Japan still amazes me just as much as it did the first time

Well, time for the fourth one! 行きましょう！

I'm 31. Software engineer at Mango and a solo founder in my free time. I love building things, music, traveling, and studying Japanese. From Andalucía, Spain 🇪🇸 Looking to connect with more founders, builders, indie hackers, and software engineers 💻

The car insurance companies CEO

It’s almost impossible to pick just 9, but here are some that have had a huge impact on me.

There’s no better place in the world than Andalusia

@GoogleAI @googledevs

AI is going through a small revolution right now with the arrival of Jev, a new type of model created by TypeSafe AI, founded by Diogo Almeida, a former OpenAI employee. 🔸 And the interesting part? Jev is NOT an LLM. 🔸 So what makes it different? Jev is not designed to generate text. It is designed to make structured, probabilistic decisions inside software. A simple way to think about it: An LLM is someone you ask to think and answer. Jev is someone you give several options to and ask to decide. Example: customer support. Imagine a user says: “You charged me twice this month and I want my money back.” With an LLM: Message ↓ Generate an explanation about the issue ↓ Interpret / parse the answer ↓ Route it to Billing With Jev: Message ↓ Which department? Billing → 97% Support → 2% Sales → 1% ↓ Billing No generated explanation is needed. Just a structured decision. 🔸 So why is this interesting? ⚡ Much faster Jev does not generate a response token by token like an LLM. It returns a structured decision directly. For example, videogames. Instead of asking an LLM: “What should the character do?” You define the possible actions: ← → ↑ ↓ A B And the model decides which one to execute. 💸 Much cheaper for decisions at scale If your system needs to make millions of small decisions, calling GPT or Claude for every single one can become expensive. TypeSafe currently advertises pricing of around $42 per billion input tokens. 🎯 Constrained outputs by construction This is probably one of the most interesting parts. Instead of hoping the model returns exactly the JSON, action or value your software expects, the possible outputs are defined beforehand. The model chooses from the allowed options. That can mean: → more predictable outputs → easier integrations → less parsing → fewer invalid responses → smaller hallucination surface → more deterministic behaviour 🔸 Does Jev replace LLMs? No. They solve different problems. Models like Jev become interesting when you need to classify, route, choose, rank or trigger actions. Have you had a chance to try it yet? What use cases can you think of??