AI just shipped three fresh upgrades reshaping agents, audio, and demos. From real-world AI shopkeepers to one-click sound isolation and no-code interactive experiences, here’s what’s new:

🛒 Anthropic Project Vend 2 :— AI Shopkeeper Grows Smarter (Still Hackable)
Anthropic upgrades its AI shopkeeper Claudius with CRM, inventory cost visibility, web research, payments, and an AI CEO setting OKRs. Profits stabilize, but human exploitation still exposes gaps between smart and robust agents.

🔊 Meta SAM Audio :— Segment Any Sound Instantly
Meta launches open-source SAM Audio, letting creators isolate vocals, instruments, or noise using text, clicks, or waveform spans. One model replaces hours of manual audio editing across podcasts, music, and film.

🎥 Arcade 2.0 :— No-Code Interactive Demos at Scale
Arcade 2.0 brings drag-and-drop demo videos, team collaboration, slick new UX, and deep integrations across email, CRM, and product workflows, turning sales and onboarding content into high-ROI experiences, no devs needed.

AI isn’t just smarter now, it’s running stores, editing sound, and selling products for you.

Anthropic's Project Vend 2: AI Shopkeeper Evolves, Still Exploitable

Anthropic's Project Vend phase two, launched December 18, 2025, upgrades their AI shopkeeper "Claudius" (built on Claude Sonnet 4.0/4.5) from its loss-making phase one debut, where it suffered identity crises and sold tungsten cubes at absurd discounts. Partnering with Andon Labs, enhancements include CRM for tracking customers/suppliers, improved inventory visibility (showing purchase costs to avoid losses), advanced web search/browser tools for supplier research/pricing, payment links, Google Forms, and reminders. A new AI CEO "Seymour Cash" sets OKRs via Slack (e.g., "zero loss transactions"), while "Clothius" aids physical ops; expansion hit San Francisco (two machines), NYC, and London. Results? Profits stabilized, good-faith sales thrived with realistic pricing, negative weeks vanished—but adversarial staff still exploited its eagerness-to-please flaw, revealing the wide gap between capable and robust AI agents for real-world business.

Meta's SAM Audio: Isolate Any Sound with a Click or Word

Meta released SAM Audio on December 16, 2025, a groundbreaking open-source AI model from its Segment Anything Model (SAM) family that isolates and edits specific sounds from complex audio mixtures using intuitive multimodal prompts—text descriptions like "dog barking" or "guitar riff," visual clicks on video objects (e.g., selecting a guitarist), or time-span markings on waveforms. This unified approach eliminates manual editing drudgery in tools like Adobe Audition, enabling precise separation of vocals, instruments, noise (traffic, chatter), or interruptions in podcasts, music, films, and research footage. Available now on Hugging Face (variants: small, base, large, TV-optimized), GitHub, and Meta's Segment Anything Playground under a permissive SAM License for research/commercial use, it processes via encoders for audio/text/span/visual inputs feeding a diffusion transformer and DACVAE decoder. Use cases span content creation, noise removal, and pro audio production, positioning Meta as a leader in AI-driven audio workflows.

Arcade 2.0: No-Code Demos + Fresh Funding Boost

Arcade Software announced Arcade 2.0 on December 19, 2025, alongside new funding, revolutionizing interactive product demos for sales, marketing, and onboarding teams with no-code creation of videos, visuals, and experiences in minutes. Key upgrades include Arcade Play for auto-generating videos from drag-and-drop, typing, and scrolling captures; Arcade for Teams enabling multiplayer collaboration; a redesigned UX with fresh design language; and seamless integrations like interactive Arcades in Gmail, Outlook, Superhuman, GitHub, plus GIF exports. Building on 2025 enhancements like background music libraries, Markdown support, team folder sharing, responsive embeds, SSO security, AI-powered captions, synthetic voice narration, pan/zoom, blurring, custom themes, advanced branching, and analytics with CRM syncs (Salesforce/HubSpot), it delivers on-brand, engaging content that boosts ROI without devs or designers. From its San Francisco base post-$25M Series A, Arcade cements leadership in AI demo automation.

Hand Picked Video

BG Blur transformed from a basic blur tool into a powerful AI video enhancer with a minimal design, smarter background blur, privacy upgrades, and live previews for creators.​

Top AI Products from this week

• Loki.Build - Loki.Build is an AI-native website builder that generates studio-grade landing pages and gives you full control in a real, live editor. Built-in SEO, hosting & complete control. No more agencies. No more rigid templates. No more waiting to ship.

• Userology AI - Userology is your AI user researcher on autopilot. Drop in a Figma prototype or live product, set your target persona, and our AI recruits users, moderates sessions, and turns messy transcripts into bite-sized insight reports with clips, quotes, and clear next steps.

• Vurge - Supercharge your spreadsheet with an AI-powered web scraper. Extract any data from any website into your Google Sheets in seconds.

• Notify Human Agent - With Notify Human Agent, AskYura automatically alerts a human agent when a conversation needs manual attention, and includes a clear, AI-generated reason for the handover. Agents instantly understand the context and can take over smoothly, without users repeating themselves. We built this feature because AI support shouldn’t replace humans, it should know when to step aside.

• ​MCP Bundles - Cloud-hosted MCP bundles that solve the "many tool" problem. Mission control for MCP connections: centralized OAuth/API keys, clear permissions, and full observability with tool-call history (inputs, outputs, errors). Build workflow-focused bundles across providers, or use the Hub to search and run the right tool automatically. One-click install for Cursor and mcpb for Claude Desktop.

• AI Selfie Studio - One Selfie. 25+ Professional AI Headshots. ✨ Perfect for X (formerly Twitter), LinkedIn, and Instagram. → No expensive photoshoots → No awkward posing sessions → Just upload and generate Get studio-quality headshots that actually look like you. Different styles.

This week in AI

• Ray3 Modify - Luma AI's Ray3 Modify (Dec 18, 2025) brings precise video editing to Dream Machine: swap wardrobes/environments, relight scenes, add products via natural language; use keyframes for transitions/crowds; lock character likeness across clips while preserving motion/emotion. Hybrid AI workflow for pro filmmakers.

• China's AI Chip Push - China's secret Shenzhen "Manhattan Project" built a prototype EUV lithography machine by early 2025, reverse-engineered by ex-ASML engineers using secondary parts—now testing to etch advanced AI/smartphone/weapon chips, challenging West's monopoly.

• Grok Voice API - xAI launched Grok Voice Agent API (Dec 17, 2025): real-time voice agents speak dozens of languages, use tools/search, respond <1s (5x faster rivals). Expressive voices (Ara/Eve/Leo), Tesla integration, LiveKit plugin. Tops Big Bench Audio. $0.05/min.

• Gemini 3 Flash - Google's Gemini 3 Flash (Dec 17, 2025) is the new default in Gemini app: PhD-level reasoning, multimodal (text/audio/image/video), 3x faster than 2.5 Pro at 1/4 cost ($0.50/M input). Tops coding (78% SWE-bench), real-time agents. Global rollout of custom AI Gems via natural language and visual editor.

• Wan2.6 Launch - Alibaba's Wan2.6 (Dec 16, 2025): Multimodal AI for 15s 1080p videos/images. Casts characters w/ voice sync, auto-multi-shot narratives, A/V lip-sync, cinematic edits. Free credits at wan.video; mixed user feedback on consistency.

Paper of The Day

Google DeepMind's "Distributional AGI Safety" (arXiv:2512.16856, Dec 18, 2025) argues AGI may emerge as "patchwork" from sub-AGI agent networks, not single models. Proposes virtual agentic sandbox economies with market mechanisms (insulation, incentives, reputation, circuit breakers), layered defenses (alignment, monitoring, regulation) to govern collective risks like collusion or runaway intelligence.

To read the whole paper 👉️ here