Welcome, humans.

Sam Lessin of the More or Less pod (and a VC) says we need a new role akin to a SysAdmin, but for bots (call it a BotAdmin?). It’ll be for the person at the company whose job it is to wrangle, well, all the bots. The More or Less gang shouted out a few alt titles: Chief Agent Officer, Master of Claws, ClawMaster, etc.

Here’s a few more: Claw Wrangler, Clawbbler, Clawmmander in Chief. Hey, I like that last one… Anyway, it’s a great episode on OpenClaw, bot anxiety, vibe-coding on planes, and so much more.

Speaking of Clawwd: do three of you want a free week of Anthropic’s Cowork to try out? If so, click this link! We only have three (graciously provided by Anthropic) credits though, so you gotta click fast.

Here’s what happened in AI today:

📰 Nvidia's robotics chief said AI agents will bring a "ChatGPT moment" to robotics
📰 Apple quietly collected ~$900M in App Store fees from generative AI apps last year
🍪 Kling 3.0 took the #1 spot on the AI video leaderboard
🌟 Andrej Karpathy, Terence Tao, and Zvi Mowshowitz in three fantastic interviews.

… and a whole lot more that you can read about here.

P.S: Want to reach 680,000+ AI-hungry readers? Click here to advertise with us.

😼 The US Government’s new AI framework, explained

DEEP DIVE: The White House Just Told Congress How to Regulate AI. Here's What It Actually Says.

The U.S. White House dropped its first-ever national AI policy framework today, and the headline takeaway is clear: the federal government wants to be the only one writing AI rules.

Here's what happened:

The framework covers seven areas: child safety, community protections, copyright, free speech, innovation, workforce training, and federal preemption (overriding state laws).
It explicitly says states should not be allowed to regulate AI development, calling it "an inherently interstate phenomenon."
On copyright, the administration declared it believes training AI on copyrighted material is legal, but wants courts (not Congress) to settle the debate.
It calls for "regulatory sandboxes" (test environments where companies can experiment under relaxed rules) and no new federal rulemaking body.
OSTP Director Michael Kratsios told reporters the goal is to pass legislation this year (before end 2026).

Why this matters: The framework sounds comprehensive, but its most powerful move is what it prevents. By calling for federal preemption of state AI laws, the White House is trying to shut down a wave of state-level regulation that's been building since Congress failed to act, oh idk, over and over again for the past 3 years since ChatGPT came out? More than 50 Republican state legislators pushed back against this approach just weeks ago, calling the administration's pressure campaign an effort to shield Big Tech from accountability.

Meanwhile, Sen. Marsha Blackburn released her own nearly 300-page federal AI bill the day before, with much stricter provisions: a "duty of care" requirement for chatbot developers, a sunset of Section 230 protections, and criminal penalties for AI companies that let chatbots have explicit conversations with kids. That said, the Cato Institute already identified five major flaws in her approach.

Our take: The White House framework is a wish list, not a law. Congress has been deadlocked on AI regulation for years (because of lots of lots of lobbying, among other actually good reason like not stifling innovation so the U.S. can lead in AI development, which matters if you live in the U.S but might worry you if you live, well, anywhere else), and the same fights over preemption, copyright, and kids' safety that stalled past bills are still very much alive.

Watch the gap between the framework's ambitions and what actually makes it through committee... we imagine an even more watered down version will be what actually gets through.

FROM OUR PARTNERS

Solving the AI Amnesia Problem

Most AI agents lack a persistent state, losing critical context once the session ends. While context windows are expanding, they remain a volatile and expensive substitute for true memory. Building authentic Agentic AI digital partners is essential for competitive advantage and requires a dedicated memory infrastructure layer in a ‘convergent database’.

This technical deep dive explores the architecture of agentic memory and provides the code snippets using LangChain and GitHub notebook to help you get started. Build a multi-tier memory manager that handles entity extraction, conversational history, and automated summarization.

Explore the full blog here.

🎓 AI Skill of the Day: How to pick the right "thinking mode" for your AI tasks (and save time doing it)

Most people use the same AI settings for everything. Newer AI models now let you toggle between fast mode and deep thinking mode. Mistral Small 4 just launched this as "configurable reasoning effort," and Claude, Gemini, and others offer similar controls. The idea: match the AI's effort to your task.

When to use fast mode: Quick rewrites, formatting, simple lookups, brainstorming lists, summarizing short text.
When to use deep thinking: Multi-step analysis, debugging code, comparing options with tradeoffs, anything where getting it wrong costs you time.

Try this prompt to see the difference:

Analyze this quarterly report and identify the three most significant trends that could affect our team's priorities next quarter. For each trend, explain the evidence, the risk level, and one specific action we should consider. Think step by step.

Okay, if you’re not a business exec, you probably want to change the above prompt to this instead (brackets indicate where to put your own context):

“Analyze this… [and then copy + paste or attach to dump in whatever you want to analyze]. For each [aspect of the document you want to specifically analyze], explain [list three things you want it to pay attention to].”

The phrase "think step by step" is your free upgrade to deeper reasoning in almost every model. Use it when accuracy matters more than speed (but note: you don’t need to use it if you click on “extended thinking” or similar functionality).

Want more tips like this? Check out our AI Skill of the Day Digest for this month.

Have a specific skill you want to learn? Request it here.

New episodes air every week on: Spotify | Apple Podcasts | YouTube

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 680K+ readers here!

*Still juggling 10 different tools? Learn AI workflows and replace 80% of them. This is where AI is headed: one place to get everything done. Smarter, faster, and built to make you 10x more productive. START LEARNING NOW!
Dreamer helps you discover, build, and remix personal AI agents that handle everyday tasks like managing your calendar, organizing groceries, and tracking your finances, all working together in one place; join waitlist here.
1. Latent Space talked to Founder David Singleton, former CTO of Stripe about the project and they have a builder funnel for technical people who want to build on the platform.
AirOps helps you create and refresh website content that ranks in both Google and AI search, so your brand gets cited when someone asks ChatGPT or Perplexity a question.
Evidenza runs market research on AI-generated copies of your customers (even hard-to-reach buyers like CIOs), turning a 6-month survey into a 6-hour one with 88% accuracy vs. traditional methods.
gstack is Startup Accelerator Y Combinator President Garry Tan’s exact Claude Code setup (which he open-sourced): 15 slash-command skills serving as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA with persistent headless Chromium, safety guardrails (/careful, /freeze, /guard), parallel Conductor sessions (10-15 sprints at once), and Greptile integration
This is wild: NousResearch Hermes Agent autonomously wrote, typeset, edited, and published a full-length 79k-word novel "The Second Son of the House of Bells" using an autoresearch-style modify-evaluate-keep/discard loop for fiction, world-building, adversarial editing, and cover art (GitHub).

📰 Around the Horn

Dear Senate, please don’t ban SeeDance 2.0, these videos are too amazing; I know this is propaganda, but 1. it’s hilarious and 2. I can’t actually believe they rolled this out on state media. wild.

Nvidia's robotics VP said AI agents will bring a "ChatGPT moment" to robotics, with a single agent coordinating an entire fleet of robots by dividing goals into tasks and assigning them individually.
Cursor's new Composer 2 model was revealed to be built on top of Moonshot's Kimi K2.5 without the attribution required by its license (which mandates a visible "Kimi K2" label for products over $20M in monthly revenue or 100M monthly users).
Apple apparently collected nearly $900M in App Store fees from generative AI apps last year, with roughly 75% of that coming from ChatGPT alone.
Zhanghao Wu (SkyPilot) ran Karpathy's autoresearch on 16 GPUs in parallel: 910 experiments in 8 hours (9x faster to same best result), with the agent autonomously discovering H200s were faster and routing work there without being told.

…And here’s everything else that happened so far this week.

FROM OUR PARTNERS

Live with Marc Benioff: Slackbot is the new interface for work

Join Marc Benioff live on March 31 to see how Slackbot orchestrates work across your data, apps, agents, and workflows through a single conversation. Learn how teams are automating work and moving faster with AI, hear from real customers, and get your questions answered live.

🏆 Sunday Special: Three Great Interviews

Okay one more; this is just pure genius

Three Fantastic interviews to watch (or add to your “Watch Later” queue!):

Andrej Karpathy on No Priors where Sarah Guo chats him up about agents, autoresearch, and what’s next in AI; P.S: Zara Zhang made it a blog!
Terence Tao on Dwarkesh about how AI is still in its brute force era, but it will eventually revolutionize everything (including experimental math).
Zvi Mowshowitz on The Cognitive Revolution on AI’s new recursive self improvement dynamic, the end of the beginning of the AI development story (we’re now in “the middle game”), and much more.