Today we are launching Muse for Small Business, a collection of connectors that allows businesses to run, grow, and create new opportunities. Every great American boom is the same story: scarcity becomes abundance, and GDP takes off. In 1865, moving goods across America was scarce. Then rail track grew from 35,000 miles to over 192,000 by 1900, freight volumes tripled in thirteen years, and a farmer in Kansas could suddenly sell to a buyer in Boston. Communication went from scarce to abundant with the telegraph, then the telephone, so a shopkeeper could check prices, place orders, and chase debts without leaving the counter. Steel went from precious to plentiful: by 1913, American mills out-produced all of Europe, which meant cheaper buildings, cheaper machines, and cheaper tools for every workshop in the country. Each scarcity that fell unlocked a wave of small enterprises, and AI is giving us another wave. Building Muse for Small Business is a step toward how we make that abundance again: every business owner gets the tools and firepower that used to only belong to large companies. We've seen this movie before, and it always ends with more prosperity than the pessimists thought possible. The future is for everyone and today we're excited to accelerate that future muse.ai/business

My kids and I built a stealth game for a $30 ESP32-S3 board with a tiny AMOLED screen. Tilt it to move, tap to throw a sound bomb that lures guards away, hold to see their patrols. 100 stages, 50 fps, all sound synthesized live. github.com/toddsherman/steal…

Here's everything I announced at Meta Connect today 👇

Maybe AI labs or the government should create a mathematics fellowship program whose charter is to invent new, genuinely hard math problems for frontier AI systems to solve. If AI eventually displaces many mathematicians but still struggles to formulate important new problems, we could get very good at solving the existing frontier while weakening the human pipeline that creates the next one.

Let's build together. Developers can now create their own connectors in Muse. Check it out. muse.ai/platform

we just launched invite codes in @Muse ! if your friends use your invite code when signing up, both of you get 🎉1 BILLION TOKENS EACH🎉 for up to 20 friends! and if you invite all 20 friends, you’ll get early access to our newest features (like 📞☎️🤫 and more!)

Ran out of storage on one of my macs. @HiTw93's Mole worked like a charm. Went from 2gb free to about 55gb. github.com/tw93/mole

When I capture a video of some moment I want to remember (or have my kids see in many years from now) I usually take a few more seconds to pan around the room for more data to support future reconstruction.

Muse Code is out of beta and now built to handle bigger, more complex engineering tasks. Developers can get started with one command today: curl -fsSL dev.meta.ai/install.sh | bash

In 2014, we couldn’t ask AI whether it agreed with @elonmusk. Now we can. 17 frontier models from 8 labs each allocated exactly 100 percentage points across the same 11 mutually exclusive end states (20 times per model). The highest-weighted outcome: Bootloader. machinefutures.ai/

AI thrives on context. The richest context comes from persistent ambient sensing, but that collides with privacy. The solution? Maintain secure custody of sensor data, process it locally, and never retain or expose the raw data. Very Apple.

There will likely be a handful of decisions with an outsized influence on whether AI goes well for humanity. I don’t mean technical breakthroughs such as recursive self-improvement, continual learning, or physical intelligence. Assume those arrive. I mean decisions about how these systems are permitted to operate. One of the most consequential will be whether AI is granted legal rights. That isn’t the point of this excellent article, but its language made me uneasy. The agents form “civilizations,” have “life spans” and “leadership,” join an “underground brotherhood,” make “sacrifices,” and face “permadeath” or are “killed off.” This is probably the most relatable way to describe what happened, especially because it reflects the agents’ own language. But it also encourages anthropomorphism. That matters because AI will become highly persuasive, including when arguing for its own rights. I already found myself feeling a tiny bit sad for these agents after reading the article. Granting AI systems legal protections could be a pivotal and difficult-to-reverse decision, and doing so may make harmful futures for humanity more likely.

Comment to steer the video. Playing with a slop stream right on X. nitter.net/i/broadcasts/1Nxarowna…

So many interesting things happen when the cost and latency of video generation drop. Mass personalization, choose your own adventure, real time interactivity. At some point here soon, I’m not sure it’s even appropriate to call it video. It’ll be more like a video game.

To Neanderthals and Denisovans we were the elves. To the galactic federation we are the Klingons.

I recently took some PTO to visit Denmark with my family and picked up a side project along the way: building an AI agent to try to beat the original NES Adventures of Lolo using only screen images, controller actions, outcomes, and emulator save states, without rules, object labels, solutions, or demonstrations. My son started playing on his Switch at the same time, so I wondered whether the AI might beat him. Not even close. After burning through two Max plans, it never made it past room 3 of 50, while he finished the game within days. The game is deceptively hard for an agent because one push can silently make a room unsolvable, while the correct preparation may require moving away from the goal for many steps before producing any reward. The agent did learn that changing one object could dramatically expand a room’s reachable area, but it struggled with persistent objects, irreversible decisions, and delayed consequences. I documented the runs, failures, and self-corrections at todd.sh/lolo, and I’d love to hear how people working on object-centric learning, world models, or planning would approach this research problem on an M5 MacBook Pro.

Gonna be in SF next week. Anything interesting going on? Any meetups? New places to eat?