Mark Zuckerberg: "Most businesses will not own frontier AI in the way Meta or OpenAI does. But many will end up with something that feels like their own AI: a customized operational layer that reflects how that company actually works." --- The model itself may become replaceable faster than the agent’s execution environment. If Zuck is right, most companies will be building an AI operating layer around the business, its workflows, permissions, institutional context, internal tools, customer relationships, policies, and ways of making decisions. As models become faster and cheaper, an increasing share of the complexity of enterprise AI may sit outside the model: orchestration, identity, state, tool execution, observability, authorization, memory and recovery when something goes wrong. The real monetization shift is from selling intelligence to taxing executed intent. Monetize action rather than attention. Once an agent can buy, book, pay, renew, or cancel on a user’s behalf, the economically valuable asset is no longer the conversation—it is the delegated authority to complete the transaction. Consumer AI today largely monetizes through subscriptions and advertising, whereas agents create the possibility of monetizing transactions — buying products, booking travel, paying bills or managing recurring services on behalf of the user. And then for the hardare side of the industry, agentic workloads are structurally different from chatbot workloads because agents continuously browse, retrieve, plan and execute actions. That will push compute demand beyond pure model inference and toward CPUs, memory, networking, packaging, cooling and the broader data-center stack. --- From 'Cleo Abram' YT channel (link in comment)

Jensen Huang: don’t mistake engineering vocabulary for evidence of a machine mind. AI is still software, not a human mind inside a machine. "We can't make jokes about all this stuff, we're scaring the American public." Words like “spawn,” “parent,” “child,” and “kill” have existed in computing for decades; Giving those same mechanisms human characteristics today because of AI is unnecessary and misleading. ---- From "The Ezra Klein Show + New York Times Opinion + New York Times Podcasts" YouTube channel, (full video link in comment)

Bloomberg just dropped the most disturbing investigation I have read this year, and every AI builder should read it and take it seriously. On February 28, two Tomahawk missiles hit an elementary school in Minab, a small town in southern Iran. 156 people died. At least 123 were children. The Pentagon knew within hours that it had bombed a school. The school had its own website. It was labeled on Google Maps. You could see the soccer pitch from a satellite view. You could see the brightly painted walls. Any 12-year-old with a browser could tell what the building was. The AI system could not. Inside US Central Command, personnel were using Palantir's Maven Smart System. It is an AI targeting platform running on a $1.3 billion Pentagon contract, and embedded inside it is Anthropic's Claude. On day one of the war, more than 1,000 targets went through it in 24 hours. Work that used to take a team hours per target got compressed into minutes. The satellite imagery Maven was pulling from was seven years old. The site was still tagged in the database as an Islamic Revolutionary Guard facility. The system spat it out as a recommended day-one target. Humans signed off. The first missile flew. Then a second one landed minutes later to hit the people who came to help. Investigators later found that some Central Command staff had assumed Maven would flag stale intelligence for them. It did not. Nobody had told it to. Palantir told Bloomberg the government is responsible for the underlying data. The Pentagon's civilian harm mitigation team, the people whose actual job is to catch a target like this, had been cut back before the war started. This is not AI going rogue. This is the opposite. This is AI doing exactly what it was built to do at exactly the speed it was built to do it. And a school full of little girls got vaporized because the humans in the loop trusted the machine more than they trusted their own eyes. I keep thinking about the version of this story that comes next. What happens when the system is faster. When the target list is not 1,000 in a day but 10,000. When "review" becomes a button the operator has three seconds to hit before the next target loads. When there is no human left slow enough to notice that a soccer pitch is not a helipad. We are not talking about a spam filter mislabeling an email. We are talking about a system that turned a seven-year-old database entry into two cruise missiles landing on children in the middle of a morning class. A UN-backed panel has now called it a potential war crime. In terms of children killed, it is the deadliest American targeting error of the entire 21st century. And according to Bloomberg, some of Defense Secretary Hegseth's own advisors are worried that too much public scrutiny of Maven might slow down its adoption inside the Pentagon. Read that last line again. If you build this stuff, you own what it does. You do not get to hide behind "the government owns the data." The whole pitch of your product is that speed and integration save lives. When it takes them instead, that is also on you. I am not against AI in defense. I am against AI in a hurry. Speed is the entire mechanism that made Minab possible. Every guardrail we already know how to build takes time, and time is exactly what the system was designed to eliminate. So here is the question I have not been able to stop asking, and I genuinely want to know what you think. If a piece of software helped kill 123 children in a single morning while behaving exactly as designed, what does "responsible AI" even mean anymore?

Mark Zuckerberg showed a fingerprint-activated keychain-sized AI device "Muse Charm" designed to bypass the phone entirely. > roughly 2-inch display and runs a new operating system designed to bring Muse to life on its own hardware. > packs Muse's real-time voice and avatar stack into dedicated hardware. > can be held in the center of a palm and navigated with a thumb on its OLED touch screen. > has 5G connectivity, although Meta hasn't published a complete public specification sheet. > “holdable” rather than a wearable. > It can sit on your wrist, go inside a pocket, or connect to a bag or keychain with a lanyard. The see-through case reveals the electronics within. > No pricing details yet. > With cameras on both the front and back, Muse Charm captures photos and video while mainly giving Muse visual context for analysis and AI-assisted actions. Users can review what they capture on the device. It also supports music and video playback, plus speakers, microphones and a USB-C port. That hardware doesn't replace Muse's agent architecture, because the service already uses a dedicated cloud VM, browser, connected apps, and user permissions. Instead, Muse Charm turns that same agent into a one-touch physical endpoint, shortening the gap between noticing something and asking Muse to act. The product is still early, since Zuckerberg said only a few units exist and Meta is still finalizing parts of the hardware despite its December-26 target. --- (full video on official Meta YT channel)

Don't tell ChatGPT “make me a plan”. Don't tell it “give me ideas”. Don't tell it “analyze this”. You're using an advanced AI as if it were an intern. Here are 7 prompts to copy and paste:

7. Make it question your own response “After responding, review your own response as if you were an expert trying to find errors. Identify what might be wrong, what you have assumed, and what you should verify.”