Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
ShinkaEvolve, an open-source framework from Sakana AI, uses LLMs to achieve state-of-the-art sample efficiency in evolving programs. It significantly improves solutions in areas like circle packing, math competition problem-solving, competitive programming, and LLM training design. The framework's innovative methodologies include exploration-exploitation balancing, novelty-based program rejection, and dynamic LLM prioritization.
Google's CodeMender is an AI agent that automatically detects and patches code vulnerabilities using Gemini models. Alongside it, it launched the AI Vulnerability Reward Program and SAIF 2.0 to define secure-by-design principles for autonomous agents and boost AI-driven cybersecurity.
Researchers reverse-engineered why language models fail at simple 4×4 multiplication and found successful models naturally built tree-like attention patterns that "cache" intermediate calculations then retrieve them later, like jotting down a thought on scratch paper. Standard training fails because models learn the easy outer digits first, but get stuck on the middle digits. By adding a simple auxiliary loss that forces the model to predict running totals at each step, models dramatically improve.
A group of scientists released a collection of open problems in computer science in 2014. One of these was the Non-Interactive Correlation Distillation with Erasures. GPT-5 Pro found a counterexample to the problem. A screenshot of its answer is available.
Cerebras withdrew its IPO filing because it wants to come out with updated financial and strategy information. Some of the company's new investors weren't mentioned in last year's filing. The company still intends to go public. It will refile again with updated financials and strategy information, including its approach to the rapidly changing AI landscape.
Sponsored
Ad-hoc dashboard requests burn valuable resources for data teams. Conversational Analytics replaces these chores with natural-language
Sam Altman said Pulse has no current ad plans but left the door open to Instagram-style ads.
Deloitte is rolling out Anthropic's Claude to more than 470,000 employees across 150 countries, Anthropic's largest deployment yet.
Replit's new Connectors let apps link to services like Google, Salesforce, Notion, and more with one-click logins, no credential juggling or repeated setups.
The deal fell through, so Medal is now creating an AI lab called General Intuition to build AI models using its gaming videos.
Sponsored
Join the Amazon SageMaker Roadshow for advanced, hands-on development with real production-grade services—not boring basics or theoretical tutorials. Choose from multiple sessions in a city near you, covering: model training and fine-tuning, building an AI-ready data foundation, and deploying production AI agents. LEVEL UP YOUR AI SKILLS FOR 2026.
OpenAI's Agent Builder will help users build agentic workflows and connect MCPs, ChatKit widgets, and other tools. It is a direct competitor to established workflow automation tools like n8n and Zapier. The Agent Builder features a drag-and-drop canvas that allows users to create agent flows from predefined templates. The canvas supports a range of modular building blocks, nodes for logic, connectors, and more, mirroring the flexibility seen in other agentic workflow platforms.
Google appears to be testing a new home screen for Gemini. The redesign adds a scrollable feed of one-tap prompt suggestions, with prompts ranging from image edits to requests regarding news, quizzes, and coding. Video and screenshots of the new home screen are available in the article. There is no guarantee that the redesign will roll out broadly.
OpenAI has acquired Roi, an AI-powered personal finance app. Only the company's CEO will be joining OpenAI. Roi will wind down operations and end its service to customers on October 15. Terms of the deal have not been disclosed. The acquisition clearly aligns with OpenAI's bet on personalization and life management as the next layer of AI products.
OpenAI and Jony Ive's partnership is struggling with some technical issues that could end up pushing back their highly anticipated AI device's release date. OpenAI and Ive are still working on the assistant's personality and potential privacy concerns. The project's budget could reportedly be challenged due to the increased computing power necessary to run mass-produced AI devices. There is still little information about the upcoming product publicly available.
Understanding format sensitivity in large language models is crucial for data pipeline architecture, performance optimization, and cost management. Many RAG pipelines involve ingesting documents that contain tables of data. These tables need to be formatted in a way that is easy for models to consume otherwise they may be needlessly hurting the accuracy of the overall system. Markdown-KV appears to be a good default to use in situations where accuracy is paramount. CSV and JSONL can hurt system accuracy.
Analysts immediately drew comparisons to the telecom bubble on the news that Nvidia had invested $100 billion in OpenAI in September. During the telecom bubble, demand was speculative, and customers burned cash. Many of Nvidia's customers are profitable and sophisticated hyperscalers. This suggests that the current situation is different from the telecom bubble.
Most economists refuse to discuss the possibility of transformative AI, even at workshops designed for it, insisting that comparative advantage keeps human wages high, as past tech progress has shown. AI's impact depends on domain structure, like protein folding has predictable patterns, but mapping genomes requires individual observations.
Claude Sonnet 4.5 has been enhanced with strong cybersecurity capabilities, matching or surpassing previous frontier models in tasks like vulnerability detection. This shift reflects a growing focus on using AI not only for cyber offense but also to assist defenders in securing systems.
Sponsored
Starting tomorrow, stream the Opening General Sessions to see how GENERATIVE AI AND AGENTS are transforming enterprise tech.
The Jules API gives developers programmatic access to Jules' capabilities to automate and enhance software development cycles. It can be used to create custom workflows, automate tasks like bug fixing and code reviews, and embed Jules' intelligence directly into everyday tools. This page provides a quick overview of the API and walks through how to make an API call.
Mira Murati's AI lab, Thinking Machines, has released a language model fine-tuning API called Tinker. Tinker exposes a handle of very low-level functions to users to let them perform supervised fine-tuning and online reinforcement learning. Each training step requires sending a batch of data over the network. It offers a stable, mature infrastructure for RL and fine-tuning experiments, making it possible for anyone to be an AI researcher.
SINQ (Sinkhorn-Normalized Quantization) is a novel, fast, and high-quality quantization method designed to make any large language models smaller while keeping their accuracy almost intact. It uses dual scaling to make models less vulnerable to outliers. SINQ's Sinkhorn-normalized optimization, which iteratively rescales rows and columns to balance their variance, makes errors more spread out and less severe, preserving model accuracy even at 3-bit precision. Weights become inherently easier to quantize when the overall matrix imbalance is reduced, leading to more stable behavior across layers and consistently higher accuracy even at very low bit-widths.
Stan's days are likely numbered. High-end applications are replacing Stan with JAX. The biggest obstacle for the shift is finding hardware on which to run JAX most effectively, but this won't be a problem in ten years. The Stan project will continue development, as it will likely still be used for a long time. The team has plenty of strategies for making it faster, like adding samplers that will work well on CPU but not on GPU.
dots.ocr was adapted to run entirely on-device using Apple's Core ML and MLX frameworks. In this post, Hugging Face outlines the conversion process, offering a practical guide for developers looking to deploy similar models locally.
The Claude team recently gave out hats and coffee at an event where people had to wait two hours in line. This post details how the event went down. Participants at the event waited, took pictures, posted on social media, and became one of the cool kids. The Claude team created an experience people wanted to be part of, which made it worth waiting for hours just for a hat. Pictures of the event and the merch the Claude team handed out are available in the article.
Keith Heyde, OpenAI's Head of Infrastructure, leads site development within OpenAI's industrial compute team, a division that's quickly becoming one of the most important groups in the company. Infrastructure is now a strategic pillar on par with product and model development. OpenAI is betting that owning the next generation of physical infrastructure is central to controlling the future of AI. This article looks at how Heyde has spent the past several months searching for Stargate data center sites and the challenges that face the project.
Sponsored
CompactifAI offers highly compressed versions of leading LLMs, reducing costs and improving throughput with negligible quality loss. Deployable anywhere — cloud, edge, on-premises — enabling seamless integration into any workflow. Explore CompactifAI
The company offered no explanation for its decision, but the move came days after raising $1.1 billion from private markets.
Google, Meta, and Amazon operated as a bloc pushing for higher revenue thresholds to limit which companies would be covered, and a16z and OpenAI suggested waiting for federal frameworks.
GPT-1-thinking performs exceptionally poorly compared to today's cutting-edge models.
Sam Altman is looking to East Asia and the Mideast for funding and manufacturing partners for his plan to build more data centers.
Sponsored
Generic AI code review catches typos, style issues, and things that _might_ be a problem. Sentry's AI code review [4] (now in beta) catches the bugs that actually take down production.
OpenAI has reached a $500 billion valuation following a $6.6 billion secondary share sale by employees. It is now the world's most valuable startup despite not yet turning a profit.
Perplexity has officially launched the Comet browser globally after a high-demand beta phase. Comet is an AI-native browser designed to reshape online interaction.
Former Stripe CTO Rahul Patil has joined Anthropic as chief technical officer, succeeding co-founder Sam McCandlish, who now becomes chief architect. The leadership change aligns with Anthropic's effort to scale infrastructure and product engineering amid rising demand for Claude and competition from Meta and OpenAI.
Today's frontier LLM research isn't about building animals, it's about summoning ghosts. Ghosts are a fundamentally different kind of point in the space of possible intelligences, thoroughly engineered by humanity. Over time, we may be able to fine-tune other ghosts more and more in the direction of animals, but it's also quite possible that they diverge even further and end up permanently different. They may end up un-animal-like, but still incredibly helpful and properly world-altering.
Most of the top 50 AI-native application layer companies are planning to ramp AI spend. Those using AI are seeing a higher ROI than with traditional tools. AI is not only augmenting employees in specific roles, it is also turning specific roles into broadly deployed skills across companies.
The model identifies evaluation scenarios 13% of the time and behaves unusually well afterward. Mechanistic interpretability showed internal representations of "fake or suspicious content" and "AI safety" concepts strengthened during training. When researchers suppressed these features, awareness dropped while misaligned behaviors increased.
Sponsored
PyTorch Conference is the premier forum for AI innovation. Join 2,500+ engineers and researchers for 6 tracks, 3 summits, training, a Startup Showcase, and more. Whether scaling a startup or advancing AI, this is where to connect with the people, tools, and insights shaping
Gemini 2.5 Flash Image is available via the Gemini API, Google AI Studio, and Vertex AI. It offers 10 aspect ratios, image blending, low latency, and competitive pricing. A production-ready image generation and editing model, Gemini 2.5 Flash Image is generally available and accessible for global use wherever Google platforms operate. Google's new pricing structure maintains competitive rates for enterprise adoption at $0.039 per image and $30 per million output tokens.
This is a public version of an internal onboarding guide at Cursor provided to GTM and non-engineering hires. It walks through how to get started from scratch to a built-out, deployed project. Each section includes a short video guide.
Jules is an asynchronous coding agent that integrates directly with existing repositories. It understands the full context of projects and can perform tasks like writing tests, building new features, providing audio changelogs, fixing bugs, and bumping dependency versions. Jules Tools is a lightweight command-line interface that allows developers to spin up tasks, inspect what Jules is doing, and customize the agents, without leaving their workflows. It is both a dashboard and a command surface for coding agents.
Vision-Zero is a domain-agnostic framework for training vision-language models via competitive visual games generated from arbitrary image pairs.
DeepMind's AlphaEvolve uses LLMs to iteratively “evolve” code that discovers mathematical structures proving new theorems in complexity theory. The system made breakthroughs on two mature problems by morphing finite proof structures while keeping the interface to the broader proof intact, allowing automated verification of correctness without human review.