Real world interactions will be a big thing. Sensors and what not. Think smart hone but not big tech smart, private smart. I've already explored how it could help me protect my boat and help me manage onboard tools. There's a lot of scope but it's going to take a while for the really interesting stuff to start happening, I reckon. This is early fun and experimental.

Good morning 🌄

Pura vida 🫂🤙🏻💜

What model were you using? Not all models are the same.

The future. Good morning.

As someone that already paid for and used AI every day for work, I just see this as an extension of that. Perhaps some people have more use cases than others for now.

With self-hosted models...

Do we have a session this Monday?

I only use Opus. Sonnet and Codex probably have successes too. Everything else will give unpredictable results.

Yeah or @Maple. That needs to be said

You can run local models.

Gm

Really good point! 🤙

It isn't magic. You plug in some LLM and usually the same Big Tech LLMs to do the AI bits. That it communicates with all those insecure platforms and including your credentials and keys for various services you want to automate is a source of many possible privacy and security problems. Some of those platform allow others to inject prompts against your stuff exfiltrating your stuff and credentials or having it do things they want as if they were you. It is a very interesting tool but it is not "your AI" cleanly working "only for you".

do you use opus directly or through @Maple ? Ideally we would also run a local model but maple ai seems to be a good compromise, at least better than to link all data to your anthropic account. I also run anthropic directly but I think I will try opus through maple ai max subs for next month. Was also looking into local models but Kimi 2.5 seems to be a ressource hog, at least nothing you can run on < 10000 $ hardware afaics. Maybe GLM 4.7 but that seems to be not near opus and kimi level. Tried sonnet-4.5 for a few days but results were not that great, not worth the wasted back and forth. But I hope as these things get more efficient and smarter we will be able to run a decent model locally on consumer hardware.

Meh. I already did that last year while toying around, and it actually got old / distracting after a while, and now I don’t use it at all. I wonder if others will go through the same process.

Maple is the way forward but needs tighter integration. They are also working on Sage but its early days

If only we could plugin private LLMs and have stable agents