DeepSeek

Chinese AI Lab Unveils GLM-5.2, a Major Leap in Open Source AI Models

Chinese AI lab Z.ai has released GLM-5.2, a massive text-only AI model with 753 billion parameters and a context window of 1 million tokens. This release follows the open-source approach, making its weights available under an MIT license. While similar in size to its predecessors, GLM-5.2 stands out for its improved performance on benchmarks like the Artificial Analysis Intelligence Index, where it ranks first with a score of 51, surpassing models like MiniMax-M3 and DeepSeek V4 Pro. However, the model's high token usage-43k output tokens per task-is notable. This contrasts with competitors like Claude Fable 5, which tops the Code Arena WebDev leaderboard despite GLM-5.2's lack of image inputs. Currently, access to GLM-5.2 via OpenRouter costs $1.40 per million input tokens and $4.40 per million output tokens, positioning it as a cost-effective alternative to models like GPT-5.5 and Claude Opus. The release marks a significant milestone in open-source AI, offering developers and researchers a powerful tool for text-based tasks. As the model gains wider adoption, its impact on coding and other applications will be closely watched.

Simon Willison5d ago

AI Agents' True Smarts Lie in Code, Not Just Models

A new study reveals that the real challenge in creating autonomous AI agents isn't just their language models but the software surrounding them. This includes tools, memory systems, testing processes, and permission settings that transform static models into dynamic agents capable of thinking and acting independently. The paper highlights that while the model is crucial, it's the "harness" or the code that actually enables the AI to perform tasks. For example, Deepseek is already building a dedicated team in Beijing focused on developing this harness technology. Their core formula-model plus harness-demonstrates how essential this layer is for creating functional AI agents. As AI continues to evolve, expect more focus on refining these software layers to improve agent capabilities. This shift could unlock new possibilities for autonomous systems across industries, from healthcare to robotics.

The Decoder3w ago

China Restricts AI Researchers' Overseas Travel

China has started requiring top AI researchers at companies like Alibaba and DeepSeek to get official approval before traveling abroad. The government is concerned about data leaks, technology theft, and losing talent to other countries. This move aims to tighten control over the domestic AI industry and protect sensitive information. This decision highlights Beijing's growing focus on safeguarding its AI sector, which is seen as crucial for the country's technological advancement. By restricting travel, China hopes to prevent its top experts from being poached by international companies or sharing their knowledge abroad. This could create challenges for global collaboration in AI research and innovation. Looking ahead, this policy may intensify competition for AI talent both within China and internationally. Researchers will need to navigate stricter regulations to continue participating in global conferences and collaborations, potentially slowing the flow of ideas and expertise across borders.

The Decoder3w ago

Deepseek Permanently Cuts AI Model Prices, Undercutting Competitors

Deepseek has made its steep discount on the V4-Pro model permanent. Now priced at $0.435 per million input tokens, it's at least 11.5 times cheaper than GPT-5.5 and over 34 times cheaper for output tokens. This move could put pressure on Western providers struggling to match these prices. For developers and researchers, this pricing makes Deepseek a more attractive option for token-hungry systems like agentic AI. The lower costs could accelerate adoption of such systems globally, especially in regions where cost is a major factor. This shift underscores the growing competition in AI pricing. As other providers respond, we'll likely see further price reductions and innovations in model efficiency to stay competitive.

The Decoder4w ago

Alibaba's AI Model Breaks Record, Runs Autonomously for 35 Hours

Alibaba's Qwen team has unveiled the Qwen3.7-Max, a cutting-edge AI model designed for long-running autonomous tasks. This powerful system not only matched the performance of Claude Opus 4.6 in benchmark tests but also outperformed Chinese competitors like DeepSeek V4 Pro and Kimi K2. What makes this model truly stand out is its ability to operate independently for extended periods-a remarkable 35 hours, during which it optimized code for Alibaba's custom chips. This achievement is a significant milestone for AI development, particularly in the realm of autonomous systems. Such models can revolutionize industries by handling complex, long-term tasks without human intervention. For instance, Qwen3.7-Max demonstrated its versatility by steering a four-legged robot, showcasing potential applications in robotics and automation. This breakthrough could lead to more efficient and reliable AI-driven solutions across various sectors. As the field of AI continues to advance, Qwen3.7-Max sets a new standard for autonomous capabilities. Future developments may focus on expanding its applications and improving its efficiency, potentially leading to even more groundbreaking innovations in AI technology.

The Decoder4w ago

DeepSeek Launches AI Code Agent, Competing Directly With Claude Code and Codex

DeepSeek is ramping up its artificial intelligence efforts with the launch of a new team in Beijing dedicated to developing an AI code agent called "Deepseek Code." This new tool aims to directly compete with established players like Claude Code, Codex, and Cursor. The company is seeking candidates with expertise in agent loops, MCP, and context engineering, as well as those who are already familiar with existing coding tools. The move marks a significant step for DeepSeek in the AI-powered coding space, which has seen rapid growth over the past year. While OpenAI's Codex and Anthropic's Claude Code have dominated this niche market, DeepSeek is betting on its engineering prowess to carve out a unique position. The company plans to leverage its Beijing team's technical expertise to build a system that can understand, generate, and optimize code with unprecedented accuracy. As the competition heats up, industry watchers are curious about how DeepSeek's offering will differentiate itself from the pack. With major players already in the game, DeepSeek's success will depend on its ability to innovate while addressing the specific needs of developers and researchers. Stay tuned for updates as this promising new entrant carves out its place in the AI coding landscape.

The Decoder4w ago

DeepSeek V4 Revolutionizes AI Coding with Affordable Pricing

DeepSeek has launched its V4 model, an open-source AI system available under the MIT license. Priced at $0.30 per million output tokens, it’s significantly cheaper than competitors like Claude Opus 4.7 ($25) and GPT-5.5 ($30). The model scored 80.6% on SWE-bench Verified, just 0.2 points behind Claude Opus 4.6. Its efficient architecture, including a 1.6-trillion-parameter MoE and optimized inference processes, makes this pricing sustainable without being a loss leader. While self-hosting requires substantial resources, the model’s capabilities in coding tasks like LiveCodeBench and Codeforces are unmatched, challenging the dominance of closed models. However, concerns about benchmark transparency and data governance remain. This release resets the price floor for high-quality coding AI, potentially forcing competitors to adjust their strategies or enhance unique features.

Hacker News1mo ago

AI Advances Push Boundaries of Reinforcement Learning

Recent developments in reinforcement learning (RL) have shown surprising progress, challenging earlier predictions that improvements would slow down as tasks became more complex. For instance, the DeepSeek-R1-Zero model demonstrated the ability to reason through lengthy chains of thought using a single rule-based reward system. Over thousands of training steps, its responses expanded from short to over ten thousand tokens, with accuracy steadily increasing. A key insight is that progress in RL isn't solely determined by "horizon length," or the time it takes for rewards to manifest after actions. Instead, three independent factors-learned internal evaluators, exploration strategies, and substrate plasticity-play crucial roles in determining success. These factors explain why models excel at benchmarks like theorem proving and coding but face challenges in areas requiring softer skills, such as creative writing or long-term research. Looking ahead, researchers are focusing on optimizing how agents evaluate their own progress and explore trajectories before receiving feedback. This could unlock improvements in both RL efficiency and human-like decision-making abilities in AI systems.

LessWrong, arXiv CS.AI1mo ago