Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Mini-o3 is an open-source model designed for agentic visual reasoning. It supports multi-turn interactions similar to OpenAI's o3, and scaling to tens of rounds. The full training pipeline is public.
While AI has the potential to improve our lives in a lot of ways, it is important to be wary of secondary effects and unintended consequences of the technology. For example, vibe coding really speeds up iteration loops and helps developers explore more ideas, but it removes previous design processes, which can change things in ways that are hard to predict. Sometimes the friction and inconvenience is where the good stuff happens, so be very careful about removing it.
Claude autonomously developed a programming language where "slay" means function, "sus" means variable, and pointers are represented by the Among Us character. The "cursed" language compiles to actual binaries and includes editor extensions.
Sponsored
Imagine agents becoming our "collective subconscious," executing digital drudgery while we focus on higher-level thinking. See how Amazon's AGI SF Lab is revolutionizing human-agent collaboration through Nova Act.
A federal judge paused Anthropic's $1.5 billion book piracy settlement, citing concerns about the terms that give authors $3,000 per allegedly pirated book.
Developers can now create connectors and use them in chat for write actions in developer mode.
Leaked code strings suggest Claude will soon automatically incorporate information from past conversations without being explicitly asked and will support an “incognito mode” that would prevent chats from being saved, added to memory, or used for model training.
The Gemini Batch API now supports the newly launched Gemini Embedding model and the ability to leverage the OpenAI SDK to submit and process batches.
Sponsored
Snowflake's usage-based pricing has been one key driver of its market leadership. In this webinar [4], you'll get an insider view of how the team managed its pricing and billing, and how they're preparing for the Al era of monetization.
Claude can now generate and edit documents, spreadsheets, slides, and PDFs directly in the app, allowing users to turn prompts and data into downloadable files.
Claude will be integrated into Office 365 apps, ending Microsoft's exclusive reliance on OpenAI. OpenAI has continued to put strain on its relationship with Microsoft by announcing competitive products, including an AI-powered LinkedIn rival and custom chips built by Broadcom.
Nvidia has announced a new GPU called the Rubin CPX designed for context windows larger than 1 million tokens. The GPU, meant to be used as part of a broader 'disaggregated inference' infrastructure approach, is optimized for the processing of large sequences of context. It performs better on long-context tasks like video generation and software development. The Rubin CPX will be available at the end of 2026.
OpenAI's Responses API replaces the previous /chat/completions API for inference. The new API has a lot more features, but the main difference is that it is stateful. Users no longer have to pass the entire conversation history with each request, they just have to pass an ID representing the state of the conversation and the provider keeps models up to date. This allows OpenAI to keep its reasoning traces secret.
Every serious AI company will eventually train its own models. The barrier to doing so is collapsing. Distillation, fine-tuning, and post-containing get easier every month. Soon, the only way to stay relevant will be to own your own models.
Production monitoring reveals real issues that pre-deployment evals will inevitably miss, especially as AI products become more unpredictable and personalized. Evals are collections of already-known failure cases, but agents can and often do fail in ways that produce no error codes.
Sponsored
Get your data ready for the future of analytics and AI at CData Foundations 2025 [12]! One free registration gets you two virtual summit days - Analytics on September 17 and AI on September 24. Join for expert insights, real-world success stories, and practical strategies to modernize your data architecture. SAVE YOUR SPOT (FREE)
ByteDance's REER (REverse-Engineered Reasoning) is a paradigm that derives step-by-step reasoning from known good answers rather than building it forward via RL or imitation.
The Model Context Protocol Registry provides a standardized way to distribute and discover MCP servers. A community-driven project, it allows organizations to create private enterprise registries while maintaining compatibility through shared API schemas and moderation guidelines.
Gemini Web has rolled out its Tools redesign. The prompt bar is now centered with a new 'Tools' menu that lets users access Deep Research, Veo, image creation, Canvas, Guided Learning, and Deep Think. Gemini now groups suggested prompts together for a cleaner look. Screenshots of the new interface are available in the article.
AI companies face varied gross margins across sectors. Chips maintain around 70% margins, while cloud services see margins pressured by AI investments, estimated between 50-55%. Application-level margins range widely, with AI "Supernovas" starting at 25%, potentially negative, while others hit 60%, highlighting pricing strategies and diversified revenue models to improve margins over time.
Sponsored
Warp Code adds a built-in editor and code review to proven CLI and MCP tools, bringing everything together in one powerful app. Top-rated
OpenAI announced a $50 million fund to support nonprofits working on AI literacy, community-driven innovation, and equitable economic opportunity.
The main question is whether smaller models will continue to lag closed source models by 6-12 months, or if the gap will expand as capabilities advance.
Nvidia CEO Jensen Huang predicts AI and robotics will increase human workloads, countering Elon Musk's vision of more leisure time.
Sponsored
Our latest releases for Redis for AI are full of core enhancements, new integrations, and dev tools to help you build fast, reliable apps at scale—especially AI ones. Here's what's new [4]:
Some of California's biggest philanthropies, nonprofits, and labor groups are joining forces to push back on OpenAI's plan to restructure into a for-profit company. This has concerned OpenAI executives and resulted in discussions about moving out of the state. OpenAI's proposed plan is being investigated by Attorneys general in California and Delaware. The startup could potentially be sued for breaking nonprofit law and may be required to pay a settlement as a condition for the restructure.
Databricks has confirmed another $1 billion raise at a $100 billion-plus valuation. The company will use the funds to invest in its database for AI agents. The round was led by Thrive and Insight Partners, who also co-led the previous $10 billion raise. Databricks has $4 billion in annual recurring revenue.
Google reduced Veo 3 pricing by nearly half to $0.40 per second and added support for vertical 9:16 videos and 1080p HD output.
Grok is working on a background thinking feature that appears to allow users to continue chatting while the assistant thinks. A screenshot of the toggle for the feature is available.
RLaaS (Reinforcement Learning-as-a-Service) has proven performance, lower costs, is more excludable, and is safer. AI companies pivoting away from pursuing AGI towards RLaaS is good news because it promises a switch to a safer mode of AI development. While RLaaS is not without risks, the transition to an RLaaS model would reduce AI risk. The concatenation of specialized models may create or assist general intelligences.
Shawn Wang, founder of Smol AI, is joining Cognition, the makers of the Devin AI software engineer. The vast majority of Wang's equity/investor capital at Smol AI will be rolling over their investment to Cognition as part of the deal. Wang will continue to work on AI Engineer and Latent Space independently of Cognition. This post explains in detail Wang's reasons for betting on Cognition.
Anthropic details why it implemented ASL-3 safety protections against biological weapons development, noting Claude now exceeds expert performance on virology evaluations, and controlled trials showed significant uplift with Opus 4 for acquiring bioweapons compared to using just the open internet. It's obviously a low-probability risk, but the impact is so high — just one infraction could mean thousands of deaths — that it still requires serious evaluation and restriction.
Sponsored
Effortlessly deploy agentic workflows and generative AI with the Inference Cloud powered by Qualcomm. Designed for today's leading models, it offers exceptional performance and seamless scalability at a fraction of the cost.
AI coding agents are great, but they're too wasteful, they don't hold the context of the database, they only have a surface-level understanding, and they're too susceptible to context rot. These problems could be solved if agents had a semantic understanding of all functions and context about how functions are connected to other parts of code. This article looks at practical ways to make coding agents more efficient.
Set Block Decoding enables parallel sampling of multiple future tokens by combining next-token and masked token prediction. This reduces decoding passes up to 5x without sacrificing accuracy.
Reinforcement Learning environments are sandboxes for AI models to learn from. They keep track of an internal state, prompt AIs to take actions to complete a task, and assign scores based on the outcome. More and more startups are offering increasingly challenging environments under the assumption that machine learning is solved and the only remaining work is to provide models with simulations to learn from. Frontier AI labs pay a lot of money for these environments, but the supply side is set to become increasingly crowded - it likely won't be a long-lived business.
A federal judge expressed serious concerns about Anthropic's proposed settlement over the use of 500,000 pirated books to train an internal version of Claude. He questioned how the settlement money would be distributed: “I have an uneasy feeling about hangers on with all this money on the table.”
Sponsored
Ready to build enterprise-ready AI agents that can handle complex tasks without breaking? Get your hands dirty with production-grade frameworks and LLMs in free builder workshops.
Nebius will provide Nvidia GPUs and AI cloud services to Microsoft out of their new New Jersey data center over 5 years.
Perplexity introduced a secure, zero-data-usage AI platform for U.S. government agencies, alongside a $0.25-per-user Enterprise Pro offering tailored to public sector needs.
Anthropic endorses California's SB 53, a bill mandating transparency in safety practices for frontier AI systems.
Cognition raised over $400M at a $10.2B valuation, in a funding round led by Founders Fund with participation from Bain Capital Ventures and others.
Sponsored
GitHub Universe returns October 28–29 in San Francisco, and the full agenda is now live! Early Bird pricing ends September 17, so this is your last chance to save $400 [4]. This year's focus is AI agents, automation, and the tools reshaping how we build software.
A federal judge ruled in June that it was legal for Anthropic to train on copyrighted materials but that it was illegal to acquire the materials from pirate sites like LibGen. Anthropic has offered to settle the class action suit from authors by offering $3,000 per work.