Intel Chief Technology Officer Pushkar Ranade stated on the Dwarkesh podcast that even with sufficient funding, global leading-edge wafer production is difficult to expand tenfold, with a more realistic growth ceiling of approximately twofold; this is primarily limited by EUV lithography machines (ASML produces fewer than 100 units per year), equipment delivery cycles (approximately 18 months), and the synchronized expansion capability of all segments in the entire semiconductor supply chain.
Microsoft and Hugging Face jointly released ThinkingBox-Training, proposing a method that directly uses executable verification of database final states (rather than tool call sequences) as reinforcement learning reward signals for post-training agents; based on offline policy-relative difficulty filtering of 187 suitable tasks, after 50 steps of GRPO fine-tuning on Qwen3.8-27B, the number of tasks passing all 20 attempts in ThinkingBox-Bench increased from 38 (7.50%) to 107 (21.10%), and average success rate rose from 51.70% to 60.78%; code, data, and training stack have been open-sourced at microsoft/thinkingbox-training, the benchmark dataset ThinkingBox-Bench contains 507 real backend-driven tasks, and the paper is published on arXiv:2608.19741.
Consumer Agent Testing Tool (CATT) v0.2.1 is an open-source testing tool used to verify how agents (such as Muse/Instinct-type assistants) call your MCP services, embedded MCP apps, websites, or HTTP APIs.
Canva has released its proprietary image generation model, Canva Craft. It adopts a two-stage architecture that first plans the design and then generates the image, supporting up to 2K resolution and over 30 aspect ratios. It can be used for scenarios such as social media, marketing content, and digital art. The model is trained on millions of human-created templates from the platform and ranks fourth on the Design Arena leaderboard. It is currently open to global users, with access via the 'Dream Lab' option in the left-hand menu.
Hexal is a modal text editor written in C#, highly aligned with Helix in terms of key mappings and behavioral logic. It supports writing plugins via Lua, with an API sufficient to implement full LSP functionality. It also allows integrating external LSP servers and formatting tools through .hexdef files. It provides VS Code-style features such as a command palette and file browser. Build releases are published in the GitHub repository navid-m/hexal-dist. Its only dependency is NLua.
Cohere has launched a confidential computing feature based on NVIDIA GPUs in Model Vault, now available for public preview on Microsoft Azure; this feature provides verifiable confidential AI inference through hardware-level isolation and remote attestation, protecting prompts, documents, responses, and model execution processes, while encrypting request payloads and restricting intermediate infrastructure access to plaintext data.
Plasma AI has open-sourced a Linear A decipherment research system named kaphtor, which includes 1719 inscription records, 502 dictionary entries, 439 complete translations, and proposals for 34 symbol phonetic-semantic values; accompanying tools support symbol position retrieval, paradigm testing, arithmetic verification, and cross-language matching based on 3799 Semitic roots; all code (Apache-2.0) and raw data (CC BY 4.0) have been published to GitHub.
Forked is an open-source AI agent state management tool that provides fork, rewind, replay, and effect log replay capabilities, ensuring that side-effect operations such as payments and emails are executed strictly once during retries and rollbacks. It supports three deployment options: a Rust library, a self-hosted server (SQLite), and forked cloud (Cloudflare). It offers a TypeScript SDK, a LangGraph checkpointer, and an MCP server, with all core components open-sourced under the Apache-2.0 license.
BYOSM is an open-source tool that allows users to input a single sentence describing a task and automatically builds the corresponding small specialized model. It can run on any CPU, outputs in ONNX format, and supports deployment via Python, HTTP, Docker, or Hugging Face. The project has released several example models to Hugging Face, including byosm-doctype, byosm-intent, and byosm-vad.
OpenAI has fully opened the Decisions API to all customers. Rebuilt based on ChatGPT Luna, the API reduces inference latency from 1–3 seconds to 100–180 milliseconds; it has added voice and visual modalities and optimized global regional latency. Real-world tests by customers such as Box, Datadog, Ramp, and Speak show improved throughput, with Speak reducing action intervals from 1.4 seconds to 0.13 seconds. OpenAI is also using it to improve the response speed of its Dots voice assistant.
Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, turning the EGM framework used by internal engineers into an integrable development tool.
Buildnotes is a developer-oriented CLI tool that supports automatically converting PRs, issues, and reviews in code repositories into technical narrative posts, self-playing videos, or clickable slides via the /buildnotes command; it supports four output types (post/video/deck/figures), depends on Node 22, git, and Chrome, is installed via npx skills add buildinternet/buildnotes, and is suitable for coding agents such as Claude Code, Codex, and OpenCode.
Decido is a LoRA fine-tuned model based on Gemma 4 26B-A4B, released in 8-bit GGUF format. It can run locally on laptop CPUs (such as Apple M4 Pro) and supports yes/no, pick-one, and rating type decisions. The output format is compatible with the TypeSafe Jev 'System One' API, providing an HTTP service interface (POST /v1/systemone) that supports three types of questions: noul/choice/score. The model and all training, evaluation, and deployment code are open-sourced on GitHub under the Apache-2.0 license.
Amazon Polly's Generative engine has released a major upgrade, providing more natural, accurate, and low-latency voice output. It adds 14 generative voices covering 9 regions, and for the first time officially supports Hindi (Arjun, Kiara) and Finnish (Suvi) in the generative engine. The upgrade is a seamless replacement (drop-in), with VoiceId and SynthesizeSpeech API call methods remaining completely unchanged; existing applications can enable the new voices without modifying any code. The new engine is now generally available in 11 AWS regions worldwide (including AWS GovCloud US-West).
IskayPet uses Amazon Bedrock managed knowledge bases to build a multilingual, multi-country enterprise knowledge assistant without needing to self-build vector databases or embedding pipelines. By using native SharePoint connectors to directly index documents (including image content), adopting an independent knowledge base design for each business domain, and combining document-level permission control with hybrid retrieval, the time employees spend looking up operational information was reduced from 3–5 minutes to about 7 seconds. The solution's frontend and AWS architecture were co-developed by Kiro and Claude Haiku 4.5.
Amazon SageMaker introduces the Training Session API, enabling users to write complete training loops (such as SFT, RL, and preference optimization) in local Python processes. It drives models and LoRA adapters on managed GPUs via API calls like forward_backward, optim_step, and sample, while the service handles GPU allocation, distributed execution, fault tolerance, and checkpointing.
Microsoft published a paper in Science claiming that its GPT-5-based Diagnostic Orchestrator multi-agent system achieved a diagnostic accuracy of 83% in medical case tests, significantly higher than the 36% for general practitioners with reference literature and 22% without references; it reached an 80% accuracy rate on emergency data from Beth Israel Deaconess Medical Center, exceeding the doctors' average of 47%; the system uses the self-built benchmark SDBench and plans to open-source it; currently validated only on historical cases, real-world clinical environment validation has not yet been completed.
Amazon Quick is a HIPAA-eligible agentic workspace launched by AWS, supporting the processing of PHI in regulated healthcare environments; its capabilities cover chat agents, knowledge base retrieval, Quick Sight analytics, Quick Flows automation, and Research report generation. However, HIPAA compliance requires customers to complete responsibilities such as signing a BAA, creating ACL-aware knowledge bases, registering customer-managed KMS keys (irreversible), upstream data masking (e.g., Glue DataBrew), and configuring audit logs.
An AWS blog post introduces how to stream sentences generated by AI agents in real time to the Kokoro-82M text-to-speech model hosted on Amazon SageMaker AI, enabling speech synthesis to start immediately after the first sentence is generated, rather than waiting for the complete response.
Modal CEO Erik Bernhardsson 在 Latent Space 播客中披露，公司已完成 7.5 亿美元 Series D 融资，估值达 160 亿美元，年化营收突破 10 亿美元，并已启动近三十亿个沙盒；其 GPU 采购策略已从早期依赖云厂商 Spot 实例转向与近 25 家新云厂商签订三年期合约，同时正评估自购硬件可能性；沙盒产品最初增长缓慢，后因编码智能体、后台智能体和强化学习需求爆发，现占公司约三分之一营收。
Microsoft security researchers stated that China-linked hackers breached the cloud backup service of IT management provider ConnectWise, stealing customer application credentials, which they then used to log in with legitimate identities to access emails and files in customers' Microsoft-hosted environments; the attackers used the open-source AI framework ARTEX to drive AI agents, breaching approximately 2250 organizations at a rate of nearly 3000 attempts per minute; ConnectWise has suspended its cloud backup service and launched an investigation.
NVIDIA's official technical blog published a tutorial demonstrating how to use the KvikIO Python library to read files directly into GPU memory on Linux systems that support GPUDirect Storage (GDS).
This article introduces how to build a measurement-based auto-scaling strategy for Kokoro-82M (an 82 million parameter, non-autoregressive TTS model) on Amazon SageMaker AI real-time endpoints.