Claws are now a new layer on top of LLM agents

Posted by Cyphase 1 day ago

Claws are now a new layer on top of LLM agents(twitter.com)

https://xcancel.com/karpathy/status/2024987174077432126

199 points | 650 commentspage 4

deadbabe 8 hours ago|

Instead of posts about claws I would like to see more examples of what people are actually doing with claws. Why are you giving it access to your bank account?

Even if I had a perfectly working assistant right now, I don’t even know what I would ask it to do. Read me the latest hackernews headlines and comments?

hoss1474489 23 hours ago||

It’s a slow burn, but if you keep using it, it seems to eventually catch fire as the agent builds up scripts and skills and together you build up systems of getting stuff done. In some ways it feels like building rapport with a junior. And like a junior, eventually, if you keep investing, the agent starts doing things that blow by your expectations.

By giving the agent its own isolated computer, I don’t have to care about how the project gets started and stored, I just say “I want ____” and ____ shows up. It’s not that it can do stuff that I can’t. It’s that it can do stuff that I would like but just couldn’t be bothered with.

fogzen 4 hours ago|

Curious… why not just use a workflow engine like n8n? Seems most people are just creating workflows but without any deterministic execution.

throw03172019 11 hours ago||

What are people using Claws for? It is interesting to see it everywhere but I haven’t had any good ideas for using them.

Anyone to share their use case? Thanks!

unixfg 11 hours ago||

My favorite use so far has been giving it a copy of my Calibre library. After having it write a few scripts and a skill, I can ask it questions about any book I’m reading.

This week I had it order a series internally chronological.

I could use the search on my Kindle or open Calibre myself, but a Signal message is much faster when it’s already got the SQLite file right there.

cryptoegorophy 10 hours ago||

I am sorry to sound dumb but can’t cursor ai do this same thing? They have .md files with skills and knowledge

cap11235 9 hours ago|||

I'd imagine you could (never used Cursor myself though). I do a similar thing with my collection of books, but I just use Claude Code.

qup 9 hours ago||||

What's the relevance?

nsonha 9 hours ago|||

from your phone?

krackers 7 hours ago||

As far as I can tell it's mostly use-cases like "externalized claude code", accessible on mobile. Maybe the "agentic harness" is slightly tweaked for longer running tasks, but if it's really better claude code will copy the tweaks anyway, so I don't really see what the hype and point is.

daxfohl 7 hours ago||

I don't think AI will kill software engineering anytime soon, though I wonder if claws will largely kill the need for frontend specialists.

solaire_oa 6 hours ago||

To clarify, you mean that we're entering a post-HTML world, correct? As in, why spend effort on the aesthetics if a human will never see it, correct?

Because that is also my worry; a post-HTML and perhaps even a POST-API world....

daxfohl 5 hours ago||

Instead of "User eXperience", a new profession "Agent eXperience" will arise.

jesse_dot_id 5 hours ago|||

The LLM paradigm will never lead to AGI and to attach something other than AGI to all of your personal data and files — and setting it free whilst you sleep — is about as dumb as anything I can imagine.

The frontend will remain a requirement because you cannot trust LLMs to not hallucinate. Literally cannot. The "Claw" phenomenon is essentially a marketing craze for a headless AI browser that has filesystem access. I don't even trust my current browser with filesystem access. I don't trust the AI browsers when I can see what they're doing because they click faster than I can process what they're doing. If they're stopping to ask my permission, what's the point?

Mark my words, this will be an absolute disaster for every single person who connects these things to anything of meaning eventually.

daxfohl 6 hours ago||

And will there be a corresponding specialty that optimizes your "website" for claws to navigate. (Beyond just providing API access)

jesse_dot_id 5 hours ago||

I'd be kind of shocked if this didn't trigger the most harmful worm of all time eventually.

ramoz 5 hours ago|

AI is set to do that on its own given containment + alignment problems.

alecco 9 hours ago||

> Bought a new Mac mini to properly tinker with claws over the weekend.

Disappointing. There is a Rust-based assistant that can run comfortably in a Raspberry PI (or some very old computer you are not using) https://zeroclawlabs.ai/ https://github.com/zeroclaw-labs/zeroclaw (Built by Harvard and MIT students, looks like)

EDIT: sorry top Google result led to a fake ZeroClaw!

rane 8 hours ago||

This zeroclaw.org has to be some kind of malware.

This is the official repo https://github.com/zeroclaw-labs/zeroclaw and its website: https://zeroclawlabs.ai/

alecco 7 hours ago||

Oof! Thanks for the catch. I fixed the links. I swear it's what I get as top Google results for both "zeroclaw" and "zeroclaw github".

yjftsjthsd-h 9 hours ago|||

I assumed that was for running the actual LLM locally?

subarctic 9 hours ago|||

Looks interesting but I haven't seen it discussed much yet. How did you find out about it?

mbil 8 hours ago||

Well it's mentioned in the tweet this thread is about

> Anyway there are many others - e.g. nanobot, zeroclaw, ironclaw, picoclaw (lol @ prefixes).

mikert89 9 hours ago||

dude nobody cares about school prestige, the whole value in openclaw was that its an innovative idea, not that its written in Rust

alecco 7 hours ago||

From their GitHub repo: "Runs on $10 hardware with <5MB RAM: That's 99% less memory than OpenClaw and 98% cheaper than a Mac mini!"

nunez 5 hours ago||

I guess it's relieving to know that us developers will never get good at naming things!

Angostura 5 hours ago|

Don't worry, Microsoft will eventually name theirs something worse, probably pre-prepended with 'Viva'

... actually, no - they'll just call it Copilot to cause maximum confusion with all the other things called Copilot

_boffin_ 12 hours ago||

I just realized i built open claw over a year, but never released it to anyone. Should have released it and got the fame. Shucks.

thih9 14 hours ago||

How much does it cost to run these?

I see mentions of Claude and I assume all of these tools connect to a third party LLM api. I wish these could be run locally too.

kube-system 9 hours ago||

You can run openclaw locally against ollama if you want. But the models that are distilled/quantized enough to run on consumer hardware can have considerably poorer quality than full models.

Veen 7 hours ago||

Also more vulnerable to prompt injection than the frontier models, which are still vulnerable, but less so.

zozbot234 14 hours ago|||

You need very high-end hardware to run the largest SOTA open models at reasonable latency for real-time use. The minimum requirements are quite low, but then responses will be much slower and your agent won't be able to browse the web or use many external services.

hu3 13 hours ago||

$3k Ryzen ai-max PCs with 128GB of unified ram is said to run this reasonably well. But don't quote me on it.

panda888888 5 hours ago|

I really don't understand what a claw is. Can someone ELI5?

fogzen 4 hours ago|

It’s basically cron + LLMs + memory connected to their discord or WhatsApp to control remotely. A persistent personal agent that just does stuff for you. People have been running on their own machines letting the LLM access their shell, browser, whatever.

More comments...