Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Kimi K2.5 is a multimodal model capable of visual reasoning, code generation from UI and video inputs, and agentic task orchestration using a swarm-based architecture. It's built on 15T vision-language tokens through continual pretraining.
Google Search now features Gemini 3 as the default AI model for enhanced AI Overviews, enabling users to find detailed answers quickly. Users can seamlessly engage in follow-up conversations with AI Mode directly from these overviews for more in-depth inquiries. These upgrades aim to provide a fluid search experience, combining quick snapshots with the option for deeper exploration.
An INT4 quantization-aware training pipeline that enables rollout of ~1TB-scale models on a single H200 GPU. By combining fake quantization during training with real quantization at inference, it achieves BF16-level stability and opens a path to cost-efficient, high-performance deployment.
The "DeepSeek Moment" predicted to trigger an AI Winter turned out to be minor, with NVIDIA's stock quickly recovering from an initial 17% drop. DeepSeek influenced a shift towards cheap, open-source AI models, particularly benefiting China. The upcoming 'V4' model from DeepSeek could further impact the balance in AI advancements, especially with the US and China navigating tensions over NVIDIA chip sales.
OpenAI is planning to take two approaches to monetizing its free tier: classic intent-based ad formats and affiliate fees on native checkout. Long term, the Ads opportunity for the company is massive and could even be a Google, Meta, or TikTok scale opportunity. A shift to advertising to agents could change the fundamental nature of what ads look like. The winning ads will probably work less to persuade humans and look more like offers that make the user better off.
Sponsored
Context graphs aren't retrieval systems; they're production world models. They allow you to ask "what breaks if I merge this PR?" so you can fix defects _before_ they hit production. PlayerZero [12] builds these production world models so your AI actually understands how systems behave.
The Kimi Agent SDK is a set of multi-language libraries that expose the Kimi Code agent runtime in applications. It can be used to build products, automations, and custom tooling. The SDK clients reuse the same Kimi CLI configuration, tools, and MCP servers to stream responses in real-time, surface approvals and tool calls, and let users orchestrate sessions programmatically. Go, Node.js, and Python are supported.
DeepSeek-OCR 2 is a significant upgrade to DeepSeek-OCR. It maintains high visual token compression while achieving meaningful performance improvements. The model is powered by DeepEncoder V2, which implicitly distills causal understanding of the visual world through the integration of both bidirectional and causal attention mechanisms. This leads to causal reasoning capabilities in the vision encoder and marked lifts in visual reading logic.
AI2 released open-source coding agents and a training recipe that makes it affordable and practical to build coding agents for private or internal codebases. Its method reproduces top-tier performance at drastically lower compute costs compared to previous approaches.
SOAR is a self-improvement framework that uses meta-RL to help pretrained models learn from problems they initially fail at. By grounding synthetic problem generation in student model progress, it enables learning on sparse-reward tasks without requiring correct solutions upfront.
SoftBank is in talks to invest up to $30 billion more into OpenAI. OpenAI is seeking up to $100 billion in new capital in a round that could value it as much as $830 billion. It needs substantial funding to continue developing its AI models, pay for its vast computing needs, and retain top researchers in an increasingly competitive market. The company is aiming to raise capital from Middle Eastern sovereign-wealth funds and other venture capital funds.
Sponsored
Stop predicting workflows. Start observing them. BAAPs discover, assemble, and deploy agents automatically - based on how your
Clawdbot's persistent memory system, which maintains 24/7 context retention, remembers conversations and builds upon previous interactions indefinitely.
Google has expanded availability of AI Plus, which includes access to Gemini models and AI productivity tools, to 35 more countries.
FLORA unifies the best creative AI models across text, image, and video into one workflow to accelerate the creative process.
Manus AI's Agent Skills open standard enables modular AI workflows for professionals, integrating Python and Bash scripts with a focus on efficiency and interoperability.
Sponsored
Most product launches start the same way: a blank page and a deadline.
Can't join live? Register anyway [4], and we'll send you the recording so you can watch on your own time.
Anthropic has introduced MCP app integration into Claude's AI assistant, enabling seamless use of tools like Asana, Slack, Figma, and Box within its interface. This update enhances real-time interaction, setting Claude apart by embedding live app experiences directly into chats, boosting workflow efficiency. The move supports open standards and encourages developer-created integrations, aligning with Anthropic's ecosystem goals.
Qwen3-Max-Thinking is a reasoning model for complex math, coding, and multi-step workflows. Available on Alibaba Cloud's Model Studio, this model excels in tasks that require evidence gathering and deep verification, offering adaptive tool-use and built-in web search features. Early testers highlight its benefits for developers and enterprises needing long-context reasoning and tool-using agents.
Dario Amodei argues that as AI approaches the capability of a "country of geniuses," humanity faces critical challenges in safety, misuse, and economic disruption. He highlights the risks of allowing powerful AI to concentrate in authoritarian regimes, potentially leading to irreversible totalitarian control or unprecedented destructive capability in the wrong hands. Amodei advocates for strategic controls and regulations to align AI development with democratic principles, emphasizing the urgent need for informed global cooperation to prevent catastrophic outcomes.
A three-phase approach to building robust evaluation workflows for voice systems, helping teams move from demos to production by structuring datasets, grading mechanisms, and feedback loops.
Model benchmarks matter less than compute access, infrastructure scale, and long-term demand for intelligence. The race is not zero-sum, and all major labs will become massive as AI adoption, enterprise agents, and compute needs grow exponentially.
Co-do showcases how browsers can act as sandboxes for running AI-powered applications safely. It utilizes layered sandboxing by restricting file system access with the File System Access API, controlling network requests via strict CSP policies, and running code in isolated WebAssembly environments. This approach highlights the potential for browsers to serve as secure execution environments for untrusted code, but reveals gaps like dependency on LLM providers and the limitations of current browser security features.
Sponsored
Auth0 for AI Agents provides a more secure foundation for developers to build AI agents without compromising security or innovation. This solution connects your AI agents to apps and data, applying the correct user permissions in every scenario.
ChatGPT Code Interpreter underwent a massive upgrade in the past few months. It can write and then test code in 11 new languages, find files online and download them into a container, and install packages via pip and npm to help it solve problems. There is no official documentation for these new features. This post provides an overview of the update, along with a full list of tools that ChatGPT now has access to.
SWE-Pruner introduces a task-aware context pruning framework that cuts token usage in code agents by over 50% without harming performance. It dynamically selects relevant code using a lightweight model, optimizing workflows in multi-turn engineering tasks.
Endless Terminals provides an automated pipeline to generate realistic terminal tasks without human labeling, enabling scalable reinforcement learning for terminal-based agents.
DeepVerifier is a test-time verifier that improves deep research agents by applying rubric-based feedback during inference.
Yann LeCun, one of the world's leading experts on artificial intelligence, has become increasingly vocal in his criticism of Silicon Valley's approach to building AI. He says the technology industry will eventually hit a dead end in its AI development as LLM technology has its limits. The herd effect in Silicon Valley leaves no room for other approaches that may be more promising in the long run. If everyone were open, the field as a whole would progress faster.
NVIDIA has invested $2 billion in CoreWeave to accelerate the development of over 5 gigawatts of AI factories by 2030.
GPUs have 3-5 year economic lives, which creates extremely steep depreciation curves, making long-term debt risky.
Sam Altman sits down with builders from across the AI ecosystem to answer questions and talk about the future of building with AI.
RadixArk, spun out from SGLang, has been valued at $400 million after securing investment from Accel, reflecting increased interest in AI model optimization tools.
Sponsored
Voice is multi-dimensional. Tone and timing can change the meaning of a conversation and transcript-reliant LLMs miss it all.
Apple plans to unveil a Gemini-powered Siri assistant in February, leveraging Google's AI models to enhance task completion using user data. An even more advanced version that offers more conversational capabilities akin to ChatGPT will be revealed at the Worldwide Developers Conference in June. This development marks a shift in Apple's AI strategy following key leadership changes and its partnership with Google.
OpenAI is adding a dedicated shopping cart section to ChatGPT. It will serve as a centralized space for reviewing selected items and finalizing purchases. The company is also adding support for personal responses in temporary chat sessions to balance privacy with customized assistance. There is no clear timeline for public availability of these features.
Codex CLI is a cross-platform local software agent designed to produce high-quality, reliable software changes while operating safely and efficiently on users' machines. This is the first part in a series of posts that explores how Codex works. It focuses on the agent loop, which is the core logic responsible for orchestrating the interaction between the user, the model, and the tools the model invokes to perform work. The post provides a view into the role the Codex harness plays in making use of a large language model.
Building AI agents for financial services requires rigorous technical infrastructure and data normalization to avoid costly errors. Key insights include the essentiality of sandbox environments for secure multi-step workflows and the transformation of raw financial data into structured, searchable context through markdown and metadata. Skills systems, using markdown instead of code, allow dynamic, user-specific instructions for models, which is crucial as AI models rapidly improve and reduce the need for complex scaffolding.
Sponsored
How do Perplexity, NVIDIA, and Ahrefs get live search results from Google and Bing? SERPAPI is the industry-standard web search API for AI -- the fast, easy, and complete way to scrape search results. Get in touch to claim a 50% discount on your first 3 months [10] (new
Large language model inference is hard. The primary challenges are memory and interconnect rather than compute. This paper highlights four architecture research opportunities: High Bandwidth Flash for 10X memory capacity with HBM-like bandwidth, Processing-Near-Memory and 3D memory-logic stacking for high memory bandwidth, and low-latency interconnect to speedup communication. The paper focuses on datacenter AI, but it also reviews the architectures' applicability for mobile devices.
TTT-Discover applies reinforcement learning during inference, allowing LLMs to adapt to each task on the fly. It sets new performance benchmarks across math, biology, algorithms, and GPU kernels.