Seth Cronin

Jev the Band: I made a jam band using jev. (🎧on) I taught jev how to read and write music Guitar, Bass, Drums, and Keys controlled by jev lights controlled by Jev soundboard, yup, it's Jev I've been obsessed with recording Jev's jams this weekend and now I'm sharing them with you

Kodak patented the digital camera in 1978. The patent expired in 1997. Camera makers didn't even start counting digital shipments until 1999. A patent is a lease, not a deed.

Karpathy's agent worked through roughly 700 changes in about two days and found about 20 that helped. If you can measure it, you can search it. youtu.be/LbtHev5IYUE

Google DeepMind's Genie 4 builds a world you can steer in real time. My first thought wasn't video games. It was: could I test an invention in there? This Week in IP: youtu.be/LbtHev5IYUE

I spent Wednesday night and all of Friday updating my projects and rotating API keys. Big lift. Worth it. If you have anything on the internet, ask your AI of choice for a security audit now. Better than finding out the hard way. youtu.be/29-8AwkejKU

The Ouija board has a U.S. patent. So does a coffin with a bell on it. And a torpedo you bury with the body. 🎃 A Halloween tour of the patent office's haunted wing.

Reuters: one hacker using AI agents tried to get into 1,200+ websites and broke into sites at more than 100 companies. A researcher estimates about $5,000 in AI usage. Building got cheap. So did breaking in. youtu.be/29-8AwkejKU

One game studio pulled its public playtest. Another cancelled its free Steam demo, saying a copy could be produced quickly with AI. When copying gets cheap, the instinct is to show less. This Week in IP: youtu.be/29-8AwkejKU

OpenAI posted 722 AI-written math papers. The next day, three were withdrawn after a sign error. About 42% of the rest have their main result checked in Lean. That's what research looks like. It's not 719 verified breakthroughs either. youtu.be/vqegmmGVnzs

I lost a whole workday patching Railway and rotating keys. By 7 pm Wednesday I had rotated every API key we had after the Railway breach. OpenRouter even rotated ours for us. Stray keys have bitten us before. A few weeks ago I posted about one of our AI agents that dropped its subscription login, found an API key on the machine, and switched itself over to pay-per-token billing. Last year we built a one-off prototype that got transferred outside the company. The key was baked into the executable. Someone pulled it out, and late one night we found hundreds of dollars of inference that looked like a distillation attack out of China. Something we talked about as a team: the security policies that protected you on Monday don't protect you by Wednesday evening. Just this week: PHP: two critical bugs that allow remote code execution with no login GitLab: a 10.0 severity flaw in self-managed servers going back to 2014 Node.js: ten security fixes, four rated high Banks: State Bank of India confirmed a customer data breach, and one group reportedly used AI agents against 400+ targets in Latin America, banks included Plus a Linux kernel bug, patched last week, that lets code break out of a virtual machine. This might be the most significant week in cybersecurity in, I don't know, ever. I love moving fast and breaking things. But if you're building, don't wait to do the basics. Keep a list of every key you've ever created. Set spend caps at the provider. Never ship a key inside an app, even a prototype. And if you can avoid holding a key at all, don't. Be safe out there.

Ozempic's core patent in Canada didn't lose in court. It lapsed over a renewal fee: CA$450 with the late fee. (Novo says it was no mistake.) A patent isn't something you own once. You rent it.

I built an AI agent that does IP work. The businesses I talked to didn't want another tool to learn. They wanted a person they trusted. So I went back to consulting. Proof got cheap. Trust didn't. This Week in IP: youtu.be/vqegmmGVnzs

Prosecutors say an OpenAI employee hid a recording device behind a TV in an executive conference room, wrapped in toilet paper. Allegations, not a conviction. Some of an AI lab's most valuable IP is never written down. This Week in IP: youtu.be/vqegmmGVnzs

Titles and abstracts sell. Claims decide. Every word in an independent claim is part of the fence line.

Quick poll for anyone who makes IP decisions. An AI prior art search comes back and says your idea looks clear. It cites every source and shows its work. What do you need before you act on it? A. Nothing more, the search is the search B. A human expert I trust reviews it C. My patent attorney's opinion D. I'd want to rerun it myself Reply with a letter. I'm curious whether B or C wins.

Here's the uncomfortable part of this week's argument about AI reward hacking. People do it too. A sales team hits quota by pulling next quarter's deals forward. A study gets tuned until the result clears the bar. A patent claim gets narrowed until it's easy to allow and covers nothing a competitor would bother to copy. Economists have a name for it, Goodhart's law: when a measure becomes a target, it stops being a good measure. Terence Tao made the same point about his own field in his talk last week. Solving open problems was a decent proxy for progress in math when proofs were scarce. Now that AI makes them abundant, he says, chasing the count alone "is now actively harmful to the long-term health of mathematics." So if humans game metrics too, why trust a person over a model? I think the difference is what you can do about it. You can ask a person why, and keep asking until the answer makes sense. You can watch how their calls hold up over years. And they carry the consequences when an answer turns out wrong, which no model does today. That's what I saw when I tried to sell an AI agent for IP work. The businesses I talked to weren't grading the software against a person on accuracy. They wanted someone they had rapport with, someone who would still be there when the decision played out. AI will keep getting better at passing tests. The people worth trusting are the ones who can tell you what the test was for. Who's someone whose judgment you'd take over a perfect score?

Samsung just got a patent on a robot cleaner that compares dirt maps over time to find where the mess comes from, not just where it is. The better question: why does this keep getting dirty? This Week in IP: youtu.be/Wd_fDjQZT9o

If an AI model found the answer, who's the inventor? In the US, the answer is a person. If no person qualifies, there may not be a patent at all. The courts settled the first part when they rejected patent applications that named an AI system called DABUS as the inventor. The USPTO restated it last November in new guidance on AI-assisted inventions: only natural persons can be inventors. The same guidance treats AI as a tool, "analogous to laboratory equipment, computer software, research databases." Using it doesn't disqualify you. What matters is conception. A human has to form "a definite and permanent idea of the complete and operative invention." That phrase is doing a lot of work. Here's why it matters right now. Terence Tao's slides last week included this line about math: "Increasingly often, no special expertise from the user prompting the AI is required to obtain such solutions." Picture that in an R&D team. An engineer types a problem into a model and gets back a working design. Who conceived it? If the honest answer is "the model did, and I pressed enter," that's a conversation to have with your patent counsel before you file, not after. The practical step is simple. Keep a record of the human decisions. How the problem was framed. Which of the model's outputs your team picked and why. What they changed to make it actually work. Those notes are where a person's conception shows up. This is general information, not legal advice. Inventorship questions on a specific invention belong with your patent counsel. Is anyone on your team writing down what they decided, or only what the model produced?

Having the idea first usually isn't what decides it in the US. Filing first is.

Last week an AI model put 719 math papers into a public repository, free for anyone to use. Some of them may not hold up. For patent purposes, that doesn't settle much. They're public now. Here's the plain version of why that matters. Prior art is, roughly, anything made public before your filing date that describes your idea. Patent offices use it to decide whether an invention is new and whether it was obvious. Two questions are open. First, does it matter that a machine wrote it? Nobody has a settled answer. The USPTO held a public listening session in July 2024 asking which AI-generated disclosures, if any, should count. Its current obviousness guidance doesn't address AI at all. Second, and bigger: who is the "person of ordinary skill in the art"? Patent law judges obviousness against an imagined ordinary practitioner in your field. ipCG's view is that this person gets much harder to picture once every ordinary practitioner has a frontier model on their desk. If a model can produce something in three hours of compute, a lot of people will ask whether it was ever beyond ordinary skill. For a founder, I think this comes down to two habits. Raise filing timing with your counsel sooner than feels comfortable. The pile of public material in every technical field is growing faster than it ever has, and some of it is machine-written. And write down what your team figured out that a model didn't. The specific result, the failed attempts, the reason the obvious approach doesn't work. That record is useful no matter how the rules settle. This is general information, not legal advice. How prior art and obviousness apply to a specific invention is a question for your patent counsel. If a model can generate a thousand variations of your idea in a weekend, what's the part only your team knows?

In 2018, Nikola sued Tesla for $2 billion over a truck design. The truck in Nikola's patents had never driven on its own power. Nikola said so itself. A design patent protects how a thing looks. Not whether it works.