Get tomorrow's brief in your inbox
Today: Anthropic's Claude Cowork expands to mobile and web, letting business tasks run in the cloud even with your laptop closed. SambaNova raised $1 billion at $11B valuation as JPMorgan Chase picks its chips for on-premises AI inference. Meta's new Muse Image generator lets users manipulate photos from any public Instagram account, raising privacy concerns.
Anthropic launched Claude Cowork on mobile and web this week, breaking its desktop-only limitation and adding cloud-based background execution. The expansion means you can start a research task at your desk, check progress updates on your phone during lunch, and pick up the finished briefing when you get home. The key shift is that Cowork tasks now run in Anthropic's cloud by default, continuing even when your laptop is closed.
Claude Cowork mobile launch | TechCrunch coverage
Usage data from 1.2 million sessions shows Cowork's strongest use case isn't coding (only 8.7% of usage). Instead, 33.4% of sessions involve "business process operating": pulling scattered updates into reports, building onboarding checklists, and reconciling spreadsheets. Another 16.4% is content creation like drafts, slide decks, and proposals. This is the work that takes time but isn't anyone's core job, which is exactly where AI assistants deliver value without replacing specialized expertise.
If you run a service business and spend hours each week compiling client updates or building proposal decks, Cowork can handle the synthesis work overnight. The mobile expansion means you can review and approve the output during your commute instead of blocking out desk time. The desktop app remains the full-featured version with local file access, but web and mobile access removes the installation barrier for teams testing AI workflows.
Slackbot gets smarter with team memory and cross-app integrations. Slack upgraded Slackbot with persistent memory that learns your team's context and workflows. The bot now connects to Google, Atlassian, and DocuSign, letting you pull information across tools without leaving Slack. New voice commands let you dictate updates hands-free. One Salesforce engineer built a skill that logs customer calls automatically, saving his team 43 minutes per week. Read more
If you're already paying for Slack, the upgrades are free. The memory feature means you don't waste time re-explaining your processes every time you ask for help. The cross-app integrations are the real value for small teams juggling too many tools.
Microsoft Foundry now runs Hugging Face models with managed GPU infrastructure. Microsoft's Foundry platform added support for 3 million open-source models from Hugging Face, with managed GPU compute that handles container updates and security patches automatically. You pick a model by parameter count and context length, and Foundry handles the GPU topology underneath. All models share a single endpoint, authentication, and billing system. Announcement
This matters for businesses that want control over their AI stack without managing infrastructure. Open models let you fine-tune on proprietary data and run inference inside your own network boundaries. If you've been avoiding open-source models because deployment looked complicated, Foundry removes that friction.
Google's Gemini API adds background execution and remote tool integration. Google shipped managed agents that can run tasks in the background for hours without holding an HTTP connection open. Agents can now connect to remote Model Context Protocol (MCP) servers, letting them access private databases and internal APIs without building custom middleware. Details
The background execution is critical for long-running work like report generation or data analysis. If you're building internal tools that need to query your CRM and accounting system in one workflow, remote MCP integration makes that possible without exposing credentials to client code.
Amazon SageMaker adds one-click onboarding from Hugging Face. You can now click "Customize on SageMaker AI" or "Deploy on SageMaker AI" on supported Hugging Face models and land directly in a pre-configured SageMaker Studio environment with permissions already set. The flow includes GPU quota visibility so you know which instances are available before requesting them. Walkthrough
This removes the friction between discovering a model and running fine-tuning experiments. If you've been intimidated by AWS setup, this is the streamlined path to production model deployment.
ZML releases free inference software that runs across Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc. French startup ZML launched LLMD, an inference server that optimizes open-source models for multiple chip architectures at maximum speed. The goal is to break vendor lock-in and let enterprises mix cheaper or lower-power chips in production. Launch story
This is free for now while ZML measures adoption. If you're running inference at scale and Nvidia costs are squeezing margins, ZML might let you diversify hardware without rewriting infrastructure.
Savi Security launches AI scam detection app for iPhone and Android. Savi raised $7 million to protect consumers from voice-cloned kidnapping scams and AI-generated phishing. The app analyzes suspicious texts, emails, and calls in real time. Founded by former Cisco security executives after a scammer used AI to clone a family member's voice and demand ransom. TechCrunch coverage
AI scams cost consumers $3.5 billion in 2025, triple the 2020 losses. If you have elderly clients or family members, this is a practical defense against increasingly sophisticated social engineering.
Simon Willison releases sqlite-utils 4.0 with database schema migrations. The popular SQLite toolkit now includes Python-based migration management with tracking for applied changes. Migrations use the powerful table.transform() method for schema changes SQLite doesn't natively support. Details
If you're building internal tools or prototypes that store data in SQLite, this gives you Django-style migration management without the ORM overhead.
Zapier's marketing exec built a portfolio of AI agents without writing code. Amanda Natividad demonstrated "vibe coding" by building internal tools using Zapier Copilot, Lovable, Cursor, and Bolt. Her journalist response agent filters PR pitches, uses a RAG model to draft responses based on her existing content, and shares qualified opportunities in Slack. A news-to-social agent creates LinkedIn and X posts from trending AI automation stories. She also built a Slack group landing page, a restaurant social network, and a Chrome writing assistant. Full portfolio
The takeaway: non-engineers can build functional automation and tools with today's AI coding assistants. If you have repetitive workflows that require context-aware decisions, AI agents can handle the grunt work while you approve the output.
Salesforce engineer saves team 43 minutes per week with one Slackbot skill. Raveesh Raina built a Slackbot skill that logs customer call activities automatically after meetings. His colleagues adopted it, and the workflow proliferated across the team because it solved a universal pain point. "That's what it looks like when one person's best process becomes everyone's capability," he said.
The pattern is: identify a task everyone does manually, automate it once, and share it broadly. The return on investment compounds as more people adopt the workflow.
Claude Cowork usage limits doubled through August 5. Anthropic extended doubled Cowork limits for Max subscribers during the mobile and web rollout period. Standard limits return after August 5.
Microsoft shifts Office 365 from OpenAI/Anthropic models to its own MAI models. Excel and Word now use Microsoft's homemade models for a percentage of user prompts to cut costs. This follows a broader trend of companies reducing reliance on expensive third-party AI after a brief "tokenmaxxing" phase earlier this year. Amazon, Uber, Meta, and Accenture have made similar moves. Bloomberg report
If you're building on Microsoft's AI features, quality may vary as the company shifts underlying models. Watch for changes in output consistency or accuracy.
Meta's Muse Image is free for "everyday creation" but requires a subscription past a certain usage limit. Exact thresholds weren't disclosed. The model powers AI image generation across Meta AI, Instagram, and WhatsApp. Announcement
SambaNova raises $1 billion at $11B valuation, plans 2027 IPO. The AI chip startup closed the first tranche of its Series F led by General Atlantic, with more investors joining soon. JPMorgan Chase selected SambaNova as its inference infrastructure partner, deploying the SN40L and SN50 chips for secure on-premises AI. CEO Rodrigo Liang said the capital secures supply chain capacity against surging demand. CNBC exclusive | TechCrunch story
The JPMorgan win signals that banks are building private inference infrastructure instead of relying entirely on cloud providers. If you work in financial services or regulated industries, on-premises inference is becoming a competitive requirement for handling sensitive models.
Even Realities (China smart glasses startup) hits $1B valuation, says Shenzhen beats Silicon Valley for hardware. CEO Will Wang, a former Apple executive, raised $150 million from Meituan and Tencent. He argues that Silicon Valley no longer rewards hardware founders, while Shenzhen offers deep engineering talent and supply chain proximity. Even's G2 smart glasses have no camera and deliver notifications, navigation, and translation through a heads-up display. CNBC interview
Meta faces privacy backlash over Muse Image feature that manipulates public Instagram photos. Muse Image lets users tag any public Instagram account to pull photos into AI-generated images without notifying the account owner. Meta's policy states users "will not be notified about content created using AI features at Meta," though settings exist to opt out. The feature is already raising privacy concerns, fitting a pattern of opt-out-by-default data practices. The Verge analysis | TechCrunch story
If you have a public Instagram business account, check your settings under AI reuse controls. This is live now across Meta AI, Instagram Stories, and WhatsApp.
Discord admits AI moderation bug wrongfully banned 8,000+ users over harmless images. The platform's AI moderation system flagged spreadsheets, chessboards, game textures, and blank backgrounds as harmful content and immediately banned accounts before human review. All affected accounts are being restored, but the incident highlights risks of automated content moderation at scale. Details
Automate journalist outreach responses with a RAG agent. Build a Zapier workflow that monitors PR request emails, scores them for relevance, and drafts responses using your existing content stored in a vector database. Share the best opportunities in Slack for quick review and send. This saved Amanda Natividad hours each week and turned press opportunities into qualified leads.
Use Slackbot voice commands to log meeting notes immediately after calls. Instead of blocking time later to document client conversations, dictate the summary to Slackbot as you walk out of the meeting. The bot can log the activity directly in your CRM or project management tool.
Check your Instagram AI reuse settings if you have a public business account. Meta's Muse Image can now pull your photos into AI-generated content without notification. Go to Instagram settings and disable AI content reuse if you want control over your brand imagery.
This week shows AI agents moving from experimental desktop apps to production tools that run across devices and in the background. Anthropic's mobile expansion, Google's background execution, and Microsoft's cross-platform integrations all point toward AI that fits into existing workflows rather than requiring dedicated sessions.
The spending patterns reveal a two-tiered market forming: frontier models like Claude and GPT-5 for discovery and complex reasoning, and open-source models for production workloads that have been refined. Companies are cutting costs by moving mature use cases to cheaper models while still relying on expensive frontier models for new problems. This isn't a zero-sum competition; it's a life cycle where both tiers serve different needs.
Privacy and moderation challenges are emerging faster than safeguards. Meta's Instagram photo manipulation and Discord's mass false-positive bans show the friction between AI's capabilities and user trust. Businesses adopting these tools need to audit data handling practices and moderation policies before they become PR problems.
Watch for more one-click deployment integrations like the Hugging Face-to-SageMaker flow. The pattern of "discover a model, deploy it instantly with pre-configured infrastructure" removes adoption friction and will accelerate enterprise experimentation with open-source AI.