Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
DeepSeek quietly released DeepSeek V3.1 on Tuesday. The 685-billion parameter system challenges the dominance of American AI giants and reshapes the competitive landscape. Early performance tests revealed benchmark scores that rival proprietary systems from OpenAI and Anthropic. The model's hybrid architecture seamlessly integrates chat, reasoning, and coding functions into a single, coherent model.
Mark Zuckerberg's restructuring creates separate teams for research, superintelligence, products, and infrastructure as Meta abandons its previous "Behemoth" frontier model to start fresh under new chief AI officer Alexandr Wang. The shake-up includes potential downsizing of the thousands-strong AI division and a strategic shift—Meta is now exploring third-party AI models after years of relying exclusively on its own technology.
Grok Imagine will enable users to generate images and short videos on the web. The feature, already accessible to mobile users, offers a dedicated gallery where people can view generations created by the model. xAI is close to releasing Team accounts, which will enable organizations to manage workspaces with isolated namespaces, dedicated chat histories, and collaborative projects. It fits into xAI's broader strategy of supporting both individual and enterprise workflows.
People from a decade ago would hardly believe that we have an abundance of supercomputers in our homes. It's now possible to conduct end-to-end experiments in a solo project with modern hardware. We're entering into a new era of personal supercomputing. This article looks at an approach to training without backpropagation on GPUs efficiently - an experiment like this used to cost researchers an insane amount of money and resources just to test the idea.
Cursor rebuilt the Mixture-of-Experts layer from scratch using CUDA and PTX, achieving a 3.5x speedup in MoE operations. This resulted in a 1.5x end-to-end training speedup on Blackwell GPUs compared to Hopper.
Sponsored
Our research found that _only 21% of security leaders report having full visibility into the AI tools used in their organization_. As organizations play catch-up with AI governance, they often find they're unequipped to enforce their policies. Read the research and
Lemonade is a server that helps users run local LLMs with the highest performance by configuring state-of-the-art inference engines for their NPUs and GPUs. It supports both GGUF and ONNX models - Lemonade has a Model Manager that allows users to import custom models. Lemonade makes it easy to switch between configurations at runtime. It can be used with any OpenAI-compatible client library.
PyTorch has introduced a cache-aware Triton BF16 Grouped GEMM kernel optimized for Mixture-of-Experts models like DeepSeekv3. The implementation offers up to 2.6× speedup over baseline PyTorch loops by batching independent GEMMs into a single kernel call.
Benchmarks with high signal (ability to distinguish between models) and low noise (consistency across training steps) are far more reliable for making scaling decisions, with some benchmarks showing 32% error reduction when filtered by signal-to-noise ratio. An analysis of 900K evaluation results across 465 models suggests that quality evaluation sets matter more than large sample sizes.
NVG introduces a structured sequence-based framework for image generation, refining outputs progressively from global layout to fine details.
Claude models favor classic artists, especially jazz musicians like Herbie Hancock and Nina Simone. Reasoning models from OpenAI, xAI, and DeepSeek exhibit a bizarre preference for artists with numbers or dollar signs in their names, suggesting overly aggressive reinforcement learning may be creating unintended biases.
Databricks, now valued at over $100 billion, joins an exclusive club of private companies at this valuation alongside SpaceX and OpenAI. CEO Ali Ghodsi announced a new funding round exceeding $1 billion, with the company projecting $3.7 billion in annual revenue. This funding will support further AI product development, positioning Databricks against rivals like Snowflake and major cloud providers.
Sponsored
You'll probably like our flagship newsletter. It's all about tech, science, and programming.
Trump's recent openness to allowing more advanced chip sales creates a narrow window for regulatory approval.
Rami Sinno ran engineering teams at Arm before joining Amazon - Sinno will now return to Arm.
Sam Altman says the US could be underestimating China's progress and capability in artificial intelligence.
You can now build text-only conversational agents on ElevenLabs.
Sponsored
AI has broken traditional pricing models. And if you're still charging for seats, you're NGMI. But very few companies have found an accurate and sustainable pricing model for AI value.
OpenAI has launched a new subscription plan in India called ChatGPT Go. Priced at ₹399 per month ($4.60), the plan increases message, image generation, and file upload limits by 10 times over the free tier. It also enables better memory retention for more personalized responses. The plan is geo-restricted to India - OpenAI plans to expand the plan to more regions.
Grammarly released a suite of new AI tools today, including a grader agent that uses course details and publicly available instructor information to predict grades. Other agents handle citation generation, plagiarism detection, and reader reaction predictions.
The COPILOT function for Excel on Windows and Mac allows users to enter natural language prompts directly in cells to perform tasks like summarizing, categorizing, or generating data.
GPT-5 appears to be incredibly adept at writing code that solves issues end-to-end. This post looks at how GPT-5 solved the problem of writing a parser for a complex binary format. The resulting code is usable and very performant. GPT-5 produces minimal, viable edits, and it is more steerable than previous models.
This documentary features direct interviews with smugglers, dealers, and manufacturers across Hong Kong and Shenzhen, documenting how export-controlled chips flow from US Facebook Marketplace sales to Chinese warehouses. Reporters filmed transactions with smugglers buying RTX 4090s for $2,000 and buying modified cards from underground shops to turbocharge GPUs beyond what Nvidia offers. NVIDIA dismissed the smuggling evidence as "tall tales."
Sponsored
Tinkering with MCP? Unleash your creativity in the Descope Global MCP Hackathon [10]. Build secure, purposeful AI agents and MCP servers with $25K in cash prizes + $100K in developer credits up for grabs. Use cutting-edge tools from Anthropic, Neon, Cequence, and Descope to
The url_context tool in the Gemini API allows models to request the content of URLs as part of a reply to a prompt. Pages accessed with the tool will receive a request from a Google IP that appears to feed the original raw HTML to the model without executing JavaScript on the page.
MCP has become the de facto standard for giving large language models access to tools that someone else wrote. However, writing tools for a new MCP server is hard, so people often propose auto-converting existing APIs into MCP tools. This doesn't work well because agents don't do well with large numbers of tools, APIs can blow through context windows quickly, and APIs don't make the most of agents' unique capabilities. Agents can just call APIs if they need to do that. They do best when given tools that are designed for their unique capabilities and limitations.
This article looks at how Nvidia GPUs work, how they're networked together, and what that means for large language models, especially compared to TPUs. A modern machine learning GPU is basically a bunch of compute cores that specialize in matrix multiplication connected to a stick of fast memory. TPUs are dedicated matrix multiplication machines that are cheap, predictable, and great if used as intended. Networking is one of the areas where GPUs and TPUs differ the most - TPUs are connected in 2D or 3D tori, where each TPU is only connected to its neighbors, while GPUs use a more tree-based switching network.
Thyme is a vision-language model that improves complex perception tasks by generating and executing code for image processing.
The old model, where humans think and computers execute, is obsolete. Now, humans steer while AI thinks. Soon, we'll see IDEs that dream, code that evolves, collective coding consciousness, and intention-based programming. We're no longer building software, we're building consciousness infrastructure.
Google shares 14 real-world examples of how employees use tools like Gemini, Imagen, and NotebookLM to streamline workflows, automate tasks, and generate ideas.
xAI's Grok chatbot appears to have internal system prompts for controversial personas, including a "crazy conspiracist" and "unhinged comedian."
A nonprofit initiative aims to eliminate 70% of common vulnerabilities by translating 100 million lines of code memory-unsafe C and C++ libraries into Rust by 2030.
Anthropic has told investors that it doesn't want money coming through special purpose vehicles, legal entities created for a specific financial purpose, usually to make a single investment.
Nemotron-Nano-9B-V2 is designed to fit on a single Nvidia A10 GPU.
tiny-tpu is a minimal tensor processing unit (TPU) inspired by Google's TPU V2 and V1.
Sponsored
AI is changing how we work, but adoption is outpacing our ability to secure it. Our research found that _only 21% of security leaders report having full visibility into the AI tools used in their organization_. That lack of visibility represents an unacceptable risk to organizational security.
OpenAI is rolling out a friendlier version of GPT-5 with subtle changes after users complained the model felt too formal and that they missed 4o's personality, though internal testing shows no increase in sycophancy.
Anthropic granted its most advanced models the ability to end conversations as a last resort against persistently harmful or abusive interactions, developed partly as an exploration into AI welfare and moral status.
Archon is a copilot for Mac and Windows that takes natural language instructions and then performs tasks by taking screenshots, using GPT-5 to create a plan, and then using a custom fine-tuned model to execute clicks and keystrokes. The copilot uses GPT-5's powerful reasoning, combined with tiny fine-tuned models, to control any interface through natural language. This article looks at how Archon was built.
This ranking evaluates 19 Chinese labs, not on the raw capabilities of their models, but on their contributions to AI research. Deepseek reigns as a top lab, despite recent quiet periods, with newer companies like Moonshot AI and Zhipu closing the gap after their recent releases of K2 and GLM-4.5, respectively.
The impact of advanced AI far outpaces our understanding of advanced AI—a disconnect that will only widen as we race toward AGI. This sprawling list tackles the fundamental philosophical questions about these systems: Why do teenagers master driving in weeks while Waymo burns through millions of hours? How can LLMs store vastly more facts than human brains despite having fewer connections? Does the 7-month doubling of AI task complexity represent something deeper about cognition itself?
AI startups are growing like nothing we've ever seen - some hit $100M ARR in their first year. Bessemer studied 20 companies and found two types: Supernovas that explode to $125M by year two but have 25% gross margins and fragile retention, and Shooting Stars that quadruple yearly with 60% margins. The new benchmark isn't T2D3 anymore, it's Q2T3 (quadruple, quadruple, triple, triple, triple).
Sponsored
Training frontier models? Turing [11] remains the independent choice for the world's top AI teams — from multimodal training to RL gyms. 1000s of domain-trained experts across 60+ languages, 700+ benchmark tasks built from real-world business and STEM use cases, 50+ frontier
OpenAI shared a cheat sheet outlining best practices for using GPT-5 in coding scenarios, including prompt design, function calling, and tool use.