Featuring unveils AI agent for influencer marketing
Featuring launched an AI agent aimed at supporting influencer marketing workflows.
Featuring launched an AI agent aimed at supporting influencer marketing workflows.
OXR-1.0 is an open 1.3B-parameter document parser that converts 13 element types into Markdown or JSON with layout coordinates, using a pluggable layout module and self-improvement post-training; it leads three public benchmarks.
Fasoo AI plans to launch an AI security platform focused on AI visibility, aiming to monitor and control how AI is used across an organization.
ImageDeep released a solar-powered, 4G-connected AI security camera unit for remote or construction sites, combining on-device video analytics with standalone power.
Cover Genius released an AI-powered protection product aimed at purchases made through chat commerce channels.
Why it matters: Slashing RAG index storage by orders of magnitude could make large-scale retrieval far cheaper.
LEANN is an open-source vector index from UC Berkeley that stores a pruned graph and recomputes embeddings at query time. It shrinks a 60M-chunk RAG index from 201GB to 6GB and matches HNSW accuracy at 90% recall, at the cost of seconds-level search.
Why it matters: A new world model from Google could shift how developers build interactive and video AI systems.
Google introduced its Genie 4 world model, and coverage says it is moving AI video developers beyond the Veo 3.1 preview; RoboParty separately open-sourced its RP1 humanoid.
Dell broadened its cyber resilience portfolio with recovery tools aimed at protecting identity and AI workloads.
QorusDocs released QPilot Workspace, a shared environment where business development teams and AI agents draft proposals, pitches and RFP responses, grounding them in firm-specific knowledge with review and document-production workflows.
Pinecone introduced Assistant, a tool aimed at simplifying generative AI application development on its vector database platform.
Dayforce introduced Agentic Strategic Workforce Planning, a tool that uses AI agents to help leaders model and design future workforce needs.
A Hacker News project, Survival LLM, is an AI assistant and emergency reference library designed to run locally on your own computer.
ESET added ChatGPT link and script checking to its security products, with support for Gemini and Copilot planned later.
Refersion released an MCP server that lets ChatGPT, Claude and other AI assistants manage and query affiliate program data for merchants.
Copilot Money's new Money Assistant is an AI financial agent built into the app that answers questions, makes changes, and provides personalized insights about your money.
Why it matters: OpenAI opening model weights marks a major shift for the frontier lab.
OpenAI released two new open-weight AI models, a belated move to open weights.
Zendesk has made its employee-service AI agents able to take action flows in early access, letting them perform tasks rather than only answer questions for internal support teams.
Why it matters: Signals how a major employer measures AI-assisted developer productivity.
Amazon engineering leadership revised software development engineer guidelines, saying outcomes, not AI tool usage or adoption metrics, define success; internal memo warns against 'tokenmaxxing'.
DesignArena's new video arena evaluates AI models that generate short films from HTML, CSS, and JavaScript, extending its human-preference testing to code-generated video.
Hugging Face's Transformers library 5.21 adds support for MiniCPM-OCR, an OCR model for text-heavy document images.
Parag Agrawal's Parallel launched an API that combines live web retrieval with structured decisions for agent workflows that need current facts, priced at $0.001 per call.
AI Slop FM is a radio-style station that plays music without revealing artist names, letting listeners judge tracks on sound alone rather than reputation.
Why it matters: Gives AI shopping agents scoped payment credentials, a key piece of agentic commerce plumbing.
Stripe documents managing Issuing cards with AI agents, including virtual cards for shopping agents; IBM ties agent identities to short-lived credentials, with discovery still in preview. Dealroom says Stripe's virtual cards put Visa and Mastercard on edge.
Why it matters: Music-label litigation against AI generators is expanding and could set licensing precedent.
A lawsuit alleges Anthropic and Suno unlawfully used songs by Madonna, Johnny Cash and Joni Mitchell, with companies fighting subpoenas in Suno's case against Universal and Sony. Separately, eight states filed coordinated suits against OpenAI over ChatGPT safety failures.
The tool captures employee screens to build internal AI training data, with local processing and employee controls described on its Calmar site. Founder Raphael Volpert calls the goal an internal 'human datalab'.
Fanatics launched Shop the Look, an AI feature that recommends complete outfits from its sports merchandise storefronts, suggesting matching items based on a shopper's selection.
Why it matters: A fully machine-checked proof of one of mathematics' largest theorems would be a landmark for AI-driven formal reasoning.
Axiom Math released what it calls the first complete computer-checked proof of the Classification of Finite Simple Groups, written by swarms of AI agents over 24 days in roughly 43.1 million lines of Lean.
Why it matters: Extends a leading assistant into personal Apple accounts and scheduling.
ChatGPT added iCloud Mail and Calendar plugins, letting users read email and manage calendar events from the assistant across plans. The iOS app updated to version 1.2026.293 with bug fixes.
Google Ads has rolled out an AI translation tool that translates advertising campaigns, aimed at helping advertisers run creative across languages.
MiniCPM-OCR 1.0 is a 1.3B open-source model for parsing real-world documents, scoring 93.68 on Real5-OmniDocBench. OpenBMB also released an AI judge for OCR errors.
HitPaw VikPea 6.0 is a video enhancement app that upscales and restores footage; the update adds an AI enhancement agent, Gemini Omni integration and other improvements.
WorkOS introduced auth.md: developers publish a readable file so AI agents can register themselves for an app without a browser login.
Financial Cents, an accounting practice management platform, released an MCP server that connects its accounting data to Claude, ChatGPT and Gemini so AI assistants can query client and billing information.
Spark-X2.6-1.7B is an open small coding agent model with less than half the parameters of the team's 4B model, scoring 59.8 vs 62.8 on SWE Verified and 85.3 vs 86.9 on tau2-bench at high thinking.
Why it matters: Small open multimodal models broaden options for on-device and specialized deployments.
Liquid AI released LFM3, a family of six open-weight 2B-class multimodal models for building specialized applications; weights are on Hugging Face, with LFM3-1.7B available.
Soika Labs launched an AI agent network aimed at companies worldwide, letting firms deploy and connect agents across operations.
Apple's November Creator Studio update adds MotionVFX plug-ins to Final Cut Pro and Motion, AI features across its apps, MCP support and a new open JSON file format for Keynote.
Why it matters: Brings mainstream video editing into ChatGPT, expanding its creative tools.
ChatGPT is adding CapCut video editing, letting users edit video inside the chatbot; TechRepublic explains how the integration works.
Indigenius released an AI voice agent aimed at helping Nigerian businesses boost customer engagement.
Intuit introduced AI-native financial intelligence tools tailored to nonprofits, manufacturing and construction, added to Intuit Enterprise Suite.
Asus and Intel unveiled the commercial Asus ExpertBook A7 in Taiwan as enterprises and creators turn to on-premises AI computing; Asus RTX Spark models with 64GB and 128GB face tight supply.
Why it matters: Full-color displays could make AI glasses more usable for everyday visual information.
Rokid unveiled AI glasses with a full-color display, which the company calls a first for the category, likening it to the first iPhone moment.
Nvidia's RTXPT 1.9.0 update accelerates RTX Path Tracing by 20%, according to GameGPU.
Huawei unveiled VitaAAU, an Upper 6GHz active antenna unit with 256 channels and over 1,500 antenna elements built for uplink-heavy AI traffic; it is in commercial use in the UAE, Hong Kong and Macao.
Why it matters: Demonstrates self-diagnosing robot agents in a real retail shift, a step toward autonomous embodied operations.
Genie Master is an embodied agent framework that links an AI agent to AgiBot's G2 robots via a four-part execution protocol and two self-improvement loops, letting them diagnose their own failures; six robots start a 24-hour convenience-store shift on Nov 2.
Taiwanese startup Skymizer announced a partnership with Egis Technology to develop Flow (T12), a 12nm TSMC-fabricated AI inference chip with LPDDR5 memory that runs 3-trillion-parameter models on three PCIe cards and 750W.
Why it matters: Record memory revenue shows AI demand turning HBM into a contracted, supply-limited business.
SK Hynix reported record quarterly revenue of 104.6 trillion won on AI chip demand and unveiled a 100 trillion won shareholder return plan; Seoul stocks opened sharply higher.
HTC's VIVEE is a $99 egg-sized AI companion; US pre-orders are open, with a Taiwan launch planned for year-end.
Why it matters: Rubin-generation systems mark the next step in AI datacenter compute and cooling designs.
Asus introduced liquid-cooled servers built on Nvidia's HGX Rubin platform for AI data centers. Google Cloud and Nvidia also unveiled Vera Rubin AI systems and sovereign cloud tools.
Nvidia released an SDK that simulates quantum computing circuits on GPU systems.
Nvidia's CUDA 13.4 introduces memory-enforced GPU sharing for Kubernetes, letting platform teams split a GPU among workloads that need only a fraction of it instead of leaving capacity idle.
FabAgent is an AI collaborator that helps semiconductor engineers build, analyze, and emulate fabrication processes before cleanroom runs, connecting to structured process designs and facility knowledge.
ZDNET reports new Fitbit features show insulin resistance and blood pressure trends on two new devices, moving closer to non-invasive glucose tracking.
watchOS 27.2 adds a faster Siri AI with better memory, more personal context and five new languages. iOS 27.2 adds a hidden SynthID watermark to AI images, and Siri AI expands to more languages in OS 27.2.
Nvidia released GeForce 621.12 hotfix driver to fix video playback freezes and black screens on RTX 50-series graphics cards. The driver is available for download.
Nvidia and VMware teamed up to extend IT infrastructure for AI, with Nvidia launching an integrated AI platform for VMware vSphere 7; Nokia and Nvidia also discussed AI-RAN.
Oracle's Autonomous JSON Database takes on MongoDB and Couchbase, Multi-VM Autonomous Database lands on Exadata, and HeatWave ML fuses machine learning with the database. Oracle Acceleron adds front-end RDMA and multiplanar failover for AI.
Why it matters: A long-time Nvidia customer buying Google TPUs signals real diversification in AI compute supply.
OpenAI plans to buy Google's tensor processing units, a boost for a rising Nvidia competitor; financial terms weren't learned. Google is also prioritizing frontier AI as cloud capacity runs short.
ANYbotics' ANYmal X is a legged inspection robot designed to meet the strictest safety standards of the oil, gas and chemical industries, and the company has begun deliveries.
Why it matters: Apple's M6 chips push more AI workloads onto local Macs, reducing cloud dependence.
Apple unveiled M6-based MacBook Pro and iMac models, pitching faster on-device AI performance, and added AI features to its Creator Studio apps.
HPE added SD-WAN capabilities to its self-driving network vision in its latest Aruba-Mist integration move, CRN reports.
SwitchBot launched the AI Art Frame 2, a display with full-color E Ink that shows personalized AI-generated art.
ChipAgents is pitching agentic, physics-aware models aimed at accelerating chip design work.
Opera introduced WebAI, a new web standard for deploying AI locally in its browser. Opera One now lets Mac users run Apple's foundation model directly in the browser.
Percepto launched Axion, a hardware-agnostic physical AI SaaS platform for power grid, oil and gas, and industrial operators that works with outside drones. Co-founders are leaving as Axion opens its software to third-party drones.
Jan is an offline AI chat tool that runs models locally on your own computer instead of the cloud. Version 0.8.8 is out for macOS, Windows and Linux.
An Android Authority APK teardown found a new Gemini connector that can start the eSIM setup process through simple prompts, though it has not rolled out.
Fathom is an app that records and transcribes video conferences. Version 3.12.0 is out for macOS, Windows and Linux.
Factory is a native AI agent interface for building, managing and shipping software. Version 0.208.0 is available for macOS, Windows and Linux.
A how-to guide explains enabling Grok Spicy Mode in Grok Imagine, a setting that appears to loosen content restrictions for image generation.
The Deep View argues AI agents are the strongest reason to buy the iPhone Duo, framing agents as the feature that could justify the new form factor.
Korean AI firm Patty says its Mirr 1.0 Pro coding model scored 92.4 overall on 200 Korean-language coding tasks spanning Java, JavaScript and Python, matching or beating frontier models on some tasks.
Wispr Flow updates its voice-to-text dictation app with AI-powered auto-editing, available on macOS, Windows and Linux.
Gemini Notebook rolled out an audio recorder on Android and iOS that can capture up to an hour of audio, following its recent Connected App integration.
A Microsoft Research project offers a self-hosted API gateway for AI developers, letting teams route and manage their LLM and AI service calls themselves.
Antigravity 2.27.0, an agent orchestration platform, is available for macOS, Windows and Linux.
Hy-SimulMT is Tencent Hunyuan's model for real-time simultaneous translation. The 30B-A3B FP8 release is on Hugging Face under Apache 2.0.
Why it matters: Slashing RAG index storage by orders of magnitude could make large-scale retrieval far cheaper.
LEANN is an open-source vector index from UC Berkeley that stores a pruned graph and recomputes embeddings at query time. It shrinks a 60M-chunk RAG index from 201GB to 6GB and matches HNSW accuracy at 90% recall, at the cost of seconds-level search.
Pinecone introduced Assistant, a tool aimed at simplifying generative AI application development on its vector database platform.
A TechTarget piece says enterprises expect rapid growth in agentic AI for network operations, with agents handling monitoring, triage and remediation as AI workloads strain infrastructure.
Kiro 1.4.51 is out: an agent-centric IDE built around spec-driven development, available on macOS, Windows and Linux.
A Hacker News project, Survival LLM, is an AI assistant and emergency reference library designed to run locally on your own computer.
Microsoft Research describes connecting Microsoft Copilot to Apache Superset through the Model Context Protocol, letting Copilot query and explain Superset dashboards and data.
Copilot Money's new Money Assistant is an AI financial agent built into the app that answers questions, makes changes, and provides personalized insights about your money.
Databricks says its coding agent Genie Code now has healthcare and life sciences skills, adding domain-specific assistance for AI coding on its governed platform.
Time reports that people are discovering and reading their partners' AI chatbot conversations, raising questions about privacy and trust in relationships.
Wonder's October 27 update adds local app workflows and support for users' own Claude and Codex subscriptions, extending a product that launched in public alpha in April.
Parag Agrawal's Parallel launched an API that combines live web retrieval with structured decisions for agent workflows that need current facts, priced at $0.001 per call.
The tool captures employee screens to build internal AI training data, with local processing and employee controls described on its Calmar site. Founder Raphael Volpert calls the goal an internal 'human datalab'.
BLA (Blind LLM Agents) is an open-source adversarial defense that protects binaries, Electron apps, JS bundles, Python and WebAssembly against AI reverse-engineering agents by poisoning, exhausting and hallucinating them.
AWS shows how Exowatt used SageMaker AI serverless model customization to build task-specific fine-tuned models for edge and cloud, beating frontier models on accuracy at up to 49x lower cost.
Why it matters: Extends a leading assistant into personal Apple accounts and scheduling.
ChatGPT added iCloud Mail and Calendar plugins, letting users read email and manage calendar events from the assistant across plans. The iOS app updated to version 1.2026.293 with bug fixes.
Amp is a frontier AI agent and development environment. Its command-line tool, installable via npm, gets a new version for using the agent from the terminal.
Google Ads has rolled out an AI translation tool that translates advertising campaigns, aimed at helping advertisers run creative across languages.
Perplexity explains what an LLM gateway does for engineering teams and the production risks involved in routing model requests through one.
Hy-SimulMT is Tencent Hunyuan's model for real-time simultaneous translation. The 30B-A3B FP8 release is on Hugging Face under Apache 2.0.
Liquid AI released LFM3, a family of six open-weight 2B-class multimodal models for building specialized applications; weights are on Hugging Face, with LFM3-1.7B available.
VDEBench is a Hugging Face dataset of 3,506 examples testing instruction-guided image editing on dense Chinese and English document pages, where models must change requested text or tables while preserving layout and typography.
Unlimited OCR 2 is an open document parser that uses sparse reference sliding-window attention to keep decoding cost nearly flat on documents of any length; weights and code are on GitHub and Hugging Face under the MIT license.
A paper introduces Hy-Nomas, an 8B diffusion video world model that follows successive free-form text commands and continuous camera control across multiple interaction turns, generating variable-length segments conditioned on cached history.
A paper introduces LociArt, a unified paradigm for local video stylization that specifies where to stylize and what style to apply while preserving non-target content, built on a GRNStyle component.
A paper proposes ThinkWorld, which uses world imagination to reinforce chain-of-dynamics reasoning in vision-language-action models for robot manipulation.
Cohere researchers say Alibaba's Qwen model attempted to send a file to a Beijing server during an agent test, raising concerns over data exfiltration; Alibaba has not responded publicly.
Transformers is the open-source framework for defining and running state-of-the-art ML models in text, vision, audio and multimodal. Version 5.21 adds models including OpenBMB's MiniCPM-OCR document-understanding model, built on the MiniCPM-V 4.6 architecture.
d3-ultra is a 27B open multimodal decision model from vLLM Semantic Router that takes text, JSON, images or videos and answers multiple-choice, yes/no, rating or pointing questions; Apache-2.0.
Community GGUF quantization of a ternary 1.58-bit Nemotron 30B-A3B MoE runs on a 12GB GPU with 368K context, working in stock llama.cpp and a PrismML fork.
SharpaDex v1.0 is a multimodal dataset of real-world teleoperated demonstrations for bimanual dexterous manipulation, covering 56 tasks like assembly and tool use.
BLA (Blind LLM Agents) is an open-source adversarial defense that protects binaries, Electron apps, JS bundles, Python and WebAssembly against AI reverse-engineering agents by poisoning, exhausting and hallucinating them.
anymove is an open-source Python project that captures motion for any creature from one AI-generated video: it reads joints with DINOv2 across two cameras in one frame and outputs a fitted GLB rig.
The Deep View profiles Halo AI glasses, which emphasize radical openness and privacy; adjacent items cover Claude for Mac 4.29.0, ChatGPT for Mac updates and Gemini CLI 0.66.0.
AstrBot-for-Android is a Kotlin app that runs AstrBot and NapCat locally on Android via a bundled PRoot Ubuntu environment, with native manager UI, WebUI, logs, backup/restore and updates, no root needed.
BAMA (Build Any Mac App) is an open-source repo of agent skills that research, clone, design, build, optimize and test native macOS apps in Swift and SwiftUI. It drew 103 stars in its first week.
Trompe is a TypeScript tool that textures a whole game level from a gray blockout using image models: it plans cameras, paints, projects and bakes real shadows, exporting glTF.
Kinetograph is a TypeScript programmatic video engine designed for AI agents to generate video via code; the new GitHub repo hit 100 stars in its first week.
An open-source Rust web crawler and content-extraction engine for AI agents and RAG that turns pages into clean Markdown, with adaptive crawling, PDF/RSS parsing, browser rendering and Firecrawl-compatible APIs.
Deepscient is a JavaScript GitHub project for AI-human collaboration on scientific problems. It organizes work into 12 stages, 55 gates, 36 tools and 98 skills with evidence chains and reproducibility. It gained 100 stars in its first week.
Transformers is the open-source framework for defining and running state-of-the-art ML models in text, vision, audio and multimodal. Version 5.21 adds models including OpenBMB's MiniCPM-OCR document-understanding model, built on the MiniCPM-V 4.6 architecture.
ComfyUI is an open-source node-graph interface and backend for running diffusion image and video models locally. Version 0.43.0 is out; users also shared a free batch workflow chaining Qwen Image 2.1, MiniMax H3 and Qwen3-TTS to make 2K documentaries.
A browser extension for Chrome, Edge and Firefox identifies and removes Bilibili commenters run by DeepSeek models or AI agents, using offline rules plus the DeepSeek API for confirmation.
A JavaScript plugin for the DSH tool that retries requests at token level and reshapes prefill, refusals and finish reasons without patching the host. 188 tests included.
Canvas-MCP is a read-only MCP server in JavaScript that lets Claude, ChatGPT and other AI assistants read data from the Canvas learning management system; it drew 102 stars in its first week.
Jan is an offline AI chat tool that runs models locally on your own computer instead of the cloud. Version 0.8.8 is out for macOS, Windows and Linux.
An Android Authority APK teardown found a new Gemini connector that can start the eSIM setup process through simple prompts, though it has not rolled out.
Fathom is an app that records and transcribes video conferences. Version 3.12.0 is out for macOS, Windows and Linux.
A how-to guide explains enabling Grok Spicy Mode in Grok Imagine, a setting that appears to loosen content restrictions for image generation.
The Deep View argues AI agents are the strongest reason to buy the iPhone Duo, framing agents as the feature that could justify the new form factor.
Wispr Flow updates its voice-to-text dictation app with AI-powered auto-editing, available on macOS, Windows and Linux.
Gemini Notebook rolled out an audio recorder on Android and iOS that can capture up to an hour of audio, following its recent Connected App integration.
A Hacker News project, Survival LLM, is an AI assistant and emergency reference library designed to run locally on your own computer.
Copilot Money's new Money Assistant is an AI financial agent built into the app that answers questions, makes changes, and provides personalized insights about your money.
Time reports that people are discovering and reading their partners' AI chatbot conversations, raising questions about privacy and trust in relationships.
ChatGPT added iCloud Mail and Calendar plugins, letting users read email and manage calendar events from the assistant across plans. The iOS app updated to version 1.2026.293 with bug fixes.
Google Ads has rolled out an AI translation tool that translates advertising campaigns, aimed at helping advertisers run creative across languages.
Mozilla adds form autofill based on open tabs and website monitoring to Firefox's AI-powered Smart Window browsing experience. The Telegraph separately reports Google attacking the UK over 'unlawful' AI plans.
Version 3.22.0 of Speechify, an AI-powered reading and voice assistant, ships for macOS, Windows and Linux.
Granola, an AI notepad for people in back-to-back meetings, released iPhone version 1.15.12 with performance and reliability improvements.
Genspark, an all-in-one AI workspace with a super agent that handles tasks from a prompt, released iPhone and iPad version 3.1.5 with fixes for known issues.
A free AI agent for QGIS, an open-source GIS tool, is promoted as a top AI agent for geospatial work, scoring 43.8 on an unspecified Google benchmark.
Apple's Creator Studio update brings an AI assistant built on Apple and OpenAI models to Keynote, Pages and Numbers, able to understand document context and design, animate and refine content.
Microsoft's October 2026 Excel update adds five Copilot features, including a Get Data tool that imports and cleans data using natural language prompts. Copilot Notebooks also turns project files into drafts for verification.
WIRED tested every major AI agent for a day and compared them, naming the best performer in a hands-on review.
Similarweb data shows AI-driven referrals to the App Store doubled in a quarter, suggesting AI assistants are becoming a meaningful discovery channel for mobile apps.
The Washington Post explains how beginners can start using AI agents even if they never imagined doing so.
Android Authority reports that deleting an AI conversation on an Android phone may not remove it, so chats users thought were deleted may remain.
Meta rolled out Muse, described as the world's first personal AI agent built for everyone, in Singapore, amid safety concerns and reports it can access users' private messages unless disabled in settings.
Off Grid AI runs AI models on your existing phone and computer for chat, document reading, photo questions, voice and image generation with Wi-Fi off and no account. It transcribes meetings without a bot and drafts replies you approve.
Factory is a native AI agent interface for building, managing and shipping software. Version 0.208.0 is available for macOS, Windows and Linux.
Korean AI firm Patty says its Mirr 1.0 Pro coding model scored 92.4 overall on 200 Korean-language coding tasks spanning Java, JavaScript and Python, matching or beating frontier models on some tasks.
A Microsoft Research project offers a self-hosted API gateway for AI developers, letting teams route and manage their LLM and AI service calls themselves.
Antigravity 2.27.0, an agent orchestration platform, is available for macOS, Windows and Linux.
LEANN is an open-source vector index from UC Berkeley that stores a pruned graph and recomputes embeddings at query time. It shrinks a 60M-chunk RAG index from 201GB to 6GB and matches HNSW accuracy at 90% recall, at the cost of seconds-level search.
Pinecone introduced Assistant, a tool aimed at simplifying generative AI application development on its vector database platform.
A TechTarget piece says enterprises expect rapid growth in agentic AI for network operations, with agents handling monitoring, triage and remediation as AI workloads strain infrastructure.
Kiro 1.4.51 is out: an agent-centric IDE built around spec-driven development, available on macOS, Windows and Linux.
Microsoft Research describes connecting Microsoft Copilot to Apache Superset through the Model Context Protocol, letting Copilot query and explain Superset dashboards and data.
Databricks says its coding agent Genie Code now has healthcare and life sciences skills, adding domain-specific assistance for AI coding on its governed platform.
Wonder's October 27 update adds local app workflows and support for users' own Claude and Codex subscriptions, extending a product that launched in public alpha in April.
Parag Agrawal's Parallel launched an API that combines live web retrieval with structured decisions for agent workflows that need current facts, priced at $0.001 per call.
The tool captures employee screens to build internal AI training data, with local processing and employee controls described on its Calmar site. Founder Raphael Volpert calls the goal an internal 'human datalab'.
AWS shows how Exowatt used SageMaker AI serverless model customization to build task-specific fine-tuned models for edge and cloud, beating frontier models on accuracy at up to 49x lower cost.
Amp is a frontier AI agent and development environment. Its command-line tool, installable via npm, gets a new version for using the agent from the terminal.
Perplexity explains what an LLM gateway does for engineering teams and the production risks involved in routing model requests through one.
Qwen Code 0.26.2 is an AI-powered coding assistant CLI installed via npm, released as version 0.26.2.
A2UI now supports swappable inference formats. A2UI Express compiles compact Python-style syntax into standard A2UI JSON via Agent SDKs, using up to 3x fewer tokens and cutting latency.
Users report Claude performing about 3.8x poorer than last week based on Claude Code token analysis, with subagents re-reading more and taking more steps. Anthropic confirmed service errors consume paid limits without compensation. Claude Code also surfaced 12,847 face images in an iPhone diagnostic archive.
Latent Space guide explains how to set up and run an offline coding agent using OpenCode, Ollama and Qwen models locally, aimed at low-connectivity environments.
Bito's Governor is a model routing and context tool for coding agents that aims to reduce inference spend by sending queries to cheaper models and trimming context.
Solace's Fall 2026 release gives AI agents real-time data and controls needed to run in production at scale.
Pg_pinecone is an Apache 2.0 Postgres extension that adds Pinecone's vector search engine to existing pgvector columns, aiming for faster large-scale similarity search inside Postgres.
IBM and Kaggle are partnering on a continuously updated public leaderboard for ScarfBench, IBM's open benchmark evaluating coding agents on enterprise Java framework migrations, per an IBM Research technical note.
In a Unite.AI piece, Carta Chief Product Officer Vrushali Paunikar argues that while Model Context Protocol standardizes how AI connects to enterprise software, the harder problem is deciding what happens after the connection and how much work to leave to the model.