Theory: LLMs that lean towards test-driven development (TDD) are major slop factories. Astra, I’m looking at you 👀 It goes like this: 1. Say a new feature F requires a minor refactor to land cleanly 2. The model first writes a failing test against the current code to validate F doesn’t exist yet 3. The test should flip green after F is implemented correctly 4. But the problem is this: the model just created more code that depends on the old shape, so its incentive to keep it became stronger 5. It marches on and implements F on top of the old interfaces, bending over backwards in ways that a human engineer would not And that is, my friends, one out of 1000 ways that slop is summoned into planet Earth Model = 💪😏 // You = 😭🤬

Theo Von: “It feels to me like we’re in this kind of privatized communism right now that people don’t want to admit to. You want to believe in capitalism because that feels like what’s fair, but now it feels like so much is owned by just a 1%. At a certain point, if capitalism is not working for everybody and 1% own everything and have just monopolized everything and squeezed it into themselves, if there’s a level of greed that’s that big, then you have to do something to take care of people”

Theo Von: “Do you think the American Dream is still alive?” Zohran Mamdani: “It feels like you have to luck out in some big way. In many ways I feel like it’s why a lot of people play the lottery. For many it feels like their only chance. People think they have better odds playing the lottery than making it in their own life. What does that say about the country we live in?”

.@theojaffee argues the biggest automation in history created the modern world. @AndrewYang says AI is different, and 1894 is the reason why: Theo: "Automating subsistence farming, which almost every human for all of history used to do, has created a tremendous amount of new industries." "This whole thing of Jubilee videos would just not exist if we didn't have a huge amount of the tasks that we previously did automated." Andrew Yang: "Only 1% of Americans now work in agriculture and that number used to be 40 times higher. So that's a fact. AI is different than anything that has gone before. We've never had thinking machines that can do billable lawyer work in seconds." "Even in the industrial revolution, there were mass riots that killed dozens of people. Today you have many more educated people who are not going to take getting kicked to the curb gently." @jubileemedia

Theorem is heading to @token2049 Singapore 🇸🇬 We’ll be there 5-9 October at Marina Bay Sands, talking tokenized assets, secondary markets, compliant liquidity, and the infrastructure needed to bring RWAs onchain. See you in Singapore.

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

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