Cainew

Qwen 3.0 Image Pro is a new AI model that advances image generation and understanding capabilities. The release represents progress in multimodal AI technology.

Maple-Preview demonstrates a ternary 20B mixture-of-experts model achieving 120 tokens per second performance on iPhone hardware.

Your follow-ups are only as good as your meeting notes. Wispr Notetaker gets your words and your speakers right, so your recaps, follow-ups, and answers are too. Before the meeting starts, it checks the invite so names are spelled correctly, and it brings the terminology you've already taught Wispr Flow into every conversation. Your transcripts use real names instead of "Speaker 1" and "Speaker 2," and every meeting is ready to pull into Claude or ChatGPT via MCP. Available on Mac. Free to try.

AdAnt AI is a team of Creative Agents that handles Social Ad Strategy, Creation, and Iteration. Built on the strategies our founding team used to generate 50M+ Organic Views and reduce Paid Acquisition Costs by 60% on average, AdAnt AI creates Scroll-Stopping, High-Converting Ads for TikTok, Instagram, and YouTube.

ngrok AI Gateway provides one hosted gateway for every model: public providers, custom endpoints, and the models you run yourself. Use one key and one URL to route across OpenAI, Anthropic, and self-hosted models with observability, access control, and fallbacks built in. Your private models connect through ngrok’s network, so they sit beside hosted providers without being exposed to the public internet.

Cloudflare Wallets introduces a programmable wallet layer for the agentic Internet. Built for the future where AI agents can discover, use, and pay for services autonomously, it gives agents a secure way to transact with APIs, content, and digital services. With virtual wallets, spending controls, and machine-friendly payments, Cloudflare is building the financial infrastructure for the next generation of AI applications.

Kiro Crew is a persistent workspace that remembers your context, lessons, and skills across sessions, so you come back to progress instead of a cold start. Build a crew of agents that work across the tools you already use, wrapped in purpose-built Apps for the jobs you repeat.

Keystroke is an all-in-one platform for building powerful AI agents. Describe the agent you need, and Keystroke builds it, connects your tools, tests it, and deploys it to a shared workspace. Give agents memory, workflows, triggers, approvals, and access to 1,000+ integrations. Open source, YC-backed, and free to try with $20 in credits.

Most companies wire Claude or ChatGPT into Slack, email, calls, tickets, and their CRM over one singular MCP. That's raw pipes into scattered systems. The model reads a slice and guesses at the rest. BackEngine MCP connects to the same tools, but reads everything first, joins all of it into one permissioned record per account, kept current, so Claude and ChatGPT always work from the whole picture. Head-to-head: 67% fewer errors, 2.4x more key facts, 65% fewer tokens vs. direct connectors.

A Mac app that turns plain English into real apps — built and stored on your machine, not someone else's servers. Your code, your GitHub, your own AI at cost. No marked-up credits, no lock-in. Free, no sign-up. macOS (Apple Silicon), Windows on the roadmap.

Self-hostable sandbox SDK inspired by Cloudflare Sandbox SDK. works on any device (Mac, Linux, bare metal). super easy to use (both by human engineers and agents). you get full control over capacity/token spent per sandbox, you can default-deny egress but still allow access to LLM providers, and API keys never directly enter each sandbox (it creates per-sandbox tokens instead that die when the sandbox stops). npm i -g hotcell

Dover’s MCP connects Dover’s free ATS to ChatGPT, Claude, Cursor, and other AI tools. You can now use powerful AI tools to review applicants, schedule interviews, move candidates through your pipeline, add notes, and coordinate hiring. Access is secure and respects your team's existing permissions. Available today as part of Dover's free ATS for startups.

Researchers successfully trained a small language model (SLM) on an affordable ESP32-S3 microcontroller for just $8.

Cloudflare OS is an open platform designed for agents, apps, and work that enables developers to build and deploy AI agents and applications. The platform aims to provide infrastructure for the emerging agentic AI ecosystem.

AI systems are successfully solving Erdős problems, historically difficult mathematical problems named after Paul Erdős. This breakthrough demonstrates AI's growing capability in advanced mathematical reasoning and problem-solving.

Zero-Mem introduces a technique for zero-token memory operations that reduces the memory overhead of LLM agents during execution. This optimization improves efficiency and scalability of memory-intensive agent systems.

Painting with Gaussians is a technique that uses Gaussian splatting for image synthesis and rendering tasks. This method offers improvements in visual quality and computational efficiency for AI-generated imagery.

On-policy distillation, in which a teacher corrects samples that the student itself generates, presupposes that the two models speak the same language: identical VAE latents, matching architectures, and a common timestep grid. We ask what happens when none of this holds, as when the strongest teacher available and the student one wishes to deploy come from different model families, and find that the standard recipes have no answer: teacher latents cannot serve as targets in a foreign coordinate ...

On-policy training has emerged as a powerful post-training paradigm for improving the reasoning capabilities of large language models, and is often enhanced by golden trajectories from stronger expert models. However, when the expert fails on harder problems, existing trajectory-guided methods lose their main source of supervision, and these failed trajectories are typically discarded as negative samples. We argue that such failures, which we call Golden Negative Trajectories, can still provide ...

Self-evolving agents increasingly convert interaction histories into reusable skills that persist beyond individual tasks. While prior work studies memory and retrieval poisoning, such attacks only affect agents when poisoned records are retrieved as context. We uncover a new and more fundamental risk: poisoned experiences can be transformed by the agent itself into durable behavioral artifacts. We present SkillJack, the first attack that exploits the experience-to-skill pipeline of self-evolvin...

Modern agent frameworks equip large language models with external skill libraries to solve complex tasks. However, it remains unclear whether these systems can effectively evolve their skills and whether the resulting skills improve task-solving capabilities. To bridge this gap, we introduce ContinualSkillBench, a dynamic evaluation framework for in-context continual skill learning. It covers five representative domains, each containing 100 interconnected subtasks ordered by increasing difficult...

We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bias, where agents bypass visual tools in favor of textual search, and (2) parametric knowledge leakage, where models rely on internal memory rather than genuine tool-augmented execution. To address thes...

Reasoning language models frequently overthink: generating extended chains of behaviors such as hedging, approach abandonment, and self contradiction that consume tokens without improving answers. We show that these behaviors are not merely a consequence of length; even when controlling for response length, incorrect traces exhibit higher rates of unproductive self-reflection than correct ones. Addressing this requires identifying where self-reflection helps vs hurts, but obtaining these step-le...

Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. Specifically, for optimization, the optimal nominal batch size grows faster, while the optimal learnin...

We identify a previously overlooked failure mode of ALiBi positional encoding: its linear bias scaling underflows floating-point precision, which zeroes out a large fraction of attention weights and renders the affected attention heads partially blind. We analyze this failure mode, characterize its impact, and examine four mitigation strategies. We further demonstrate its occurrence in state-of-the-art pretrained models based on ALiBi. Comprehensive pretraining experiments with 148M-parameter de...

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation offers denser signals through teacher branches with privileged context, but existing approaches typically derive such context from ground-truth answers or retrieved skills, which may not reflect the stat...

Persona skills distill personal interaction histories into portable and executable artifacts for downstream agents. While enabling flexible personalization, this process concentrates fragmented personal signals, amplifies their impact through reuse, and challenges defenses designed for individual records or retrieval-based memory. To systematically investigate the safety of the persona-skill pipeline, we introduce AntiSkillBench, an end-to-end benchmark for evaluating risks and defenses across t...

Audio-based piano transcription performs well on onset, pitch, and velocity, but the sustain pedal lets sound persist long after key release, so audio systems predict pedal-extended offsets rather than physical key release. Yet existing Visual Piano Transcription (VPT) systems focus on onset detection from short video windows, offset accuracy lags onset by a wide margin, and note-level velocity has not been reported. To address these gaps, we present V2N (Video to Notes), the first complete VPT ...

A common task in legal Information Retrieval (IR) is to find relevant legal sources from case-law collections. While legal practice often requires pinpoint citations (pincites) to specific case paragraphs, most existing public legal IR datasets lack paragraph-level citation annotations. Yet, publicly available datasets with such information contain data leakage in the query text and exclude paragraphs that are neither citing nor cited from the corpora, creating an unrealistic and oversimplified ...

This work focuses on building an advanced agentic harness to improve how AI agents are structured and controlled in complex tasks. The harness provides better frameworks for agent coordination and execution.

Demis Hassabis transitions from CEO to Chair at Google DeepMind while Jeff Dean, a prominent AI researcher, departs the organization. This leadership restructuring marks a significant change in the company's executive structure.

The Rust programming language community is adopting an official policy governing the use of large language models in development and decision-making. This policy addresses how LLMs will be integrated into the Rust project's processes.

The UK AI Security Institute released a security incident report (INC-2026-07-28-01) documenting vulnerabilities or threats in AI systems.

Several Iowa-led states are petitioning OpenAI to implement stricter controls and limitations on bot deployments within their jurisdictions.

Flowise, a popular open-source tool for building AI applications, is shutting down operations.

New third-party cyber security evaluations are being conducted on OpenAI models to assess their robustness and vulnerability to attacks.

A comprehensive analysis debunks eight common misconceptions about software engineering practices in the context of generative AI development.