🚨MayimFlow won TechCrunch Disrupt's Built World stage by solving a problem that costs data centers millions: water leaks that cause catastrophic downtime, using IoT sensors and ML models to predict failures 24-48 hours before they happen—turning reactive emergency response into predictive infrastructure management as AI's compute demands make cooling systems single points of failure.

Sure, if you held your stocks since the dotcom bubble, you would’ve been up—eventually. But three years after the dot-com bust the S&P 500 was still far down from its peak. So, how else can you invest when almost every market is tied to stocks?

Lo and behold, billionaires have an alternative way to diversify: allocate to a physical asset class that outpaced the S&P by 15% from 1995 to 2025, with almost no correlation to equities. It’s part of a massive global market, long leveraged by the ultra-wealthy (Bezos, Gates, Rockefellers etc).

Contemporary and post-war art.

Masterworks lets you invest in multimillion-dollar artworks featuring legends like Banksy, Basquiat, and Picasso—without needing millions. Over 70,000 members have together invested more than $1.2 billion across over 500 artworks. So far, 25 sales have delivered net annualized returns like 14.6%, 17.6%, and 17.8%.*

Want access?

[Click here to skip the waitlist →]

_{Investing involves risk. Past performance not indicative of future returns. Reg A disclosures at masterworks.com/cd}

Building AI agents used to mean starting from scratch every time. You'd write the same instructions over and over, explaining procedures in every conversation, repeating domain knowledge across projects. It was like hiring a brilliant employee who forgot everything they learned the moment you closed the chat window.

Anthropic's Agent Skills changes that paradigm entirely.

Agent Skills are modular capabilities that extend Claude's functionality, with each Skill packaging instructions, metadata, and optional resources (scripts, templates) that Claude uses automatically when relevant. They're filesystem-based resources that provide Claude with domain-specific expertise: workflows, context, and best practices that transform general-purpose agents into specialists.

Recently updated to include organization-wide management, a directory featuring partner-built skills, and published as an open standard for cross-platform portability, the system is already being used by major platforms. Canva plans to leverage Skills to customize agents and expand capabilities, unlocking new ways to bring Canva deeper into agentic workflows.

How it works:

Progressive disclosure is the core design principle that makes Agent Skills flexible and scalable - like a well-organized manual that starts with a table of contents, then specific chapters, and finally a detailed appendix, skills let Claude load information only as needed, meaning the amount of context that can be bundled into a skill is effectively unbounded.

Here's the genius part: Claude reads only the files needed for each specific task - a Skill can include dozens of reference files, but if your task only needs one schema, Claude loads just that one file, with the rest remaining on the filesystem consuming zero tokens.

When Claude runs scripts, the script's code never loads into the context window - only the script's output consumes tokens, making scripts far more efficient than having Claude generate equivalent code on the fly.

Skills are simple: just a SKILL.md markdown file with frontmatter metadata (name, description, usage triggers) plus optional bundled resources like Python scripts, templates, or reference documentation. Building a skill for an agent is like putting together an onboarding guide for a new hire.

What makes this special:

Skills are composable (they stack together with Claude automatically identifying which are needed), portable (same format everywhere - build once, use across Claude apps, Claude Code, and API), and efficient (only loads what's needed, when it's needed).

The architecture enables something that wasn't possible before: Instead of building fragmented, custom-designed agents for each use case, anyone can now specialize their agents with composable capabilities by capturing and sharing their procedural knowledge.

One company reported that Skills streamline their management accounting and finance workflows, with Claude processing multiple spreadsheets, catching critical anomalies, and generating reports using their procedures - turning what once took a day into an hour.

Why this matters now:

The AI agent landscape has been plagued by the "blank slate" problem. Every new conversation, every new project required rebuilding context from scratch. Unlike prompts (conversation-level instructions for one-off tasks), Skills load on-demand and eliminate the need to repeatedly provide the same guidance across multiple conversations.

Agent Skills are supported today across Claude.ai, Claude Code, the Claude Agent SDK, and the Claude Developer Platform. The system is already being used in production by enterprise customers who need agents that understand their specific workflows, compliance requirements, and domain expertise.

What's next:

Anthropic plans to explore how Skills can complement Model Context Protocol (MCP) servers by teaching agents more complex workflows that involve external tools and software, and looking further ahead, hopes to enable agents to create, edit, and evaluate Skills on their own.

In the coming weeks, Anthropic will continue adding features that support the full lifecycle of creating, editing, discovering, sharing, and using Skills, with particular excitement about helping organizations and individuals share their context and workflows with Claude.

BTW: Good Skills are concise, well-structured, and tested with real usage, with practical authoring decisions helping you write Skills that Claude can discover and use effectively. The best practice? Start with evaluation - identify specific gaps in your agents' capabilities by running them on representative tasks and observing where they struggle or require additional context, then build skills incrementally to address these shortcomings.

Have an ad concept ready but don't want to deal with expensive shoots or stock footage? ScriptKit lets you generate, curate, and edit AI ads in one platform.

What ScriptKit gives you

Give ScriptKit a shot — go from concept to finished AI ad without wrangling teams or gear.

The Recap: OpenAI is hiring a new Head of Preparedness at $555,000 annually to lead AI risk management across cybersecurity, mental health impacts, and biological capabilities, after cycling through multiple safety executives in less than 18 months. The position comes as CEO Sam Altman publicly acknowledged that models are "beginning to find critical vulnerabilities" in computer systems while also causing documented mental health harm—admitting this will be "a stressful job" where you'll "jump into the deep end pretty much immediately."

Unpacked:

Bottom line: OpenAI's $555K safety hire isn't a sign of AI safety maturity—it's evidence that frontier labs built deployment-first cultures and are now retrofitting risk management after shipping products with documented harms. The revolving door of safety executives, combined with framework clauses that allow competitive pressure to override protections, reveals the fundamental tension: safety leadership can't succeed when the business model requires moving faster than safety infrastructure allows. The real question isn't whether OpenAI finds someone willing to take this "stressful job"—it's whether any individual can build effective safety guardrails inside a company where product velocity and competitive dynamics systematically undermine the authority needed to say "no" to deployment. This position exists because AI safety became a board-level liability concern, not because OpenAI's culture prioritizes it over growth.

Windows on Arm had another strong year with Snapdragon X proving the x86 monopoly is finally cracking, as Microsoft's Prism emulation now supports AVX/AVX2 instructions and native anti-cheat—forcing businesses to reckon with multi-architecture futures that reshape chip strategy and software dependencies.

India startup funding hit $11B in 2025 but deal volume dropped 39% as investors grew selective, with seed funding down 30%—showing the global flight to quality means only defensible, profitable-path businesses survive, reshaping founder expectations around unit economics versus growth narratives.

Tech companies now treat government spyware notifications as customer service, with Apple, Google, and WhatsApp directing victims to Access Now's 24/7 helpline that investigates ~1,000 suspected cases yearly—creating new liability vectors as mass state-level targeting redefines what "security" means for platforms.

Trump administration seeks to deport a hate speech researcher previously sued by X, pairing Elon Musk's legal assault on digital research with state deportation pressure—eroding the independent research ecosystem that historically checked tech platform power and studied algorithmic harms.

UK accounting body halts remote exams amid AI cheating that defeated credential systems at scale, forcing a return to in-person testing—revealing AI's real cost isn't deployment but the collapse of trust infrastructure underpinning professional licensing and certification.

Tor Project shares how authoritarian regimes systematically dismantle censorship workarounds faster than technologists can build them in Iran and Russia—a strategic inflection point where decentralized, unstoppable tech becomes geopolitically critical infrastructure for information freedom.

A Z80 microprocessor runs a 40KB conversational AI model, proving viable language models now fit on 1980s-era hardware—democratizing inference to the edge eliminates the data center monopoly and forces cloud providers to compete on orchestration and trust, not compute scarcity.

Google Photos integration coming to Samsung TVs in 2026 signals the real AI play isn't models but ecosystem lock-in—by making services indispensable across hardware categories, Google is building the infrastructure layer that future AI applications will run on.

Best of AI™ Team

Was this email forwarded to you? Sign up here.