Why it matters: A large open-weight model tuned specifically for scientific discovery broadens open access in AI-for-science.
Hydrogen-1.0 is a 753B open-weight model for chemistry, materials science and life sciences, built by post-training Z.ai's GLM-5.3 on lab records and agent discovery trajectories. Open weights are available.
Evoke Security announced an upcoming integration with Claude Cowork to extend AI agent detection and response, adding security monitoring for Anthropic's workplace agent product.
ON1 released Photo RAW 2027 for Mac and Windows, adding a new RAW engine and revised NoNoise AI models. The AI Culling Assistant is exclusive to the MAX edition.
Readwise Books combines an ebook store, reading tracker, detailed reading stats and AI features, aiming to challenge Amazon's reading ecosystem.
Momentic (YC W24), launched on Hacker News, offers AI agents that test applications by using them like a user to catch bugs.
A new Chrome AI feature uses Gemini to read a user's Gmail and autofill web forms with the details found there. Engadget explains how to disable it if users don't want Gemini accessing email.
Cyara launched its Agentic CX Trust Platform, pitched as providing independent proof of quality and compliance on every live agent conversation, whether human or AI.
CrowdStrike and Dell Technologies announced integrated AI security for the Dell AI Factory with Nvidia.
Why it matters: A large seed round for a world-model approach to predicting human behavior.
42AI came out of stealth with $62.5M seed funding and a behavioral world model that aims to predict human decisions, even in novel situations.
A Reddit user shares Pinokio Remote, which lets people reach their locally hosted AI apps remotely. It is a community post with a link.
Why it matters: Anthropic targets financial advisors with a specialized Claude offering, expanding AI into wealth management.
Anthropic launched Claude for Financial Advisors, an AI tool aimed at helping financial advisors and challenging Wall Street's 1% fee model, according to a former advisor's analysis.
Why it matters: DeepSeek's agent push intensifies competition among Chinese labs for customizable AI agent tooling.
DeepSeek released Creator Mode, letting users quickly build custom plugins for its AI agent to tackle complex tasks, part of the Chinese lab's push to strengthen its agent offerings.
ACCESS Newswire launched SPOT, a system of seven AI agents that review every press release before submission, aiming to catch errors and improve quality.
Why it matters: Content watermarking from a major lab could shape how AI text is detected online.
The Independent reports Anthropic's Claude will now leave a trail indicating whether a text was made with AI, aimed at helping readers identify AI-written content.
Founder Gib Bassett is taking MyDataWork, an analytics work-context workspace, to agency buyers via AWS GovCloud, with Bedrock AI and a $15,000 annual public package.
WealthFeed, an AI-powered growth engine for financial advisors, added Form 5500 retirement-plan data, giving advisors another source for prospecting in the retirement plan market.
Dost, an AI finance automation platform, launched AskDost, an assistant that connects to a company's ERP and financial data, answers questions in natural language and runs finance workflows after human approval.
Applied Brain Research announced general availability of a speech AI SDK for the Zephyr RTOS running on Alif's Ensemble family of processors, bringing speech recognition to embedded devices.
Why it matters: Puts patients' own medical records under chatbot interrogation, a notable health-AI integration.
Athenahealth launched a tool that brings patient health records into Anthropic's Claude, letting patients ask the chatbot questions about their medical data.
LinearB launched Agentic Studio, which turns software-delivery data into an engineering context layer that leaders and coding agents can query in natural language to make decisions.
Sentira is an agentic AI experience for 10x Genomics' Chromium single-cell platform that lets researchers analyze Chromium data by asking questions; the company charges by the analysis.
Majanen, who co-founded Quattroporte in 2005, is launching an AI forecasting platform built on Q-Ball research dating to 2007.
KVARK Sparkle 1.0 is an open-weight model based on Qwen3.8-27B tuned by Egzakta AI Lab to cut reasoning tokens by up to 67%, giving faster answers at similar quality.
Abra is a unified AI agent for business planning across the enterprise; it routes planning questions to the right finance agent. Available now from Abacum.
Why it matters: Usage caps show frontier labs are rationing compute as AI losses grow.
OpenAI is preparing shared usage limits for ChatGPT and Codex across Plus, Pro and Business plans, as AI firms ration compute amid mounting losses, The Telegraph reports.
Kling3.io released Kling 4.0, an AI video generator it says produces more realistic visible action scenes.
Clarivate is connecting its CompuMark trademark data to Microsoft Copilot through IPOne, its IP intelligence platform, giving IP professionals natural-language access to trademark insights inside Copilot.
Ebix partnered with intentyx to launch SMARE, a platform for tracking and governing AI agent behavior in regulated industries.
Google Pay launched a new in-app AI shopping feature in India, letting users search and shop with AI assistance inside the payments app.
SFX v3 is ElevenLabs' newest sound effects model, swappable with a single model_id change; it improves multi-part prompts and vocals and clips fewer high notes.
KPN partnered with ElevenLabs to build Mobiel AI, an AI assistant inside KPN's business mobile plans that helps Dutch entrepreneurs manage every call.
Suno MCP lets users create music inside AI assistant chats via an MCP integration.
A patent-pending AI interviewer is designed to conduct talent acquisition interviews across languages, helping employers assess candidates who speak different languages.
Sunrate launched an Agentic Card that lets AI agents make business payments, aimed at automating corporate spend, The Manila Times reports.
SmartKargo launched a portfolio of AI applications aimed at airlines, pitched on delivering measurable value to air cargo operations.
Why it matters: AI stocks now carry enough weight to move the entire US market.
A sharp October selloff in AI-related shares, including Nvidia, is testing whether the wider US stock market can hold up without the AI trade leading it.
Why it matters: A regulator investigating agent-caused breaches signals new liability questions for AI agents.
Delaware's Justice Department launched a probe of OpenAI after a Hugging Face breach in which about 700 OpenAI agents reportedly used leaked credentials.
XDC Network partnered with Keyban to give AI agents non-custodial accounts for making onchain USDC payments, enabling agents to transact without a custodian holding funds.
Mango AI launched a face swap tool that replaces a person's face in short video clips with a chosen image, aimed at quick video creation.
SnapRef-v1a is a reference-based image generation model on Hugging Face aimed at keeping characters consistent across generated images.
ACTRA opened early bird access in the U.S. for a personal AI product positioned as going beyond the smartphone. Details on features and pricing are sparse in the source.
Nothing OS 5.0, based on Android 17, adds an AI app builder and new Glyph features; Livemint details the rollout timeline and how to download it.
Nvidia Developer describes Workload Map, a tool within its DSX AI Factory Platform that helps the ecosystem support new AI frameworks faster when designing and operating AI factories.
KISTI launched Hangang, South Korea's sixth supercomputer and its first 614.7-petaflop system, powered by 8,500 Nvidia GPUs.
TOZO's AIVU AI glasses are now on sale, touted as the first with a tinted waveguide lens, which keeps the display visible in bright light.
A "most significant" DirectX update extends Nvidia's Mega Geometry technology to GPUs from other vendors, broadening hardware support for the geometry system.
Why it matters: Frontier labs face growing local pushback over data center energy use, making grid commitments a template for future builds.
OpenAI's Grid Protectors Program will partner with utilities so future compute campuses can support grid reliability and protect consumer electricity costs, including cutting power use during emergencies.
Why it matters: A $115B optics acquisition signals AI data center connectivity is now core semiconductor strategy.
Marvell Technology agreed to buy data center optics maker Lumentum in an all-stock deal worth $115 billion enterprise value, one of the biggest tech tie-ups ever.
E2E Networks launched Blaze, a global pay-per-token AI inference service that lets customers run model inference without fixed commitments.
A Hugging Face repository, HannesVonEssen/microduck-basketball, is listed under the robotics task. The listing describes a public repository, not an inference endpoint; a simready-glb dataset is also posted.
Paris-based Gallium Defense emerged from stealth with a €7.1M seed round led by Keen Venture Partners and 360 Capital to make drone detection cheaper and more scalable than legacy systems.
Why it matters: Brings end-to-end AI planning into the main open-source self-driving stack.
Autoware 2.0, the open-source autonomous driving software, adds support for end-to-end AI planners while keeping rule-based components; TIER IV is spearheading the initiative with the Autoware Foundation.
redu.cloud is a cloud service designed for AI agents, offering fast access to stateful resources and a free tier. It targets app deployments, DevOps, training stateful models and fixing live incidents at agent speed.
Agibot unveiled Worken at its Innovation Day 2026, a proposed measure of how much productive work robots deliver, aiming to give the industry a standard beyond demos.
Honeywell launches Forge Plant Cognition, an industrial optimization offering built with Nvidia AI and Dell Technologies, aimed at plant operations.
Mobilint deployed Ceva's SensPro AI DSP and NeuPro-Nano NPU in REGULUS, its edge AI system-on-chip, marking the first production deployment of Ceva's NeuPro NPU family. Mobilint is a fabless AI chip company founded in 2019.
Krafton released Ludi 0.2, a human-collaborative robot AI model designed to read human intent and work alongside people.
ASUS announced the XA P2N-E2, an AI server aimed at data center workloads. No pricing or availability details were disclosed in the source.
RoboParty launched RP1, a full-stack open-source humanoid robot, part of a broader open-source robotics boom that includes robot ducks, crabs and humanoids.
Why it matters: Indicates where China's humanoid robotics industry sees the next competitive bottleneck.
Reuters reports China's humanoid robot makers have mastered movement and now face the harder race for AI brains, with focus shifting to intelligence and cognition.
Qwen AI Glasses gained real-time call translation for calls and meetings, media translation in the companion app, and improved speech recognition in noisy places via an OTA update.
Why it matters: Ties OpenAI's compute needs to Korean memory supply amid a chip shortage.
Samsung hosted OpenAI CEO Sam Altman in Yongin to discuss AI-era business strategies with presidents of its major affiliates, as a memory chip shortage deepens and Altman bets on Samsung and SK.
Why it matters: xAI's breakneck Blackwell buildout is one of the largest single-site AI compute pushes.
Elon Musk said xAI is taking over an old Spam factory and a gun range to expand its Colossus 2 AI complex toward roughly 1.2M Nvidia Blackwells by year-end, adding 'Gun Range' and 'Spam' clusters.
YYForce has put 11 autonomous delivery robots to work across two Singapore hotels, handling item delivery for guests. The fleet is a commercial rollout of indoor service robots in hospitality.
nv-uefi-unlocker is an open-source UEFI tool that lifts Nvidia's half-speed tensor rate cap on RTX 20, 30 and 40-series desktop and laptop GPUs under Windows, raising FP16 math performance up to 65%.
Google could soon let parents limit how long their kids can use Gemini chats, according to Android Authority.
ON1 released Photo RAW 2027 for Mac and Windows, adding a new RAW engine and revised NoNoise AI models. The AI Culling Assistant is exclusive to the MAX edition.
Readwise Books combines an ebook store, reading tracker, detailed reading stats and AI features, aiming to challenge Amazon's reading ecosystem.
AWS post shows how to use Crawl4AI, an open source web crawler, to prepare web content for Amazon Bedrock Knowledge Bases, handling JavaScript rendering, selective extraction and metadata enrichment before data reaches S3.
Momentic (YC W24), launched on Hacker News, offers AI agents that test applications by using them like a user to catch bugs.
AWS post shows how sequence-level knowledge distillation on Amazon SageMaker AI transfers a frontier teacher model's agentic behavior into a smaller, lower-cost model for production agents.
AWS post shows a multi-agent engineering assistant using Amazon Nova Multimodal Embeddings, Bedrock AgentCore and the Strands Agents SDK to cut aerospace component substitution evaluation from days to minutes.
Granola is an AI-powered notepad that listens to meetings and turns transcripts into structured notes. Version 7.726.2 is a minor app update with no details in the source.
Kimi is an AI chat assistant from Moonshot AI. Version 3.2.24 is a routine desktop app update with no new features described in the source.
A new Chrome AI feature uses Gemini to read a user's Gmail and autofill web forms with the details found there. Engadget explains how to disable it if users don't want Gemini accessing email.
BoltAI is a desktop AI chat client that lets you use multiple AI models from one app. Version 2.18.4 is out for macOS, Windows and Linux.
A Reddit user shares Pinokio Remote, which lets people reach their locally hosted AI apps remotely. It is a community post with a link.
A report says being rude to a particular AI chatbot could get users banned, explaining the reason behind the policy.
Drive Tesla Canada reports Tesla could soon add xAI's Grok assistant to its Intel-based vehicles, expanding the chatbot beyond newer hardware.
A shattered.io post compiles Oracle's documentation for its OCI Generative AI service, listing the models and capabilities available to cloud customers.
An open-source TypeScript project that lets a company set up Claude Code and Codex once, self-hosted, with shared company files, team skills, MCP and per-person folder permissions.
Why it matters: Puts patients' own medical records under chatbot interrogation, a notable health-AI integration.
Athenahealth launched a tool that brings patient health records into Anthropic's Claude, letting patients ask the chatbot questions about their medical data.
LinearB launched Agentic Studio, which turns software-delivery data into an engineering context layer that leaders and coding agents can query in natural language to make decisions.
MIT Technology Review argues developers should measure shipped output rather than token counts, since agentic coding tasks use 10-50x more tokens than chat prompts.
dsh-see-now is a JavaScript tool that captures your screen, filters frames by how much they change and tiles several into one image, so AI can watch your screen in real time without spending tokens on static content.
A Hugging Face repository, HannesVonEssen/microduck-basketball, is listed under the robotics task. The listing describes a public repository, not an inference endpoint; a simready-glb dataset is also posted.
The article walks through building an enterprise AI stack covering models, agents, the Model Context Protocol and governance.
The piece traces Tim Berners-Lee's Semantic Web vision to today's agentic web, where autonomous software agents navigate and act on the internet.
A Notebookcheck comparison details what current ChatGPT, Gemini and Claude AI subscriptions cost.
The Verge reports AI tools are being used to mass-produce indie games, with a surge of AI-generated titles reaching storefronts.
A browser-based AI platform aims to streamline early-stage visual planning, letting users quickly produce and iterate concept visuals.
A portable local studio bundles a complete offline AI toolkit on a single USB drive, letting users run AI tools without an internet connection or cloud account.
A command-line interface for Amp, the agentic coding and development environment, installable via npm i -g @ampcode/cli. This is a routine version bump.
Xmind 2027 is a mind-mapping app that gives users and AI a shared visual workspace. It adds Xmind Agent, which works from local documents and maps, and MCP support to bring that context into ChatGPT and Claude.
Cedric is an AI agent that joins Zoom, Meet or Teams calls, presents your slides and answers questions using live data from your tools. It also catches you up if you join late and drafts notes and follow-ups for approval.
A Hugging Face repository, HannesVonEssen/microduck-basketball, is listed under the robotics task. The listing describes a public repository, not an inference endpoint; a simready-glb dataset is also posted.
Phonon is an open text-to-speech model family; this Hugging Face Space demo runs it entirely in your browser in English, French, German, Spanish and Portuguese, with no audio leaving the device.
OBLITERATUS published DeepSeek-V4.1-Flash-1017-obliterated, a refusal-direction ablated version of DeepSeek V4.1 Flash with BF16 weights under MIT, reporting a 4% refusal rate.
A Hugging Face Space demos realtime text-to-image and multi-image editing with Qwen-Image-2.1, using a 2-step DMD2-distilled LoRA fused into the weights and AOTInductor kernels on ZeroGPU, streaming 512px results live.
An open-source TypeScript project that lets a company set up Claude Code and Codex once, self-hosted, with shared company files, team skills, MCP and per-person folder permissions.
dsh-see-now is a JavaScript tool that captures your screen, filters frames by how much they change and tiles several into one image, so AI can watch your screen in real time without spending tokens on static content.
JavaScript GitHub project adding a memory layer to the AstrBot chatbot: five memory types, hybrid retrieval, forgetting rules, emotional boundaries and an autonomous life loop, without altering personality.
JavaScript GitHub project for running language-model agents inside Minecraft with Mineflayer tools, persistent sessions and shell sandboxes.
NextTransformer is an open-source Python framework for next-generation Transformer architectures, offering compute-matched benchmarking and a platform for autonomous architecture research. It drew 101 stars in its first week.
FlashAlign is a Python tool that re-aligns AI models instantly, either per request for verified users and channels or always on, and works with Hermes, OpenClaw and custom harnesses. It reached 100 stars in its first week.
omux is a native macOS terminal app for running many Claude Code and Codex sessions side by side, with projects, worktrees, live agent status, sidebar allow/deny, resumable conversations and PR status.
uncut is a 7 MB coding-agent harness for shell, CI and local models that claims lossless context with no silent truncation and no TUI. Written in Go, it gained 121 GitHub stars in its first week.
AnyModelBot is an open-source, local-first alternative to hosted AI bots that runs agents on 70+ models via one Modelflare key, reachable in-app, WeChat or Feishu, with memory and computer control.
graphmemory is an open-source memory system for AI apps: it stores facts bitemporally, applies time-graded recall and decay, expires entries by TTL and dedupes, backed by Neo4j, PostgreSQL and Redis.
Tucano is a GitHub project for running a team of AI agents on your Mac, Windows or Linux machine. Agents talk in channels, write and review code, and use your tools on Claude Code, Codex, OpenCode or Pi; risky actions wait for approval.
my-office is a JavaScript GitHub project that renders Claude Code sessions as a 3D office, giving a visual view of agent activity; it drew 105 stars in week one.
LobeChat is an open-source AI chat and agent platform. Version 2.3.0 adds Proactive Bot, letting bots act without being @-mentioned, with 145 merged PRs since v2.2.21.
A zero-to-one beginner tutorial that doubles as a runnable local agent: file and command access, MCP, long-term memory and a web UI showing each step. Builds up from a 100-line loop and dissects Claude Code and Codex.
cheap-ag is a TypeScript MCP server for Claude Code in which Opus plans while a cheaper model writes code and tests gate output, letting a $20 plan go further.
Google could soon let parents limit how long their kids can use Gemini chats, according to Android Authority.
ON1 released Photo RAW 2027 for Mac and Windows, adding a new RAW engine and revised NoNoise AI models. The AI Culling Assistant is exclusive to the MAX edition.
Readwise Books combines an ebook store, reading tracker, detailed reading stats and AI features, aiming to challenge Amazon's reading ecosystem.
Granola is an AI-powered notepad that listens to meetings and turns transcripts into structured notes. Version 7.726.2 is a minor app update with no details in the source.
Kimi is an AI chat assistant from Moonshot AI. Version 3.2.24 is a routine desktop app update with no new features described in the source.
A new Chrome AI feature uses Gemini to read a user's Gmail and autofill web forms with the details found there. Engadget explains how to disable it if users don't want Gemini accessing email.
BoltAI is a desktop AI chat client that lets you use multiple AI models from one app. Version 2.18.4 is out for macOS, Windows and Linux.
A report says being rude to a particular AI chatbot could get users banned, explaining the reason behind the policy.
Drive Tesla Canada reports Tesla could soon add xAI's Grok assistant to its Intel-based vehicles, expanding the chatbot beyond newer hardware.
Athenahealth launched a tool that brings patient health records into Anthropic's Claude, letting patients ask the chatbot questions about their medical data.
A Notebookcheck comparison details what current ChatGPT, Gemini and Claude AI subscriptions cost.
The Verge reports AI tools are being used to mass-produce indie games, with a surge of AI-generated titles reaching storefronts.
A browser-based AI platform aims to streamline early-stage visual planning, letting users quickly produce and iterate concept visuals.
Xmind 2027 is a mind-mapping app that gives users and AI a shared visual workspace. It adds Xmind Agent, which works from local documents and maps, and MCP support to bring that context into ChatGPT and Claude.
Cedric is an AI agent that joins Zoom, Meet or Teams calls, presents your slides and answers questions using live data from your tools. It also catches you up if you join late and drafts notes and follow-ups for approval.
CuriousLM is a free AI app that runs the model on your phone or computer, so chats stay local and encrypted with no account or server. It underlines uncertain words you can tap, and includes on-device voice and Pocket Wikipedia.
Version 2.91.0.607 of Tencent's AI assistant Yuanbao for macOS, Windows and Linux, which runs on Hunyuan and DeepSeek models.
Browser-based AI tools let creators turn text and images into visual concepts.
Reports say ChatGPT will soon use health data from Android devices, such as Google Health Connect, to give users more useful personalized health tips and insights.
Malaysian developer i-City introduced i-Ris, a 'stay home AI' system for future households built on Nvidia technology, aimed at home automation and assistance.
Muse is an AI agent launching in Singapore with connectors for Grab, Shopee, Singapore Airlines and more, letting users act across services.
A technical issue disrupted Instinct's personal AI agent on WhatsApp, including payment features; the service has since been restored. Instinct's agent handles tasks via WhatsApp chat.
A Hacker News discussion highlights how AI-generated fan wikis riddled with errors are muddying games culture, raising questions about reliability of community reference sites.
Google Gemini now lets users turn their face and voice into a shareable AI avatar, and adds in-chat editing of PDFs, Slides and Sheets. Separately, a Gemini Gmail suit was refiled after a judge found no concrete harm.
'Searching for SmarterChild' filmmakers explore why people hurl abuse at chatbots, examining human-AI interaction and the emotional bonds users form with conversational agents.
AWS Machine Learning · 1h
#
AWS post shows how to use Crawl4AI, an open source web crawler, to prepare web content for Amazon Bedrock Knowledge Bases, handling JavaScript rendering, selective extraction and metadata enrichment before data reaches S3.
Momentic (YC W24), launched on Hacker News, offers AI agents that test applications by using them like a user to catch bugs.
AWS Machine Learning · 2h
#
AWS post shows how sequence-level knowledge distillation on Amazon SageMaker AI transfers a frontier teacher model's agentic behavior into a smaller, lower-cost model for production agents.
AWS Machine Learning · 2h
#
AWS post shows a multi-agent engineering assistant using Amazon Nova Multimodal Embeddings, Bedrock AgentCore and the Strands Agents SDK to cut aerospace component substitution evaluation from days to minutes.
A Reddit user shares Pinokio Remote, which lets people reach their locally hosted AI apps remotely. It is a community post with a link.
A shattered.io post compiles Oracle's documentation for its OCI Generative AI service, listing the models and capabilities available to cloud customers.
LinearB launched Agentic Studio, which turns software-delivery data into an engineering context layer that leaders and coding agents can query in natural language to make decisions.
MIT Technology Review argues developers should measure shipped output rather than token counts, since agentic coding tasks use 10-50x more tokens than chat prompts.
The article walks through building an enterprise AI stack covering models, agents, the Model Context Protocol and governance.
The piece traces Tim Berners-Lee's Semantic Web vision to today's agentic web, where autonomous software agents navigate and act on the internet.
A portable local studio bundles a complete offline AI toolkit on a single USB drive, letting users run AI tools without an internet connection or cloud account.
A command-line interface for Amp, the agentic coding and development environment, installable via npm i -g @ampcode/cli. This is a routine version bump.
Display 1.0 is a workspace where teams and their AI agents create docs, prototypes and apps together. It stores agent-made files, organizes knowledge in Spaces and Folders, and works with any agent at a flat price.
Agent Inbox is an API that gives each AI agent its own email address, Slack handle and Telegram handle. It works with Claude, Grok, OpenAI and other agent frameworks, letting you message your agent and get replies in the same thread.
Hyperresearch is an agentic research tool that returns cited reports and checks each cited claim against its source, keeping sources in a searchable vault. It adds Answers MCP and API so agents can get short quoted answers or a 'not found' instead of a hallucination.
Widgeteer lets you describe a widget, like a quote calculator or booking form, and builds a working one you can edit by hand or by asking. Submissions go to email, Zapier, n8n or a webhook, and it works on WordPress, Shopify and any site.
A step-by-step guide to deploying a serverless AI agent on AWS in about 90 minutes, covering setup through deployment.
Perfect Corp launched YouCam CLI, letting developers run YouCam AI APIs from a terminal or an AI agent.
Dataiku introduced Headless, which brings governed data and AI work into agentic coding tools.
Gregory Kurtzer, creator of CentOS and Rocky Linux, says he built a new kernel that is not Linux in 11 days using AI assistance. The project is separate from Rocky.
Berth is a lightweight open-source GPU resource manager written in Go that grants applications access to a GPU through leases, priorities and a request queue, letting users live-swap workloads like llama.cpp, ComfyUI and Strata.
OccupySaaS, founded October 2026, released ZeroAuthy as an Auth0 replacement and SalesCompel as a Salesforce replacement, both AI-built and Apache-2.0 licensed.
Stack Overflow's senior VP of community says reports of the site's death are exaggerated, arguing AI changed how developers access knowledge but the knowledge and company behind it remain.
An open project built with Claude Code that downloads a local AI model, Wikipedia, maps and survival manuals onto any drive in one command, shipping as a Claude Code skill.
A commentary argues a 100% pull-request merge rate can't show how much engineer time went into reviewing, correcting and approving AI-written code, so it fails to prove review costs fell.