilize these models via the PAI Model Gallery, which enables them to implement AI applications without needing to write code. Alibaba Cloud provides both complete and distilled versions of the DeepSeek R1 model, allowing for greater adaptability for various AI uses.
DeepSeek’s growth and future prospects
DeepSeek AI is seeing a rise in popularity in both China and the U.S., with major players like Microsoft and Amazon beginning to integrate its models into their services. Although DeepSeek initially trained its AI models using Nvidia’s H100 GPUs, it currently executes inference on Huawei’s Ascend 910C chip. Additionally, Huawei is in the process of developing the Ascend 920C, anticipated to compete with Nvidia’s forthcoming Blackwell B200, which will further enhance DeepSeek’s AI capabilities.
The swift growth of DeepSeek and its integration with prominent cloud services indicate China’s rising impact in the AI domain. With its budget-friendly models and upgraded hardware, DeepSeek is positioning itself as a key contender in the global AI competition.
Tags: DeepSeek, Huawei Ascend, SiliconFlow
February 4, 2025
GeForce RTX 5090 Beats Radeon RX 7900 XTX in DeepSeek Test
Last week, AMD asserted that its Radeon RX 7900 XTX could outperform Nvidia’s GeForce RTX 4090 in a DeepSeek benchmark. However, the test did not include Nvidia’s latest Blackwell-based GeForce RTX 5090, instead using the older RTX 4080 Super. In response, Nvidia has released its own benchmarks, which, as expected, highlight its products in a much more favorable manner.
Proper Labelling Matters
In contrast to AMD, Nvidia accurately labeled its Y-axis (tokens/second). It conducted tests using the Llama-bench platform with int4 quantization. In the initial test featuring 7 billion parameters, the Radeon RX 7900 XTX reached just over 100 tokens per second. The RTX 4090 outperformed it by 46%, achieving around 150 tokens per second, while the RTX 5090 surpassed it by an impressive 103%, hitting approximately 200 tokens per second.
Consistent Results Across Models
The results remain largely consistent with a model of 8 billion tokens, and when testing with a 32 billion token model, the RTX 5090’s advantage increases to 124%, generating about 50 tokens per second. It’s important to note that these benchmarks come directly from the companies and should be viewed with a degree of skepticism. Additionally, both companies seem to have designed their testing methods to favor their own results. Nonetheless, it isn’t shocking to see that the RTX 5090 outpaces the two-year-old RX 7900 XTX, particularly in a competitive environment where Nvidia has a stronghold.
Source:
Link

Tags: amd, Nvidia GeForce RTX 5090, Radeon RX 7900 XTX
February 3, 2025
DeepSeek Surges in US While ChatGPT Experiences Decline
As ChatGPT starts to show signs of losing its grip, DeepSeek, a rising Chinese AI company, is pushing hard into the global arena. Data from Semrush indicates a sharp decline in ChatGPT’s daily visits in the U.S., dropping from 22.1 million in October 2024 to just 14.9 million by January 2025. In contrast, DeepSeek is experiencing impressive growth, with visits leaping from 2.3k in October to 71.2k by January 19, marking a remarkable increase in a very short time.
Global Impact of DeepSeek
On a worldwide scale, DeepSeek’s traffic has skyrocketed past 7.12 million, and its quick rise to the top of download charts in 140 countries indicates a significant change in what consumers prefer. Yet, this success has ignited serious controversy.
National Security Concerns
Central to the argument are the national security risks associated with DeepSeek. Unlike OpenAI’s ChatGPT, which operates under different regulations, DeepSeek is governed by Chinese laws that require strict data-sharing. Experts are raising alarms about the app’s massive amount of user data, which includes everything from personal inquiries to specific industry information that may be used for strategic advantages. Issues regarding cybersecurity have already come to light, as major companies have pointed out vulnerabilities, with DeepSeek itself admitting to a serious cyberattack just last week.
Dewardric McNeal, a senior analyst at Longview Global, highlights the larger significance of this issue: “This isn’t just about stolen data—it’s about mapping public sentiment, tracking industry trends, and influencing narratives.” Concerns go beyond individual users; industries and lawmakers are wary that DeepSeek’s open-source nature could give China more insight into U.S. supply chains and technological advancements.
Former advisor to the Biden administration, Matt Pearl, expresses even greater worry, labeling DeepSeek’s privacy policy as worthless under Chinese regulations. He cautions that the app could be used for mass surveillance, monitoring users across multiple devices, and potentially embedding malware through software updates. “One bad update is all it takes,” Pearl warns.
Future of DeepSeek in the U.S.
With rising tensions between the U.S. and China, some are beginning to think that a ban on DeepSeek is likely. “If TikTok faced scrutiny, DeepSeek is an even bigger concern,” Pearl states. As the competition in AI heats up, the lingering question is: Will the U.S. permit a Chinese AI company to take over its market, or will regulatory actions put a stop to its expansion?
Source:
Link
Tags: AI competition, ChatGPT, DeepSeek
February 2, 2025
India Develops Affordable AI Models to Compete with ChatGPT
Artificial Intelligence is rapidly changing the landscape, with LLMs (Large Language Models) gaining significant traction due to their wide range of uses. ChatGPT stands out as a notable AI model, alongside new innovative entrants like DeepSeek from China. However, India is now stepping up to the plate, aiming to introduce its own AI model, potentially launching within this year.
Indian Government’s Initiative for Affordable AI
At a recent AI conference, Ashwini Vaishnaw, who is the Union Minister of Electronics and Information Technology, announced that the Indian government is in the process of developing its foundational AI model. He emphasized that this model will offer functionalities similar to those of DeepSeek and ChatGPT, but at a much lower development cost. The minister mentioned that this new AI solution could be ready in approximately 8 to 10 months.
Focus on Local Needs and Inclusivity
During the event organized by the Indian AI Mission, Ashwini Vaishnaw disclosed that researchers in the country are crafting an AI ecosystem framework aimed at supporting this foundational AI model. The goal is to create an experience that caters specifically to Indian users, addressing their linguistic and contextual needs. This initiative seeks to promote inclusivity while working to remove biases found in existing models.
Computational Strength Behind the AI Development
The Union Minister also highlighted India’s computational capabilities, noting that the domestic AI model is being developed using a facility equipped with 18,693 GPUs. In comparison, ChatGPT was trained using around 25,000 GPUs, while DeepSeek utilized 2,000 GPUs for its development.
Cost Comparison with Existing AI Models
Typically, using a well-known AI model like ChatGPT might cost about $3 per hour, but the Indian AI model is expected to be priced at just Rs 100 (approximately $1.15), thanks to government subsidies. This announcement comes on the heels of UC Berkeley researchers successfully replicating DeepSeek AI for a mere $30.
Tags: ChatGPT, DeepSeek, Indian AI Model
February 1, 2025
OpenAI Unveils Enhanced O3-Mini AI with Free ChatGPT Access
OpenAI has introduced its newest AI large language model, the o3-mini, which is quicker and more efficient than its earlier version, the o1-mini, in delivering precise responses. This model marks the first small reasoning LLM that is freely available for all ChatGPT account holders starting today.
Three Reasoning Levels
The o3-mini model offers three levels of reasoning efforts: low, medium, and high. Users with a free ChatGPT account can access o3-mini at the medium reasoning level without any cost, while those with paid accounts have the option to select either the low or high reasoning levels. Subscribers of ChatGPT Plus, Team, and Pro can utilize o3-mini right away, but Enterprise users will need to wait until February. For developers eager to build popular apps utilizing the OpenAI o3-mini API, guidance is available in a book on Amazon.
Performance Comparison
Overall, o3-mini in medium or high reasoning modes excels beyond o1-mini according to standardized AI benchmarks, especially in tasks that require reasoning. In high reasoning mode, o3-mini even surpasses the larger o1 model on certain benchmarks; however, it doesn’t have o1’s capability to interpret images or handle visual data.
Speed and Efficiency
Interestingly, o3-mini delivers responses 24% faster, which equates to approximately 2.46 seconds quicker than o1-mini. This improvement not only decreases the waiting time for ChatGPT users but also lessens the carbon footprint associated with running o3-mini. While it is more efficient, hackers might find o3-mini disappointing, as its potential for being utilized in cybersecurity attacks has been significantly limited.
OpenAI news release, OpenAI o3-mini system card
Source:
Link
Tags: ChatGPT, o3-mini, OpenAI
February 1, 2025
UC Berkeley Researchers Replicate DeepSeek AI for Just $30
AI research has long been dominated by large tech firms with substantial funding. But a team from UC Berkeley has changed the narrative. They successfully replicated the main features of DeepSeek R1-Zero for a mere $30 (yes, you read that right). Their initiative, named TinyZero, shows that sophisticated AI reasoning models can be created without hefty expenses. Plus, AI research is becoming increasingly available to everyone.
The Team’s Ambition
Under the leadership of Jiayi Pan, the researchers set out to recreate DeepSeek’s reasoning framework using reinforcement learning (RL). Rather than depending on costly cloud services or enormous computational resources, they trained TinyZero using just a basic language model, a straightforward prompt, and a reward system. Pan expressed his enthusiasm on X (previously known as Twitter), stating, “You can experience the ‘Aha’ moment yourself for < $30.” He also mentioned that TinyZero represents the first open reproduction of reasoning models, emphasizing its capability to verify and enhance its own responses.
Development Process
In order to evaluate the model, the researchers utilized a game called Countdown, where players must attain a target number through basic mathematical operations. Initially, TinyZero made random guesses, but over time, it learned to confirm its answers, seek improved solutions, and adjust its strategies. They experimented with various model sizes, ranging from 500 million parameters to 7 billion parameters. The findings were intriguing: smaller models (0.5B parameters) merely guessed answers and halted, while larger models (1.5B+ parameters) began to self-verify, refine their responses, and notably enhance accuracy.
Impressive Affordability
What really sets TinyZero apart is its low cost compared to conventional AI models. Here’s a look at the expenses:
– OpenAI’s API: $15 per million tokens
– DeepSeek-R1: $0.55 per million tokens
– TinyZero’s total cost: $30—one-time training expense
This accessibility means that anyone—not just large tech corporations—can explore AI reasoning models without financial strain.
Open for Exploration
TinyZero is open-source and can be found on GitHub, making it possible for anyone to experiment with it. While it has only been tested in the Countdown game, Pan aspires for this project to broaden the reach of reinforcement learning research. However, he acknowledged that it’s still in the early stages, stating, “One caveat, of course, is that it’s validated only in the Countdown task but not the general reasoning domain.” Yet, even with this limitation, the implications are significant: AI development need not be costly. With initiatives like TinyZero, affordable and open-source AI might represent the future.
Source:
Link
Tags: DeepSeek R1-Zero, TinyZero, UC Berkeley
February 1, 2025
TinyZero: Affordable DeepSeek AI Clone Developed for $30
While it’s quite difficult to confirm the supposed development expenses of this new AI model, even the hours put in by PhD students are likely worth much more than a mere $30. Nevertheless, this recent narrative demonstrates that the upheavals in the AI industry weren’t merely a result of exaggerated reactions. It’s worth noting that some AI models for personal use that require minimal computing power have been around for a while, even prior to the emergence of DeepSeek.
The Power of DeepSeek
Clearly, DeepSeek operates with a significantly larger database, whereas the researchers at the University of Berkeley zero in on “reinforcement learning.” Their software doesn’t rely on human-generated data, as the AI can only confirm its own findings.
Self-Verification Process
This self-checking mechanism can produce extensive chains of reasoning, but it does necessitate a certain amount of processing time. This strategy is effective for straightforward mathematical problems and programming tasks, due to the simplicity of the verification involved in these scenarios.
The AI adjusts its methods gradually to arrive at the right answer with the least number of steps possible. Users can access this AI on GitHub, where the source code and usage examples are available.
Open Access to Data
The creators have also made available all the data that was used in this programming trial. Since the model doesn’t incorporate extra data, TinyZero AI stands out as a very streamlined AI model. However, it is expected to perform comparably to more advanced models in specific tasks, such as solving a mathematical puzzle game.
Regardless of whether this AI was truly developed for just $30, once foundational technologies are made publicly accessible, there will always be individuals who can enhance or refine them. After all, today’s smartphones require only a small fraction of the computing power that early 2000s desktop PCs needed and still outperform them. In the realm of AI, this progress appears to be accelerating even more rapidly.
Source:
Link
Tags: DeepSeek, reinforcement learning, TinyZero AI