DeepSeek is powered by the open source DeepSeek-V3 model, which its researchers claim was trained for around 6m ( 4.2m) - significantly less than the billions spent by rivals. But this claim has been disputed by others in AI. Its emergence comes as the US is restricting the sale of the advanced chip technology that powers AI to China. To continue their work without steady supplies of imported advanced chips, Chinese AI developers have shared their work with each other and experimented with new approaches to the technology. This has resulted in AI models that require far less computing power than before.

The new Chinese AI that rivals US big tech at a fraction of the cost has seen its popularity soar.

The speed at which the new Chinese AI app DeepSeek has shaken the technology industry, the markets and the bullish sense of American superiority in the field of artificial intelligence (AI) has been nothing short of stunning. Venture capitalist Marc Andreessen may have said it best. "DeepSeek-R1 is AI's Sputnik moment," he posted to X on Sunday, referring to the satellite which kicked off the space race. DeepSeek was the most downloaded free app on Apple's US App Store over the weekend. By Monday, the new AI chatbot had triggered a massive sell-off of major tech stocks which were in freefall as fears mounted over America's leadership in the sector.

The launch of a new chatbot by Chinese artificial intelligence firm DeepSeek triggered a plunge in US tech stocks as it appeared to perform as well as OpenAI's ChatGPT and other AI models, but using fewer resources. By Monday, DeepSeek's AI assistant had rapidly overtaken ChatGPT as the most popular free app in Apple's US and UK app stores. Despite its popularity with international users, the app appears to censor answers to sensitive questions about China and its government. Chinese generative AI must not contain content that violates the country's "core socialist values", according to a technical document published by the national cybersecurity standards committee. That includes content that "incites to subvert state power and overthrow the socialist system", or "endangers national security and interests and damages the national image".

The DeepSeek AI chatbot, released by a Chinese startup, has temporarily dethroned OpenAI's ChatGPT from the top spot on Apple's US App Store. Also, the DeepSeek model was efficiently trained using less powerful AI chips, making it a benchmark of innovative engineering. I've tested many new generative AI tools over the past couple of years, so I was curious to see how DeepSeek compares to the ChatGPT app already on my smartphone. After a few hours of using it, my initial impressions are that DeepSeek's R1 model will be a major disruptor for US-based AI companies, but it still suffers from the weaknesses common to other generative AI tools, like rampant hallucinations, invasive moderation, and questionably scraped material. Users interested in trying out DeepSeek can access the R1 model through the Chinese startup's smartphone apps (Android, Apple), as well as on the company's desktop website.

Powerful artificial intelligence software from Chinese startup DeepSeek indicates that its engineers built a competitive model despite U.S. attempts to curtail China's tech development, raising questions about the effectiveness of Washington's trade curbs. The company's recently released R1 model, which it claims to have developed at a fraction of the cost borne by rival AI companies, sent tech stocks into a tailspin Monday as investors questioned the need to spend billions on advanced hardware. It's also sparked a debate in Washington about the best strategy to prevent China from developing cutting-edge AI, which U.S. policymakers see as a national security risk. The U.S. imposed sweeping controls on the sale of the most advanced Nvidia chips to China in October 2022, and has ratcheted up the measures each year since. But Nvidia has responded by designing new semiconductors for the Chinese market -- including those DeepSeek likely used to build R1.

Investors sold technology stocks across the globe on Monday as they worried that the emergence of a low-cost Chinese artificial intelligence model would threaten the dominance of current AI leaders like Nvidia, shaving 592.7 billion off the chipmaker's market value. Last week, Chinese startup DeepSeek launched a free AI assistant that it says uses less data at a fraction of the cost of incumbent services. By Monday, the assistant had overtaken U.S. rival ChatGPT in downloads from Apple's app store. This led the tech-heavy Nasdaq to fall 3.1% on Monday. Nvidia was the Nasdaq's biggest drag, with its shares tumbling just under 17% and marking a record one-day loss in market capitalization for a Wall Street stock, according to LSEG data.

Recent advancements in fine-tuning Vision-Language Foundation Models (VLMs) have garnered significant attention for their effectiveness in downstream few-shot learning tasks.While these recent approaches exhibits some performance improvements, they often suffer from excessive training parameters and high computational costs. To address these challenges, we propose a novel Block matrix-based low-rank adaptation framework, called Block-LoRA, for fine-tuning VLMs on downstream few-shot tasks. Inspired by recent work on Low-Rank Adaptation (LoRA), Block-LoRA partitions the original low-rank decomposition matrix of LoRA into a series of sub-matrices while sharing all down-projection sub-matrices. This structure not only reduces the number of training parameters, but also transforms certain complex matrix multiplication operations into simpler matrix addition, significantly lowering the computational cost of fine-tuning. Notably, Block-LoRA enables fine-tuning CLIP on the ImageNet few-shot benchmark using a single 24GB GPU. We also show that Block-LoRA has the more tighter bound of generalization error than vanilla LoRA. Without bells and whistles, extensive experiments demonstrate that Block-LoRA achieves competitive performance compared to state-of-the-art CLIP-based few-shot methods, while maintaining a low training parameters count and reduced computational overhead.

Vision-and-Language Navigation (VLN) tasks mainly evaluate agents based on one-time execution of individual instructions across multiple environments, aiming to develop agents capable of functioning in any environment in a zero-shot manner. However, real-world navigation robots often operate in persistent environments with relatively consistent physical layouts, visual observations, and language styles from instructors. Such a gap in the task setting presents an opportunity to improve VLN agents by incorporating continuous adaptation to specific environments. To better reflect these real-world conditions, we introduce GSA-VLN, a novel task requiring agents to execute navigation instructions within a specific scene and simultaneously adapt to it for improved performance over time. To evaluate the proposed task, one has to address two challenges in existing VLN datasets: the lack of OOD data, and the limited number and style diversity of instructions for each scene. Therefore, we propose a new dataset, GSA-R2R, which significantly expands the diversity and quantity of environments and instructions for the R2R dataset to evaluate agent adaptability in both ID and OOD contexts. Furthermore, we design a three-stage instruction orchestration pipeline that leverages LLMs to refine speaker-generated instructions and apply role-playing techniques to rephrase instructions into different speaking styles. This is motivated by the observation that each individual user often has consistent signatures or preferences in their instructions. We conducted extensive experiments on GSA-R2R to thoroughly evaluate our dataset and benchmark various methods. Based on our findings, we propose a novel method, GR-DUET, which incorporates memory-based navigation graphs with an environment-specific training strategy, achieving state-of-the-art results on all GSA-R2R splits.

Supervised fine-tuning (SFT) and reinforcement learning (RL) are widely used post-training techniques for foundation models. However, their roles in enhancing model generalization capabilities remain unclear. This paper studies the difference between SFT and RL on generalization and memorization, focusing on text-based rule variants and visual variants. We introduce GeneralPoints, an arithmetic reasoning card game, and adopt V-IRL, a real-world navigation environment, to assess how models trained with SFT and RL generalize to unseen variants in both textual and visual domains. We show that RL, especially when trained with an outcome-based reward, generalizes across both rule-based textual and visual variants. SFT, in contrast, tends to memorize training data and struggles to generalize out-of-distribution scenarios. Further analysis reveals that RL improves the model's underlying visual recognition capabilities, contributing to its enhanced generalization in the visual domain. Despite RL's superior generalization, we show that SFT remains essential for effective RL training; SFT stabilizes the model's output format, enabling subsequent RL to achieve its performance gains. These findings demonstrates the capability of RL for acquiring generalizable knowledge in complex, multi-modal tasks.