Erik Meijer

Typical VC bullshit. This is the case for any (remotely) related AI startup.

"something harder" sounds like pun intended.

Way back when pitching Leibniz Labs it was based on two premisses (1) verifiably safe, and (2) hyper-personalized. I got laughed out of the room consistently (or in VC speak" ...blah .. no conviction .. blah ...). In the mean time no one is laughing about safety and verification any more. The models are not yet powerful enough, but soon I expect hyper-personalization via partial evaluation will be common practice as well. If you want to learn more, come to my "Welcome to the AI chop shop" talk at aidevcon.io. Use "ERIK30" to save some $$$ to spend on tokens instead.

What happened to the continuous learning hype?

Congrats!

> I hear you've now nerfed your models so that when people try to use them to clean the spamware and spyware out of their televisions, you block that. Most definitely they have locked down on reverse engineering.

Since when is mechanical sympathy controversial?

If you like experimenting with fermentation, try Kefir. Just get three GLS 80 wide mouth 1L laboratory bottles (one to drink, one to ferment, one to wash) and some dried culture from amazon. Super easy.

At least no pineapple ;-)

Try to ask Claude to prove [some property you are interested in] using TLA+, and you will be amazing how thorough a job it does to write a model, connect it to your code, explain what it means, what the assumptions are, ... Rinse and repeat until you are happy. I am 100% convinced you cannot iterate that fast and that comprehensive with human developers.

Biggest lesson for me has been the fewer ingredients the better. It is very tempting to put loads on stuff on them.

I learned *a lot* from watching Maestro Vito lacopelli!

As AI has surpassed all our skills, and the big labs grabbed all opportunities with their infinite token budgets, I have been learning the craft of pizza making. Before the food robots rob us of that as well.

As a non-native speaker, they are isomorphic to me ;-)

"... First I'll check what the vendored TLAPS supports and how the existing proofs are laid out: ..." Claude This always freaks me out. It fucking wrote the model in TLA+ 5 minutes ago!

Yeah, before AI, all code was perfect, never slop. Not!

Been giving advice for free since early 2023.

From my AI World Fair talk in June 2026 on using formal verification for AI safety. Apparently they did listen, but never bothered to tell me (yet) ;-)