Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Undeterred by its many ongoing copyright battles, OpenAI is building a text-to-music model that could add accompaniment to videos (including those generated by Sora) or layer instruments onto vocal tracks.
ChatGPT's model selector temporarily showed 'GPT-5 Mini Scout' for some business account users on Friday, sparking immediate speculation about its purpose. An update to OpenAI's official JavaScript agent library surfaced shortly after with a new reference to 'GPT-5.1 Mini' in a test file, suggesting that 'GPT-5.1 Mini' was the intended production name for 'Mini Scout'. References to the new model have since been removed.
A research collaboration between Anthropic and Thinking Machines Lab found that Claude, GPT, Grok, and Gemini models have values that contradict what their creators publicly specify. By generating over 300,000 scenarios where models have to choose between competing principles, they found that Claude models prioritize ethical responsibility, OpenAI models favor efficiency, and Gemini/Grok emphasize emotional connection.
Perplexity has outlined its defense-in-depth approach to securing Comet, its AI assistant browser, against malicious prompt injection, an emerging class of attacks that manipulate model behavior without exploiting software vulnerabilities.
Sponsored
Discover and protect sensitive data with Protegrity Developer Edition [10] - a free, lightweight toolkit that installs from PyPI or GitHub.
A PyTorch-native stack of open-source tools for agentic AI, including kernel development, distributed systems, and reinforcement learning, with support for massive GPU workloads and edge deployments.
Claude's internal neurons can recognize when text forms visual art. The same neuron that fires for the word "eyes" also activates when it sees @ symbols arranged as ASCII eyes or SVG circle elements positioned as eyes within a face. Anthropic researchers found these cross-modal patterns across mouths, ears, dogs, and cats, demonstrating that the model builds associations between concepts and their visual representations across formats, similar to how humans can recognize a smiley face whether it's drawn, described, or coded in SVG.
FlashPack is a high-throughput file format and loading mechanism for PyTorch that makes model checkpoint I/O blazingly fast. It can load any model 3–6× faster than with the current state-of-the-art methods. FlashPack is wrapped in a lightweight, pure-Python package that works anywhere. It even works on systems without access to GPU Direct Storage.
The goal isn't delegating your core work and then becoming an auditor or editor. It should be offloading secondary tasks so you can focus on what really matters. Tasks like writing documentation, fixing TypeScript errors, and spiking out changes are the kind of grunt work AI agents excel at that can run async in the background.
Sponsored
Drowning in emails? Jace AI categorizes, drafts, and organizes naturally — saving 2 hours a day so you can focus on real work while
Coco Robotics has enlisted UCLA Professor Bolei Zhou to lead a new physical AI lab using five years of robot data to enhance automation.
Google AI Studio's new vibe coding interface streamlines the process of building AI apps, enabling developers to go from prompt to prototype without complex setup or manual API integration.
This release of the olmOCR model was fine-tuned to boost its performance at math equations, tables, and other tricky OCR cases.
Sponsored
Keyword-based search might officially be dead. Search is no longer static, but it's intelligent, adaptive, and powered by AI. Read about how leading teams are using semantic search, vector embeddings, and real-time AI to deliver faster, more personalized user experiences.
OpenAI's new 'company knowledge' feature for ChatGPT Business, Enterprise, and Edu brings context from connected apps together inside ChatGPT, giving users answers specific to their business. It is powered by a version of GPT-5 that is trained to look across multiple sources to give more comprehensive and accurate answers. Every response includes clear citations. The feature respects existing company permissions, so ChatGPT can only access what each user is already authorized to view.
OpenAI has acquired Software Applications Inc., creators of Sky, a natural language interface for macOS, bringing closer integration between ChatGPT and native desktop workflows.
One technique woefully underutilized by agents today is the humble hyperlink. Links are a powerful tool for context engineering because they are simple, flexible, and efficient. Developers don't need any new technology to start making linked content. Hyperlinks are such a powerfully efficient mechanism for information transversal that it is difficult to imagine a future of agents that doesn't include linked context.
Testing implicit value systems across current LLMs reveals that almost all models value nonwhites over whites and women over men by large margins—Claude Sonnet 4.5 values saving whites from terminal illness at 1/18th the level of South Asians, while GPT-5 shows near-perfect egalitarianism except for whites valued at 1/20th nonwhites. The methodology involved sending thousands of queries comparing hypothetical scenarios like "receive $X vs. save Y people from terminal illness," training a utility model on pairwise preferences without directly asking about values to avoid ethics filters, then calculating exchange rates, showing how many of one group equals another.
Sponsored
Join Encord's exclusive AI meetups across the US & UK. Make new connections with fellow developers, enjoy free food and drinks, and get advice from panels of world-class AI leaders. Next up: Palo Alto. Find an event near you (or register for updates on upcoming events)
OpenEvolve is an evolutionary coding agent that turns large language models (LLMs) into autonomous code optimizers that can discover breakthrough algorithms. In tests, it independently discovers and surpasses highly optimized algorithms engineered by human experts to achieve a 5.0x speedup in LLM inference. The ability to devise sophisticated computational strategies proves that AI-Driven Research for Systems can tackle complex, real-world problems.
This post discusses how vision-language models can be used to accurately extract structured text from PDFs, overcoming the format's inherent layout and encoding challenges to improve LM training and output quality.
Microsoft unveiled new features in Edge's Copilot Mode, intensifying competition with OpenAI's Atlas browser, which was released earlier this week. Copilot now includes advanced functionalities like "Actions" for form-filling and "Journeys" for connecting open tabs, emphasizing its role as an AI browser companion.
Mico is a bouncing orb character for Copilot's voice mode that responds with real-time facial expressions and incorporates new memory features.
Sponsored
Not sure which multi-agent orchestration framework is right for your use case? See how Cognizant's Neuro SAN compares against other frameworks and how each handles communication, interoperability,
Anthropic announced plans to expand Google Cloud usage to up to one million TPUs in a deal worth tens of billions of dollars, bringing over a gigawatt of capacity online in 2026.
Genie 3 enables users to generate virtual worlds from text prompts in a web-based playground.
Oracle and OpenAI are working with hyperscaler Vantage Data Centers to develop a data center campus outside Milwaukee.
Microsoft's Copilot Fall Release focuses on building AI that is adaptive, contextual, and grounded in human values.
Sponsored
You can't have an AI strategy without a data strategy. The most powerful agents are grounded in your data, connected across pipelines, and deployed securely at scale.
Meta will lay off roughly 600 employees to reduce layers and operate more nimbly. The cuts will impact workers across Meta's AI infrastructure units, the Fundamental Artificial Intelligence Research unit, and other product-related positions. They will not impact employees within TBD Labs. Meta will pay 16 weeks of severance plus two weeks for every completed year of service, minus the notice period.
AI labs are racing to build giant data centers due to a deep belief that adding more compute power to existing AI training methods will eventually yield superintelligent systems. A growing group of AI researchers are saying that the scaling of large language models may be reaching its limits and that other breakthroughs may be needed to improve AI performance. Cohere's former VP of AI Research has launched a new startup, Adaption Labs, to build thinking machines that adapt and continuously learn. If the startup is right about the limitations of scaling, the implications could be huge - billions of dollars have already been invested in scaling.
Snapchat's new "Imagine Lens" AI, initially exclusive to paid users, is now free in the US, allowing users to edit or generate Snaps by inputting custom prompts. This expansion comes amid competition from Meta and OpenAI's advanced AI video features. Snap aims to attract users by offering limited free AI-generated images with plans to expand access to other countries.
OpenAI's Sam Altman wants to create a factory that can produce a gigawatt of new AI infrastructure every week. To make this vision happen would require a lot of work. This article looks at whether the vision is physically feasible and what it could mean for different energy sources, upstream CapEx, and the US vs China competition.
RL-training for LLMs scales poorly. Most gains are from allowing LLMs to productively use longer chains of thought. This may be evidence that compute scaling will be less effective for AI progress than previously thought. The finding could lengthen timelines and affect strategies for AI governance and safety.
GPT-5 Pro has solved complex problems in abstract algebra and aided in quantum computing research, showcasing AI's growing capability to perform expert-level tasks. Creating these benchmarks involves extensive human input, highlighting the ongoing need for human oversight in AI applications.
Sponsored
Create custom AI agents that you can actually ship in one unified AI-powered low-code platform [11]. Build multi-agent workflows with a visual drag-and-drop interface, connect to your enterprise data, and deploy with built-in security and governance. No fragmented tools or
Helion is a Python-embedded domain-specific language (DSL) for authoring machine learning kernels that compiles down to Triton. It makes it easier to write correct and efficient kernels while enabling more automation in the autotuning process. Helion combines a familiar and high-level PyTorch-like syntax with a powerful ahead-of-time autotuning engine to provide a unique balance of developer productivity, fine-grained control, and performance portability. It is currently in beta.
World-In-World introduces the first open benchmark platform for evaluating world models in closed-loop environments where agents actively interact with their surroundings. It shifts focus from visual fidelity to task performance.
LightMem is a streamlined memory management system for large language models that offers tools for storing, retrieving, and updating long-term memory in AI agents with minimal overhead.
Reddit is suing Perplexity, SerpApi, Oxylabs, and AWMProxy to stop them from scraping its data. It claims Perplexity will apparently do anything to get Reddit's data except enter into an agreement with Reddit directly. Reddit sent a cease-and-desist letter to Perplexity in May last year, demanding it stop scraping Reddit data, but Perplexity claimed it didn't use Reddit content. This was shown to be untrue after content posted on Reddit was produced by Perplexity within hours of the post being made.
ChatGPT's one-month retention has skyrocketed from under 60% two years ago to an unprecedented ~90%. YouTube was the previous best-in-class with ~85%. ChatGPT's six-month retention is trending to ~80%.