Tracked 382 · Curated 15
Anthropic research introduces "introspection adapters" (IA), a tool enabling language models to self-report learned behaviors, including potential misalignment. The IA generalizes to detecting hidden misalignment, backdoors, and safeguard removals, allowing models to describe their own behaviors.
Stripe has released the Link agent wallet, enabling AI agents to make payments on behalf of users without accessing their real card information. AI agents can create single-use payment credentials that expire after use. Users receive notifications to approve or reject each purchase within Link. The Link CLI is now open-source, allowing developers to integrate it with tools like Claude Code, MCP clients, or their own agent workflows.
In a fireside chat at Sequoia AI Ascent 2026, Andrej Karpathy discussed three new LLM paradigms: fully LLM-driven applications like menugen, software installation via natural language, and LLM knowledge bases. He also addressed the 'jaggedness' of LLMs and the emerging agent-native economy. Karpathy highlighted the shift from merely speeding up existing tasks to enabling fundamentally new capabilities and speculated on a future where classical computing supports primarily neural computation for agentic systems.
AI Agent products feature two main interaction models: Agent-centric (e.g., Codex, Cursor) with chat lists on the left and code/docs on the right, and user-operation-centric (e.g., GitHub Copilot) where the agent interacts via a sidebar. While sidebar interactions have drawbacks, users are accustomed to them. A key design decision is whether the agent is central or supplementary. Furthermore, handling the transition from stateless to stateful interactions is crucial for innovating agent UI.
The Qwen team has released FlashQLA, a high-performance linear attention kernel library specifically optimized for NVIDIA Hopper GPUs. Built on the TileLang compiler framework, FlashQLA leverages the Gated Delta Network (GDN) mechanism for linear attention, achieving up to 3x speedup over existing Triton kernels. It addresses the O(n²) complexity bottleneck of traditional Transformer attention mechanisms, enabling more efficient processing of long sequences through techniques like gate-driven context parallelism, hardware-friendly algebraic reformulation, and fused warp-specialized kernels.
The FDA is launching a pilot program to monitor clinical trials in real time using AI and cloud computing. The agency believes this initiative could significantly shorten the time it takes to approve new drugs.
The author posits that 'structure' is the core concept behind product design, Agent development, large language model (LLM) training, human relationships, and organizational design. A well-structured product guides users naturally, LLMs perform excellently with proper structure, and teams are efficiently organized. The article criticizes AI transformation efforts in major tech companies as superficial, comparing them to 'adding stones to a riverbed' without changing the underlying structure, arguing that learning AI will be futile without organizational change. It emphasizes that changing one's destiny requires first restructuring.
RADIO-ViPE is the first online semantic SLAM system that processes any monocular video without requiring calibration, depth sensors, or prior poses. It understands natural language queries like "locate spoon" while simultaneously mapping 3D environments.
Claude Security is now available, offering security teams a way to utilize Opus 4.7 within their codebases without the need for custom tooling. The solution eliminates the requirement for API integration or building custom agents.
Cloudflare has made post-quantum encryption for its IPsec service generally available, utilizing the IETF hybrid ML-KEM (FIPS 203) draft. This advancement aims to protect against 'harvest-now-decrypt-later' attacks and has passed interoperability tests with hardware from Fortinet and Cisco. This move accelerates Cloudflare's goal of achieving full post-quantum security by 2029 and secures Wide Area Network (WAN) traffic against future quantum computing threats.
With ChatGPT remaining the top downloaded app on the App Store, it appears OpenAI will soon launch a new iPhone app.
A study in collaboration with Harvard and Stanford medical experts showcases an AI system that processes physical symptoms in real-time using live video and audio. It can analyze gait, breathing, and rashes, demonstrating diagnostic reasoning in simulations with patient actors.
When deploying agents to production, multi-tenancy is a key consideration. Auth provides the solution for this, with integrations supporting Clerk and Supabase.
Andrew Kelley, the creator of Zig, explains that while not all LLM-assisted code can be perfectly identified, the distinct nature of human vs. AI errors makes them detectable. He likens spotting LLM users to sensing a smoker in a room, stating their 'digital smell' is obvious to those who abstain. Kelley prohibits this 'smoking' in his 'house'.