Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Qwen3.5-397B-A17B is the first model in the Qwen3.5 series. The native vision-language model demonstrates outstanding results in reasoning, coding, agent capabilities, and multimodal understanding. It uses an innovative hybrid architecture that fuses linear attention with a sparse mixture-of-experts. While it contains 397 billion parameters, only 17 billion are activated per forward pass. The model supports 201 languages and dialects.
Manus Agents is a new way to access and use Manus directly inside messaging apps. Telegram is currently the only supported app, with more platforms coming soon. The agent features few reasoning, tools, and multi-step task execution. The feature makes agents accessible wherever users are.
Microsoft is testing new Researcher and Analyst agents integrated into Copilot's upcoming "Tasks" feature. This feature will allow users to schedule complex prompts, leveraging OpenAI and o3-mini models for research and data analysis. The addition of an "Auto" mode aims to streamline task automation, potentially differentiating Copilot in productivity use cases.
Anthropic's CEO, Dario Amodei, expects 'geniuses in a data center' to show up within a few years. While Anthropic's actions do not seem to fully reflect this optimism, its caution is necessary. This article contains notes from a recent podcast where Amodei discusses China, export controls, democracy, AI policy, AI risks, and continual learning.
AGI is likely possible, but it probably won't come from Transformer-based models. Transformers are very powerful, but they have fundamental limitations. Solving these limitations could take decades. This isn't to say that LLMs aren't useful - the current technology is already fundamentally changing society.
Inference costs may not be that much of a bottleneck for AI progress. The cost to reach a given capability level falls fast, so the inference cost burden is more transient than it might appear from looking at only frontier models at launch. The data on RL scaling is still thin, so it is difficult to draw conclusions yet. It will be interesting to see how quickly cheaper models catch up to frontier capability levels, and how inference costs for fixed tasks decrease over time.
900+ global decision-makers about how they're operationalizing agentic AI, with observability as the foundation for trust and control. See
Sponsored
Autonomous operations are rapidly expanding, and 74% of enterprises
Benchmark performance gives biased estimates of out-of-distribution generalization if LLM training data is polluted with benchmark test data. Typical decontamination filters fail to detect semantic duplicates. This suggests that recent benchmark gains are confounded - the prevalence of soft contamination means gains reflect both genuine compatibility improvements and the accumulation of test data and effective test data in the growing training corpora.
Alibaba's ZVEC is an open-source, in-process vector database enabling rapid, scalable similarity searches using Alibaba's PROXIMA engine. It supports dense and sparse vectors with hybrid searches and can be deployed across various platforms, including notebooks and edge devices. Installation is straightforward via Python or Node.js, offering a lightweight solution for handling vector data efficiently.
Spreadsheet Arena is an open platform for evaluating LLM-generated spreadsheets. Formatting and structure often influence user preference more than formula complexity. There are significant differences in domain-specific preferences, with academic models suffering from heavy formatting and finance models benefiting from professional color coding. Crowd preferences often diverge from expert ratings, particularly in color coding and formatting.
Micron Technology, the largest American maker of memory chips, is rushing to add manufacturing capacity to break the memory bottleneck. The company is spending $50 billion to more than double the size of its 450-acre campus. It will build two new chip factories, the first of which is expected to start production of DRAM in mid-2027. Micron also recently broke ground on a $100 billion fab complex in New York, and it announced a $9.6 billion fab investment in Japan last year.
Flapping Airplanes aims to revolutionize AI by developing data-efficient training methods, reducing reliance on vast datasets. Backed by $180 million, the founders emphasize diverging from traditional methods, drawing inspiration from the human brain without replicating it exactly. They focus on creativity and fresh perspectives, employing a team oriented towards groundbreaking research rather than incremental improvements.
Anthropic has a big consumer marketing problem. Its story, while well known in Silicon Valley, isn't widely heard or legible to the greater cultural conscience. It is puzzling how a company so good at aesthetics and narrative engineering can be so bad at this.
Sponsored
Let MongoDB for Startups partner with you as you build the next big thing with the tools, support, and go-to-market opportunities you need
LLMs can assist with decompiling Nintendo 64 games up to a certain point - this post describes a developer's attempt, how their workflow evolved as the project matured, what helped, and where they're currently stuck.
US productivity increased roughly 2.7% for 2025, a near doubling from the sluggish 1.4% annual average that characterized the past decade.
Nearly all models tested responded that they should walk.
AIs optimized as “reward-seekers” might be influenced not just by local training incentives but also by distant retroactive rewards or simulated scenarios administered later or by powerful actors.
AI is constrained by silicon, supply chains, and economics.
Model labs have a structural cost advantage that pure inference providers will struggle to match.
Sponsored
OpenClaw (formerly ClawdBot and MoltBot) is an open-source agent assistant that operates with the same permissions as the user who installs it - and can act on their behalf.
>> WANT TO GET CERTIFIED IN AI SECURITY? With the _Foundations of AI Security_ [5] series, you can learn how to deploy secure AI across the enterprise directly from experts at Zenity and AWS. Attend all three virtual sessions to earn earn the Foundations of AI Security Professional Certificate!
OpenClaw's creator will join OpenAI to develop accessible agents and ensure OpenClaw remains open and independent within a new foundation. OpenAI's support will provide access to cutting-edge models and resources, aligning with the creator's vision of making these tools widely usable. OpenClaw's community will benefit from a supportive structure that encourages innovation and data ownership.
OpenAI introduced an optional Lockdown Mode for higher-risk workflows and added “Elevated Risk” labels for capabilities in ChatGPT, ChatGPT Atlas, and Codex that can increase exposure to prompt-injection attacks.
xAI's Grok Build plans to introduce Parallel Agents and Arena Mode, enhancing its capabilities from a coding assistant to a full IDE. Parallel Agents allows up to eight coding agents to work simultaneously, while Arena Mode facilitates competition among agents to determine the best output. The update also includes a new UI resembling a browser-based IDE, dictation support, and nonfunctional GitHub integration, with Grok 4.20 expected soon.
Both CEOs of OpenAI and Anthropic claim that human-level AI is just around the corner. However, Transformers have their fundamental limitations. While they are very powerful and have taught us a lot about what general intelligence is, overcoming these problems will require research that could take decades. The advances that made modern GPTs possible were discovered gradually over several decades, and each discovery came with new and more nuanced issues.
OpenClaw's recent security issues highlight a larger problem with agent ecosystems as new attack surfaces, where malicious skills can exploit agent + tools + marketplace combinations. Current safeguards like prompts are inadequate. Agents need sandboxing, scoped credentials, restricted tools, and proper logging to manage risks.
Anthropic and OpenAI have unveiled new "fast modes" for LLM inference, offering differing performance enhancements. OpenAI's model speeds exceed 1,000 tokens per second using Cerebras chips, but rely on a less capable model, while Anthropic's fast mode supports real models with up to 2.5x speed through low-batch-size inference. Despite OpenAI's technical achievement, the utility of faster but less capable inference remains debatable, focusing more on OpenAI's exploration of Cerebras' potential.
Text has to pass through a tokenizer before GPT-5.x is able to understand any input. A tokenizer is a compression layer that converts raw text into a sequence of integer IDs. Every design decision baked into the tokenizer has downstream consequences for cost, accuracy, multilingual performance, and hallucination rates. This article looks at OpenAI's open-source tokenizer library, tiktoken, and explains how it works.
Sponsored
Human judgment is the backbone of AI quality. Yet most orgs scaling AI lack clear decision frameworks, consistent interpretation policies, and reliable oversight mechanisms. Welo Data [14] builds AI data quality systems that govern human and machine judgment. Turns out that
AI agents can tackle increasingly complex tasks, but to achieve more ambitious goals, they need to be able to meaningfully decompose problems into manageable sub-components and safely delegate their completion across to other AI agents and humans. This study proposes an adaptive framework for intelligent AI delegation. The proposed framework is applicable to both human and AI delegators and delegatees in complex delegation networks.
Personaplex is a real-time speech-to-speech conversational model that jointly performs streaming speech understanding and speech generation. It operates on continuous audio and can predict both text tokens and audio tokens auto-regressively to produce its spoken responses. Personaplex's design allows the model to update its internal state based on users' ongoing speech while still producing fluent output audio, so it can support highly interactive conversations. The model is ready for commercial use.
Minimax M2.5 is fast, cheap, and great at coding. While the team behind the model has released a technical article on how the model works, it is quite dense. This post aims to make the content of that article more accessible for readers. The core problem the team is solving is how to make RL work at massive scale for being good at Agentic AI. Training LLMs to be good at agentic tasks imposes many challenges: huge volumes of training data need to be processed quickly, training needs to be stable, and agents need to be good at a diverse range of tasks.
Composition-RL reused “too-easy” pass-rate-1 prompts by automatically combining multiple problems into new verifiable questions for RLVR training. Across 4B–30B models, the approach improved reasoning over training on the original dataset.
You have to know where you're going before you take that first step. The first thing to do is always to define your goal. Reverse engineer requests and combine that with context to create discrete, boolean, testable Ideal State Criteria. This criterion can carry through to become the verification criteria as well.
Anthropic's partnership with Palantir means that the Defense Department and federal law enforcement have access to Claude. This means it was possible that the technology was used in matters related to the US military's operation to capture former Venezuelan President Nicolás Maduro. Anthropic's usage guidelines prohibit Claude from being used to facilitate violence, develop weapons, or conduct surveillance. A company spokesperson said that Anthropic has not discussed the use of Claude for specific operations, but the company is committed to using frontier AI to support the US in national security.
Manus AI launched a 24/7 Agent via Telegram but faced immediate suspension without explanation.
Microsoft plans to reduce its reliance on OpenAI by developing its own AI models, spearheaded by AI chief Mustafa Suleyman.
An example of how an agent can be used to edit parts of software like a wiki.
The exponential growth of AI is due to a series of bottlenecks being smashed, and Groq's LPU is the next step.
openrouter/free is a router that selects free models at random from the models available on OpenRouter.
Sponsored
☀️ Imagine waking up to your issues triaged, your CI failures investigated with fixes for you to review, and two fresh PRs proposing improvements to your tests. All of that while you were sleeping.
OpenAI released GPT-5.3-Codex-Spark, a smaller ultra-fast coding model optimized for real-time use in Codex and capable of generating over 1,000 tokens per second on low-latency hardware.
Google released a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode, expanding access to Ultra subscribers and select API users. The update was developed with researchers to better handle open-ended scientific and engineering problems with messy or incomplete data.
Anthropic raised $30 billion in Series G funding at a $380 billion post-money valuation, led by GIC and Coatue with broad institutional participation.