nemath

If we have really powerful models (AGI) - would it still make sense to have subagents that have access to specific tools and has specific context?

Wouldn't the overlord model just do everything?

Hey there! Is there a third party service which routes my queries to best possible model? I find myself doing tasks like - creative writing, systems research, image generation etc. I currently just use ChatGPT pro - is there any service that based on my task routes it to best model for that task?

I spoke to a friend of a company that you all have probably heard of 10b+ val. They currently have a problem where the engineers are not able to run more than 5 instances of their server locally. They want to run more instances (10) so that AI agents can work on these and everyone is 10x more productive.

My answer to this was to remotely host instances and let cursor work on it. He also said the another team might (?) be working to build something to help internally. Why wouldn't they just use a company like fly.io or modal for this? Or just give employees GCP VMs that have more compute and they just ssh into it?

What are your top picks of some great launch videos recently which aren't just cool cinematically but also bring across what they are trying to solve more clearly?

I’ve been thinking about a parallel between cursor like IDEs and what health data unification might look like.

Today, my health “signal” is fragmented across: 1. Clinical records (EHR, labs, diagnoses, meds, doctor notes) 2. Wearables (sleep, HR, HRV, activity — high-frequency streams) 3. Life context (work schedule, stress proxies, travel, finances, routines)

What I want is something like an IDE for my body: all these sources show up as files, and when I ask a question, an agent retrieves the minimal relevant pieces and responds with citations + uncertainty.

“My resting HR is up this week — is it sleep, travel, training load, illness, or medication timing?” “After I started X medication, did my sleep/HRV trend change?”

This could also be proactive - obviously there are hard issues I'm sure - permissions, privacy, clinical safety, etc.

Curious for raw/critical feedback: Where does this analogy break, and what’s the hardest missing piece to make this real?