.@wordgrammer on why math isn't a scarce resource, and why art might be: "Math is a very powerful cultural institution. It is sacred. We care about it, not necessarily because it leads to useful outcomes, but because it is beautiful." "You stop viewing math as a competition, and you start viewing math as an unlimited resource towards which humans and AI can work together forever. Math is inexhaustible. There are an infinite number of true statements about the world." "Competing over math as if it was a scarce resource, I think is a little bit of a silly thing to do. Art, I think, is maybe not so infinite, 'cause there's a finite amount of attention that humans have."

.@wordgrammer on how “AI art is theft” became a lasting reputational problem for AI, and why the same mistake is now repeating with mathematicians: "I compare it to a leaky pipe. A lot of people in the AI industry completely forgot about the, quote, steal from artists issue, and I think that artists never quite got over that." "The damage to the reputation of the AI industry has in very large part been driven by artists who that initial starting point rubbed the wrong way. They've used all their cultural influence and capital to turn people against AI." "Something very similar is happening right now with mathematicians. It would be easy to fix it right now and address the issues of mathematicians." "If it goes untouched, over the course of a year, maybe two years, maybe three years, the public's perception of AI will be just so bad that it might be impossible to break."

.@wordgrammer says AI art should steal Adobe’s playbook: stop competing on infinite content and make Hollywood-level creation accessible to everyone. "Adobe products took Hollywood technology and they just made it accessible to the world at large. This seemed to be a really, really good business model." "The majority of the techniques used by influencers, by video editors, by marketers over the past 20 years have been cropping videos, editing B-roll. Significantly less time has been spent towards just taking a single 30-second long clip and making them fantastic." "Until every single video in the world looks like Hollywood, there's no sense competing over art as if it's a scarce resource."

SITUATION DETECTED: Alibaba plans to train a new AI model with between 5 trillion and 10 trillion parameters, CEO Eddie Wu said Tuesday at Alibaba Cloud’s annual Apsara Conference in Hangzhou, per Reuters.

Rocket Lab CEO @Peter_J_Beck reveals the “ridiculous” 24-hour constraint he gave the Neutron team, and how it forced them to engineer servicing out of the rocket: "At the time, it seemed a ridiculous constraint, but it drove all the right engineering thinking. Time will tell if we ever actually get there, but certainly the vehicle is designed to do that." "It's not 24 hours of inspections. It's literally 24 hours, load payload, and go again. Just like you would with a normal aircraft." "If you had 24 days, then you have plenty of time for inspections and replacement and servicing of components. If you have 24 hours, it engineers out all of these serviceable items." "We can't go and service the rocket engine between flights. We just have to make it go and go and go. You don't pull a turbine out the front of an airplane engine every time you land." @RocketLab

Rocket Lab CEO @Peter_J_Beck does the back-of-the-napkin math on Iridium’s ~$3B constellation and says Rocket Lab could have built it for around $350M: "Take Iridium for example. Their constellation, don't quote me on this, is around about $3 billion for them to develop and deploy. We did a back of the napkin calculation and we reckon we could have done that for around about 350 million." "The timeframe to develop the satellite and put it on orbit, typically that would take five years plus. Roll some stuff off our production line, 12 months or 24 months." "When the cost is an order of magnitude less, you can take some risk, because nobody's gonna come up with a business plan, spend $3 billion and wonder if it's gonna work." "You can put experimental business plans in orbit that if it doesn't work out, you've lost a relatively small amount of money. Your ability to innovate drastically increases." @RocketLab

Rocket Lab CEO @Peter_J_Beck says reusable rockets aren’t new, the Space Shuttle proved that in the ’70s, but the economics never worked until recently: "I think people forget the space shuttle. The space shuttle was a reusable launch vehicle, designed back in the '70s. It's just unfortunate that the reusability didn't deliver on its cost targets. It was more expensive to almost reuse it." "What's unique now is it has been demonstrated, and the economics have been demonstrated." "You have a giant rocket full of fuel, and it puts all of that energy to go up. Newton's third law, you're gonna have the exact same energy coming back down, so you have to dissipate all of that energy. It's a re-entry and control problem." @RocketLab

Rocket Lab CEO @Peter_J_Beck breaks the entire space industry into 3 simple layers, from rockets to the services we use every day: "I probably shouldn't have called it Rocket Lab. I should have called it Space Lab. The rocket always steals the show because it's this big roaring stick in the sky." "Rockets, that's about a $10 or $20 billion industry. Spacecraft, the tools in space to provide services, 20 to $30 billion. And then the reason why you do all of those things are applications in space." "Space is kinda weird that it's hidden infrastructure. You don't drive past it every day, so you don't see it. But everybody interfaces with space every single day, whether you realize it or not." "We build rockets, we build satellites, and with our acquisition of Iridium, we're also a services company from space." @RocketLab

Rocket Lab CEO @Peter_J_Beck explains what Iridium’s L-band can do that Starlink broadband can’t, including provide a GPS-like signal 1,000X louder when GPS is jammed: "Think about L-band spectrum as like a really loud voice. It can shout through the rain and it can shout through buildings. Broadband is a quieter voice, so if it's raining, you start to lose your signal." "That means it becomes really important for defense and safety critical things, where the signal just must get through." "If you look at any of the trouble zones right now, GPS is basically almost useless because it's completely jammed and spoofed. L-band can also be used for providing a GPS position, but it's a thousand times louder." "If you're sitting on the ground and you're trying to jam GPS, it has to be a thousand times more powerful than it currently is, and that's a challenging thing to do." @RocketLab

Rocket Lab CEO @Peter_J_Beck reveals the US government gave them 24 hours to launch and rendezvous with another satellite, and they made orbit in just 16 hours 42 minutes: "We're a prime contractor on a couple of these really large national security missions. It's really exciting to see the government procure these kind of missions in a much more commercial way." "The government didn't procure a rocket, they didn't procure a satellite, they procured an image of another satellite in orbit." "They gave us a call up, and we had 24 hours to integrate our payload, launch it, and rendezvous with the spacecraft in orbit. Within 16 hours and 42 minutes, we were already in orbit." "This is the new world, where commercial companies can truly respond at what you would have thought of as a nation or a defense kind of cadence." @RocketLab

Theorem co-founder @rajashree breaks formal verification into 3 problems and says AI already solved the one everyone thought was hard: "The core challenge is taking these programs and shoving them into a proof assistant, and then being able to phrase these questions. That's the theorem statement generation problem." "The second part is generating all the proofs. That one, the AIs solve fully, because they can reason about these programs perfectly." "The third part is checking this proof, which is where you need to spend your CPU cycles, not GPUs. The proof assistant is asymptotically too slow, so even though you wrote the proof, the checking time is so long that you can't actually get the answer." @theoremlabs

Theorem co-founder @diagram_chaser reveals the one-line change that took verifying real-world HTTPS code from 4,000 millennia to seconds: "The project that I worked on in my PhD was verifying the code that runs HTTPS in browsers like Chrome and Firefox." "You plot this beautiful graph that is an exponential in the number of bits in the prime, where you're like, it takes a couple seconds on my tiny toy examples, and on real-world examples, it would take over 4,000 millennia." "The thing that it's actually checking is that you wrote the same thing in two different ways. This should be very fast. It's doing the operations in the wrong order." "So you tweak one line, and then it drops down to a couple seconds." @theoremlabs

Theorem co-founders @rajashree + @diagram_chaser explain why formal verification could break cybersecurity’s endless whack-a-mole and stop reward hacking during AI training: Rajashree Agrawal: "Formal verification is asymmetric security. Currently you have this whack-a-mole problem. As the attackers get better, the defenders have to catch up. You keep doing this forever." "Once you prove this particular property holds of your program, you don't need to check that again. This feels like the asymmetric approach that you need if you're going to avoid this cyber apocalypse." "If you wanna move beyond one-off interactions, build a very complex world, lots of automations, you're going to have structured things coming out of models which looks like software. Being able to reason about software seems like a good hammer to have." Jason Gross: "If you can verify the RL environments and the graders that you're running, then the models won't have any reward hacks during training, and so you can potentially train them to not be as reward hacky in general." @theoremlabs

Across Muse, Instinct and Grok Bot, personal agents are going mainstream. We looked at how this might affect CPU demand. Here’s how it works under the hood:

Theorem co-founders @rajashree + @diagram_chaser on how you actually prove an AI agent can’t escape its sandbox, and what happens when the proof exposes a way out: Jason Gross: "Anything that the agent does inside the sandbox will not result in some canary file outside the sandbox getting changed." Rajashree Agrawal: "We intend to have one by the end of the year. We've got a verified sandbox now, and the goal is to keep adding features in collaboration with the AI labs." Jason Gross: "If the AI can guess the secret root key of the package server, then it can do anything. Either I try to prove that the AI is not going to be able to guess that, or I design the system so that the channel just doesn't allow it to authenticate that way." "You also want to prove that on most inputs there's no change in behavior. Because otherwise it could make it inescapable by saying, well, sandbox just shuts down as soon as it starts." @theoremlabs

Theorem co-founders @rajashree + @diagram_chaser explain why verified AI sandboxes are still months away despite the latest breakthroughs in automated theorem proving: Rajashree Agrawal: "The models just got good enough to prove these theorem statements, or they're still getting there. One of the costs is just tokens." Jason Gross: "Taking the recent Navier-Stokes news as a baseline, the models wrote something like 600,000 lines of Lean in about 17 hours. This works out to between one kilobyte and 30 kilobytes verified per hour." "The smallest version of Linux with a sandbox that we can make is about five megabytes. So that's still a handful of months away, even at the rate of the Navier-Stokes auto-formalization." "We need to build a pipeline from verifying the software back into RL-ing the models, so that if you want to verify all production software that exists in the world, it costs you less than $10 trillion." @theoremlabs

Smartling VP Olga Beregovaya says the reason frontier AI is still Anglo-centric is surprisingly simple: translation isn’t a priority. "I may make a controversial statement. If you look at frontier models, usually translation is not the main objective. If a model is to perform 1,000 tasks, translation is just one of 1,000." "Since translation is an afterthought, I would say long tail under-resourced languages translation is a layer below the afterthought." "With each next release, discoveries are made through evals that the languages underperform, and for most of the models, long-tail language performance starts evolving." "The more models translate, the more people read, the more people react, so the training data is just increasing. I'm gonna write on social media about it, and hey, what did I just do? I just provided additional customization data. It's a compounding effect." @smartling

SITUATION EXPLAINED: Is Jev a new kind of AI, or just a very fast classifier? • Diogo Almeida claims Jev is up to 200x faster and 400x cheaper than LLMs • Jev doesn't write anything. It returns scores and probabilities, fast and cheap • @tenobrus: "every flashy demo is something existing done 10x faster and cheaper and not actually functional, and the launch was incredibly misleading" @theojaffee: "It's very, very early days. It's like the equivalent of GPT-2 to what we have now with Astra. And it's possible that there are some tasks for which something really cheap and really fast is better than frontier intelligence."

SITUATION EXPLAINED: Grok 4.7 beats GPT-6 Astra on real-world tasks, and it's 5x cheaper. • Same price as 4.6, $2 input and $6 output per million tokens, well under Sol, Fable, and Astra • A new larger base model, which Elon has put at 2.1 trillion parameters, up from 1.5 trillion, with extra training on SpaceX data • Trained to natively understand the Grok Bot harness, the same move Anthropic made with Claude Code • Leads on electrical engineering and Harvey's legal benchmark, beats Astra Max on GDPval, trails on software engineering • @elonmusk: "Grok 4.7 places SpaceXAI as third after Anthropic and OpenAI for agentic coding. When factoring in that Grok is significantly faster and lower cost, it's a great choice for your everyday workhorse" @theojaffee: "SpaceX has a huge amount of data on real world hardware problems, real world engineering. So I bet Grok models are going to be better at rocketry engineering than any of the other models out there."