pull down to refresh
I want to invest in (1) too. I think in the short run it's still faster and better for me to do things hands on, maybe AI assisted, but with high interactivity
But I think in the long run, it will be a good investment to learn how to manage a fleet of assistants with a lower degree of needing to be in the loop
IME, where we are now, you mostly need to
- start the loop
- review the loop once it exits
You can probably judge a system's awesomeness by how infrequently you have mistakes at (2).
Next leap is probably getting rid of (1) - the loop is always running, it knows what you want from your history, from your grunts, from what it read about you online, from reading all the books you haven't read, and you say yes or no at most.
Next leap is probably getting rid of (1) - the loop is always running, it knows what you want from your history, from your grunts, from what it read about you online, from reading all the books you haven't read, and you say yes or no at most.
Would you really want that? I'd be throwing the clanker out the window after the second time I have to decline lol.
Right so the loop gets sensor input / reads over your shoulder / looks at what you're doing and then reacts to that? Maybe we should post-train GLM-5 with Elon's system prompt. I'd be really careful to give this kind of surveillance capability to any third party in any way tho.
you build a factory that builds factories.
Yes! Though why would I want some bot to bother me? This is why I am sitting here puzzled that now that I have a fully working claw, I literally have nothing for it to do. The only thing it could do for me is open an issue on forgejo for me that i dictate, but is that really a good use of tokens? I can just dictate into the issue body and tag it for "first plan then build". No hallucinations that way.
I was surprised you bothered to get one working. It sounded like you already had a bespoke system doing most of what anyone would get from openclaw.
For me, aside from how bloated and vibed together it is, it helped expose agent-system design. I had been neglecting it, and probably would have waited for some company to deliver some end-to-end experience. And, most of the end-to-end experiences are like "run it on our servers and we'll own/manage all your session memory without the ability to export."
I suspect the moat/lockin these companies seek is our session memory. At least, it's closest thing I've seen to something that might keep folks from switching.
I was surprised you bothered to get one working.
I said I would, so then I'm obligated by my word, even if I don't need it! I also want to have an opinion about it, so I need to (a) hack it and (b) run it for a while. Which is my problem right now. Nothing it can do will really help me solve something I haven't solved in a much more blissful way than the yolo vibe coders can offer! But I still want to have an informed opinion... I'm sure something will come to mind.
While I'm doing this playing, I'm having claude build me a bespoke data/document lake system [1] Maybe I can connect it to that and know everything about everything I care about... like a uniform interface to my meatspace memory backup/aide 🤔
I suspect the moat/lockin these companies seek is our session memory.
YES! Because then you can tune. I largely suspect that the training data for Claude post-4.0 is 100% the session logs: what we use it for, how we use it, where the friction is experienced, where the deadlocks happen... This is why it's still a dumb af bot sometimes when you do something audacious.
At which it is failing hard on the details rn fwiw - feels like there's some regression in Claude code atm, started noticing odd responses and more errors yesterday - or maybe they nerf'd it coming out of Friday with some tears in the ego? Too early to tell but I'm about to write some things manually because the back and forth is making me go nuts - My "openness" tracker says I have a PR open for a little over 4h30m now. That's unheard of! ↩
I said I would, so then I'm obligated by my word, even if I don't need it!
opti is lawful good.
I'm having claude build me a bespoke data/document lake system
The datalake bit is really neat. Like a digital collection representing all your tastes.
stands up I need more hardware.
~lol I'm looking for excuses to test search. This test went well. But failed for finding uber eats prediction market earlier.
It was in a video, I believe, not in text
For (1), I wonder how GUIX-y, bootstrappable, I can make an agent system. Like, what kind of agent "seed" do I need to somewhat reproducibly build me a full-blown agent system.
Two goals today:
I think (1) is where the hard part of making stuff is now. You build the factory that builds things now. Or, better, you build a factory that builds factories.