interesting detail in this DeepMind paper: in their setup the model never has to get bigger as tasks get harder. only the instructions get longer. pairs well with GEPA. makes me wonder how small a model can be if it's just good at following instructions
it so easy to get hooked to the codex app. Currently using a proxy to access all models here. I don't see myself ever return to the terminal. Honestly, I don't know why people still use the terminal
Saw this first hand at https://t.co/8PC4lz5W71, speed is the most important variable at any early stage company.
Do things that don’t scale and do them fast. You will see the scale come to you.
spoke to a real estate agent today and he was using basic chatgpt, then manually doing everything else
made me realize how much of the world still thinks AI is just chatting with chatgpt
we’re so unbelievably early. people have no idea what’s already possible, let alone what’s coming out every day
overengineering before speaking to users was always a mistake.
you can direct a codex with your voice and finish a usable prototype before your hike is over. then find the exact people with the problem, book calls and test messaging before production code exists.
but the real unlock isn’t building faster. it’s learning faster. you can learn so much fucking faster now with all these tools at our disposal.
every conversation sharpens your understanding of the problem. a month of focused calls can teach you more than a year of building in isolation.
biggest lesson from this summer: build enough to start the conversation. then let the market tell you what deserves a real build.
GLP-3s are genuinely going to make the world healthier once they’re approved. This stuff is insane.
I’ve seen startups in SF where GLP-3 use is basically part of the culture, and everyone is getting into the best shape of their lives while building the next generation of companies.
Absolutely insane to watch.
The next great AI application, driven by open source, or open weights, will be a job simulator.
How employees gain experience in a future AI world is going to be far different from today. Employees won’t have as many touch points in the company to gain knowledge and experience from. That’s where judgement has historically come from.
Much like race car drivers and pilots have software that is continually updated to enable them to experience as many scenarios as possible, trying to replicate what could happen IRL, smart companies will have their employees and stakeholders with the most domain knowledge create the simulator that takes them through every possible situation they could face and helps prepare them.
Onboarding will have a completely different meaning
Smart companies will realize that Father Time is not only undefeated in sports , but business as well. If you are not capturing what is your business is all about from ALL of your employees , you will be challenged.