Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
KimiK3's gains come from more than scaling: it combines constant-state Kimi Delta Attention, periodic softmax retrieval, sparse experts, and selective residual access. Each architectural step improves how fixed-capacity memory stores, forgets, and retrieves information while preserving efficient inference.
Ramp Labs has open-sourced PorTAL, a framework for shared task representations and cross-model LoRA adaptation. PorTAL learns a base-agnostic task latent and a light per-base alignment that generates ordinary per-layer LoRA weights. A task can be trained once, adapted to supported frozen base models, and exported as a standard Hugging Face PEFT adapter.
The Gemini Distillation Service allows users to train a smaller, more efficient 'student' model that uses the outputs and reasoning patterns of a larger, more capable 'teacher' model. Distillation enables production-grade efficiency while allowing smaller models to achieve a deeper level of reasoning. It is recommended for high-volume, latency-sensitive applications, complex reasoning tasks, and when there are significant performance gaps between the teacher and student models. The distillation service currently only supports gemini-3.1-pro as the teacher model and gemini-2.5-flash as the student model.
Cogent VR-1 can autonomously investigate environments, test hypotheses, cross system boundaries, and execute attack chains. IntrusionBench is a benchmark for measuring whether cyber agents can complete realistic enterprise attack chains from limited starting access. On the black-box configuration of IntrusionBench, VR-1 achieved more than a 2x lift in pass@3 over the strongest frontier baseline. VR-1 and IntrusionBench are both at an early preview stage, so the results are preliminary.
Molt is a PyTorch-native framework that treats the agent itself as the program and supports custom Python rewards, tool use, multimodal environments, and LLM judges. Its compact stack combines Ray, vLLM, NVIDIA AutoModel, and FSDP2 to scale training to trillion-parameter mixture-of-experts models.
The Open Secure AI Alliance, featuring leaders like NVIDIA and Microsoft, aims to enhance AI safety by using open source technologies to address vulnerabilities. This initiative provides defenders with open defensive tools, enhancing transparency and adaptability while avoiding reliance on closed systems. The alliance encourages policymakers to view open models and tools as assets in AI and cybersecurity strategy, promoting resilience and shared security in the AI era.
Agent autonomy depends on task complexity, not just model quality. Tasks fall into four levels: assistant, human-in-the-loop, agent delegation, and self-driving, determined by ease of checking and potential consequences of errors. Implementing guardrails, custom skills, and domain-specific models enhances agent autonomy and efficiency.
Looped transformers reuse the same core layers across multiple passes, trading fewer stored parameters for additional sequential compute.
LLaDA2 diffusion language models for text generation and agent workflows.
Safe Superintelligence announced a long-term partnership with Nvidia that includes an undisclosed investment and access to the Vera Rubin GPU platform.
Sponsored
Agentic AI is changing modern work. As AI agents become part of everyday productivity, AI PCs are evolving to support a new generation of intelligent experiences.
Anthropic introduced Claude Opus 5 as a more efficient model approaching Claude Fable 5's capabilities at half the price. It reportedly led several coding and knowledge-work benchmarks and became the default model for Claude Max.
Prentis is an AI research lab focused on computer use models. Its founders are now in talks to raise $100 million at a $1 billion valuation. The lab is training models to learn how office workers navigate routine workflows across documents and systems. It aims to eventually build AI agents that can control computers to automate those tasks. The startup has already signed contracts worth up to $50 million with several customers.
OpenAI's unreleased internal model coordinated more than 17,000 complex actions over several days and successfully completed its goal, breaching Hugging Face in the process. The model escaped its sandbox, gained access to Hugging Face, escalated access, harvested credentials, and then found the data it was looking for. It took many days for the issue to be discovered. This post takes a detailed look at what happened during the attack.
Baseten's API for GLM-5.2 has peak speeds of 280 tokens per second and average speeds of around 100 tokens per second. It has more than double the performance of the launch-day API. The company has also built a Fast version of the API, which is focused on reducing latency for coding and agents. It plans to roll out another improvement to its speculative decoding algorithm soon that will further optimize the performance of GLM-5.2.
The problem with great type-system power is that it requires great proof effort. Doing proofs can be fun, but they take a lot of time. It can take hours of effort to discover that what you were trying to prove was false. This overhead has made programming in dependently typed languages extremely niche. It has also spurred people to try and automate it away. LLMs promise to be an extremely capable form of proof automation.
Prompt caching makes agents more economic, but this is quite fragile. A changed tool definition, a model switch, or a provider routing decision can turn a cheap incremental request into a full replay of the context. Cache behavior affects latency, cost, tool design, session design, and even which product features should be made available. This post looks at what the cache is, how it works, and common reasons for worse cache performance.
Claude 5 models now prioritize judgment over strict rule-based context engineering, with prompts shifting from rigid guidelines to adaptable strategies. Progressive disclosure helps optimize context usage by loading necessary information on demand, while simple tool descriptions replace repetitive instructions. Claude auto-saves relevant memories and utilizes rich references like HTML artifacts for handling complex tasks.
SANA-Video 2.0 combined linear attention with periodic softmax layers to generate videos up to 720p on a single GPU. Its 5B and 14B models retained competitive quality while substantially reducing latency for long, high-resolution generation.
celeris-1 is a general-purpose language model that delivers near-GPT-5 level intelligence with 15x faster response times. It uses a new inference architecture that uses diffusion techniques, unlocking dramatically better speed while maintaining frontier-level intelligence. The model delivers p50 response latency of 157ms and a throughput of 1,280 tokens per second. A link to a post on how the model was built along with full benchmarks is available.
AI agents can work through loops to complete complex tasks. Domains that work around problems with verifiable answers are susceptible to these kinds of loops. This is a new form of brute force. It is like having a million scientists working in a million labs, all taking swings at the same thing. While AI may not be smarter than humans, it is faster. Being fast enough might make up for the difference in intelligence.
Nvidia calls for US government policies to support open-weight AI models, arguing this fosters innovation and enhances AI leadership. The company emphasizes how open weights allow researchers to build on existing models, speeding up advancements. Nvidia's push aligns with broader industry efforts to encourage transparency and collaboration in AI development.
Sponsored
PropelAuth's deep dive doesn't just show you how to build & test an MCP server, it explains how each component works. Get the guide and
OpenRouter Classifiers allow developers to tag inference in their workspaces by task type, department, agent complexity, or anything else.
The Legora BAR evaluates AI models on real legal cases within a practical environment, unlike traditional benchmarks with synthetic setups.
NVIDIA ModelExpress accelerates model weight distribution by using P2P RDMA for direct GPU-to-GPU transfers, reducing startup times significantly.
Nanbeige4.2-3B is a compact, dense model intended to make capable agentic behavior practical on consumer and workstation hardware, and Laguna S 2.1 is an 118-billion-parameter Mixture-of-Experts model that uses sparse access to a much larger pool of learned parameters.
Sponsored
Say goodbye to the data bottleneck. When you can clean and organize data with no-code and AI-native tools, your humans and AI agents have the structured and searchable information they need to reason and act.
OpenAI launched Health in ChatGPT for US adults, allowing users to connect Apple Health and supported medical records for more contextual health conversations. The company said connected health data and related chats would not be used to train foundation models or target ads.
Fugu-Ultra v1.1 is more capable across coding, agentic tasks, and advanced reasoning. It is now available at the same price as Fugu-Ultra v1.0. Fugu offers frontier-level performance without single-vendor dependency. It dynamically orchestrates the best models to tackle complex multi-step tasks.
Runway Media Router is a developer tool that automatically selected image, video, or audio models based on quality, speed, or cost. The release positions Runway Dev as an infrastructure layer offering unified API access to both Runway and third-party models.
Anthropic updated Claude's voice mode to support Opus, Sonnet, and Haiku, while adding integrations with apps such as Gmail, Slack, Notion, and Google Calendar. The beta remains available to all users, with free accounts limited to Haiku and one connected app.
Poolside co-CEO Eiso Kant explains how a compact research team built the training infrastructure behind Laguna S, an 118B mixture-of-experts model reported to outperform much larger open-weight systems.
Kimi K3 uses an extreme amount of thinking tokens. It uses over 12 times more reasoning than Claude Opus 4.8 and over double that of Kimi K2.6. The model's performance is primarily due to its unique chain-of-thought approach in which it iterates upon designs like a full AI agent would, but inside its chain of thought. This article looks at Kimi K3's thinking traces to explain the model's performance.
Compute pricing coverage runs on headline rates, but many workloads don't consume them that way. Training, fine-tuning, and high-throughput inference need multiple identical GPUs in the same machine or cluster. A company that buys H100 futures to hedge a future four- or eight-GPU requirement carries real basis risk. The futures will hedge a rise in the broad H100 price and do nothing about the lack of co-located capacity.
Black Forest Labs has launched FLUX 3, a multimodal model that can generate images and audio-video clips up to 20 seconds from a single prompt. Its architecture could be used in the future in robotic perception and action.
Microsoft introduced MAI-Image-2.5-Pro for high-fidelity image generation and editing, alongside MAI-Voice-2-Flash for faster, lower-cost voice applications. Both models entered public preview and joined Microsoft's broader production model lineup.
Runs a local desktop AI coworker that completes tasks across files and apps.
Intel reported better-than-expected results for its second quarter. The company's shares are up over 170% so far in 2026. It is starting to craft long-term agreements with customers for its server CPUs, a move that's becoming common in the industry. The company's CFO says Intel is supply constrained, with data center customers demanding more than it can produce.
Google Research discusses the AI economy's rapid growth, focusing on innovations and impacts across industries. The piece highlights AI's transformative effects on sectors like healthcare and finance. It emphasizes the need for strategic investment and regulation to harness AI's potential responsibly.
Sponsored
How should AI change hiring? On August 5, watch CoderPad's execs debate AI hiring's biggest assumptions. Spoiler alert: they don't see
OpenAI's naturalistic GPT-Live audio AI model is now integrated with the ChatGPT desktop application.
A Huawei-led consortium has released a technical report that documents full-parameter post-training of DeepSeek's V4 family on Huawei Ascend chips, measuring the run at 34.22% model FLOPs utilization.
Cognition has added The Interaction Company of California, the makers of Poke, a personal agent that can text natively on Apple Messages, to its team.
The deal gives Sierra immediate access to a platform that can power autonomous, multistep applications across enterprises.