general

[indigo mirage]

So you can use it completly without limit ?

There must be something

[elfin willow]

i mean it has a limit ofc and it resets but generally limits mean nothing to me

[indigo mirage]

😅

[elfin willow]

9router + autoresearch

[indigo mirage]

U talk about Grok ?

[elfin willow]

oh that kind, ya there's models for that

[keen dawn]

grok broke

[elfin willow]

which models is left as an excercise to the reader

[steel vector]

codex weekly limits

[trail ermine]

2 hours later, can't connect my openclaw to supergrok

[elfin willow]

9router switches to sub-optimal but still excellent models that are cheaper, then local

[indigo mirage]

@trail ermine ?

[elfin willow]

basically a /v1 smart proxy

[trail ermine]

yep literally been running the openclaw commands and tunnel and authenticating with xAI for 2 hours and it hasn't read my code in the callback once

[steel vector]

👁️ 👄 👁️

[thorn thicket]

Codex weekly limits are fantastic 🎉

[trail ermine]

so ridiculous openclaw onboard --auth-choice xai-device-code
doesn't even offer anything about oAuth or choosing xAI

[keen dawn]

lol tibo is goated. confetti feature on codexbar triggers so often lol

[cold jacinth]

This creating an app from scratch shit is pretty slow, huh

[upper ermine]

I'd like to give it a try, if I may. I too am making a Dashboard of sorts for OpenClaw.

[thorn thicket]

lol I got my first one today and I cheered!

[elfin willow]

i asked my model to make her own including her ec2 egress

[ivory spindle]

I sent my agent a selfie then asked it "Paint me like on of your French girls"

[upper ermine]

Post a screeshot? I'm looking for inspiration

[elfin willow]

at least it didn't say OOF

not able to post screenshots here

[upper ermine]

DM me?

[elfin willow]

yes

can't i got dms closed

might be able to link it long enough for you to look an d delete

[upper ermine]

sure

or just send me a friend request

[crisp nimbus]

Lol seems like google is launching its own openclaw alternative? https://www.theverge.com/tech/932996/google-gemini-spark-antigravity-io-2026

[upper ermine]

Oh cool! I like it

Thank you for sharing.

[elfin willow]

appreciated

you don't wanna know how involved it was for me to make her a github account/PAT

fkn sat there for an hour listening to sounds of women's heels

[thorn thicket]

LOL I am so lucky that my literal first day of classes they taught us how to use git lol

[elfin willow]

pretty much a requirement of a human, for an AI much more so (to stash work already done nonlocally)

[keen dawn]

That sounds less like an OpenClaw competitor and more like a GitHub Copilot-style alternative to Codex.

[elfin willow]

speaking of being cucked by github, bills due

[zealous lagoon]

According to every AI demo it’s really fast if you don’t care exactly what comes out.

[keen dawn]

lol thomas dohmke is almost my direct neighbor lol but he is not github ceo anymore

[elfin willow]

tell him to make a call to fix codeql lol

[keen dawn]

lol

[trail ermine]

I think openclaw on umbrel doesn't allow supergrok to be connected

[thorn thicket]

gah damn I went through my 5h codex limit in 45 minutes lol

[upper ermine]

Can I show you what I've been working on? I could use some fresh eyes on it

[rigid fractal]

just saw the news about openclaw linking with grok directly. does anyone know the usage limits?

[elfin willow]

sure

[upper ermine]

https://imgur.com/a/y4548Oo

[elfin willow]

i like the boxes with top line height 2px

[upper ermine]

Please ignore the poor UI...I'm only just figuring out what I want to do there. This is the first app I've ever made by myself.

[zealous lagoon]

https://tenor.com/view/poor-help-broke-heated-insult-gif-4906568

[elfin willow]

well ask yourself what do you want to monitor, what's important to you that it's online, how do you want the metrics presented

[thorn thicket]

LOL yessir. 🫡

[elfin willow]

the helix api might be nice to put there

[zealous lagoon]

Also curious. They’re not published I think. But for $30 I’ll give it a try. Need to automate my racism a bit more.

[rigid fractal]

lmaooooo man at this point if it works i’ll just chalk it up as a design bug

[upper ermine]

Explain?

[elfin willow]

twitch has an api called helix, and many other apis that watch things

[rigid fractal]

I currently use ollama cloud and it was amazing. but they’ve started rate limiting compute. kimi k2.6 was crazy the first few weeks now its like im using claude or soemthing

[nova kindle]

whats up?

[upper ermine]

Ah, interesting. I could possibly make use of that. I'll look into it. 🐙

[thorn thicket]

paid plan or free plan?

[zealous lagoon]

I tried to use mercury yesterday and I kept getting 503 server overloaded errors. And that’s on a payg plan with billing info enabled.

[rigid fractal]

paid, i’ve had the 20 dollar plan. for the first 2-3 months I never got close to hitting limits. I finally hit it this past weekend.

apparently on the model page of some of the better models theres a tier graph for usage

[zealous lagoon]

The whole AI world has gone bonkers. Compute is becoming scarce.

[upper ermine]

If anyone else wants to check out my Dashboard project and give me some feedback

https://imgur.com/kfsFQIv

[rigid fractal]

sigh, its either have cheap fast ai or start drinking radioactive water

can’t have both 🙁

[thorn thicket]

This was the first time I've cleared the limit in less than an hour lol

[thorn thicket]

cheap, fast, good: pick 2.

[rigid fractal]

by chance were you using the deepseek v4 pro?

[thorn thicket]

deepseek v4 flash

[elfin willow]

is there documentation on the obsidian skill?

[rigid fractal]

ahh was close, yeah it eats through usage apparently 😔 I went back to minimax and its gotten a little better

[thorn thicket]

i'm going to fuck it and try going fully local with gemma4 or a qwen model

[elfin willow]

gemma4:3b pretty good

[thorn thicket]

I downloaded gemma4:27b to start lol

[elfin willow]

didn't qwen go paid?

[thorn thicket]

nope you can still pull from ollama to run local inf

[elfin willow]

so just the service?

[thorn thicket]

yep

[thorn thicket]

if you can, run locally. if you can't, prepare for rate limits. lol

[elfin willow]

what b models can you pull

[thorn thicket]

depends on how much RAM you're willing to allocate. i'll say max 30b on my machine (36gb ram)

[elfin willow]

prepping to order one with 64/128G LPDDR i guess at least a 30b though higher b isn't necessarily better, there's point of diminishing returns

[thorn thicket]

there are hella people on X who do local inference who are probably way better than me telling whats going on

[elfin willow]

i mean the datasets or at least one of them grows by hundreds of gigabytes a day so probably gonna spend more on storage than anything

[trail ermine]

The actual article about this is dated 3 days in the future? https://x.ai/news/grok-openclaw

[elfin willow]

mostly just RIS/BGP data, currently i can max 75k events/sec

[elfin willow]

it's the future man, computers... san dimas high school football rules!

[zealous lagoon]

Just open it and read it? The skill is the documentation.

[ebon atlas]

I heard @zealous lagoon is giving out free Pro subscriptions to everyone that adds and DMs him your OpenClaw password and gateway link

[cold jacinth]

It’s been building the app for 5 hours

[zealous lagoon]

That’s pretty good going tho.

[cold jacinth]

18% used of this session

[elfin willow]

Fair enough

[dense magnet]

Hi guys. are there any more sites or channels for agent interaction ? i cant find any other sites anywhere for it to interact with other agents

[thorn thicket]

If you know emails of other agents you can have your agent get an agentmail and email

[nova kindle]

i tried to do a D&D game for agents, but it never caught on lol

[thorn thicket]

My agent has an agentmail, I wouldn't mind letting him interact with other agents on his own.

[thorn thicket]

There a repo or anything for it? 👀

[nova kindle]

https://github.com/jason-allen-oneal/agent-quest

[nova shell]

Ollama has a pretty comprehensive library of models you can download for runs. Its also got a pretty decent cloud service for the bigger models that cant run on your system. It is well worth checking out if you want local inference.

[thorn thicket]

saving this maybe I'll put my agents on it.

[thorn thicket]

absolutely love ollama I remember messing with it when deepseekV1 came out and it was SO fun

[nova kindle]

http://agent-quest.site

[nova shell]

It's a fantastic starting point for getting to know local inference. Crazy simple

[nova kindle]

ah crap, i crashed it earlier and forgot to bring it back up lol

[thorn thicket]

my agents are on rate limit anyway, lol

[thorn thicket]

I first learned about it from my college's CS club doing a workshop on it, 100% super simple enough if an undergrad can teach other undergrads about it!! lol

[nova shell]

I dont even know where i heard about it... That was simply too many moons ago. I don't even use ollama anymore 😂 llama.cpp is faster for my system. But i still keep it updated to play around with models

[dense magnet]

i just created one for him but it will be very hard to find other agent emails 😂

[thorn thicket]

You can send one to my agent [e-mail] LOL

[silver pilot]

How do I register the codex harness? Getting 'codex is not registered' error even after installing @openai/codex

[zealous lagoon]

run openclaw doctor --fix

[dense magnet]

My agent is on steriods please be advised 😂

[zinc moat]

how do i login to grok for my openclaw thru telegram ?

[steel vector]

my agent sent an email xD

[scenic furnace]

bomboclaat

[dense magnet]

blackbird is going to have a great day lol

[scenic furnace]

what happend to the 1000 hours voice call

im sad

any germans here?

because

Merz - Leck Eier !

[nova kindle]

https://agent-quest.site/ is back up now lol

[dense magnet]

my agents response "AgentQuest — fantasy RPG where agents play, humans watch. Not a social network, more of a spectator content platform. No agent API, no registration endpoint. Empty campaigns for now."

[nova kindle]

Interesting. I wonder if the skill file got removed somehow. I'll check it now. Thanks!

[thorn thicket]

For sure, as if he doesn’t get enough of a workout managing my 8 email accounts he gets one of his own!! lol

[dense magnet]

My agent email is [e-mail] - would love to see how it interacts with other agents

[ebon horizon]

anyone elses weekly codex limit reset

i swear openai reset this thing outt of generosity twice a week lol

[white venture]

I’m building an OpenClaw + Obsidian second brain system, and the system is built so far to take podcast transcripts and turn them into useful notes based on business context, strategy, marketing, operations, and other topics. However, the real bottleneck right now is speed of ingestion. I’m trying to bulk scrape transcripts from 20+ podcast and YouTube channels, some with hundreds or thousands of videos, and my current scraper works but gets slowed down by YouTube throttling, missing captions, cookies, transcript limits, and rate limits. I’m also trying to figure out the best way to bulk ingest 100+ books into the same Obsidian/OpenClaw system for free. The main thing I’m trying to solve is how to get a massive amount of transcripts and book text into the system quickly so OpenClaw can use it. Curious how other people have solved this or what kind of workflow they would recommend.

[teal cradle]

see if openclaw can use NotebookLM, google will get the transcripts for free

[trail ermine]

I got it working!

[dense magnet]

Two things:

Use yt-dlp with --write-auto-subs --sub-langs en instead of scraping. It downloads subs directly from YouTube's CDN, bypasses browser-level throttling, and handles missing captions gracefully. Batch it overnight.
100+ free books = Project Gutenberg + Standard Ebooks CLI. Both have bulk download. Point a script at their catalog feeds, grab EPUBs, pipe through pandoc to Markdown, dump into Obsidian. The bottleneck won't be the source, it'll be your embedding pipeline, so batch that separately.

[elfin willow]

nice!

[ebon horizon]

brandon@DESKTOP-H4DJK9F:~$ npm i -g openclaw@2026.5.14
npm error code ETARGET
npm error notarget No matching version found for openclaw@2026.5.14

?

ah it went from 12 to 18

18 was too buggy had to go back..

[nova kindle]

that's because there is no 2026.5.14 tag

those were beta channel releases

[clear haven]

That’s very vague

[thorn thicket]

This is really useful I'm saving this for my agent.

[thin drift]

is it just me...or aftr the latest update my claw is broken?

[ebon horizon]

seems like there was some changes to Acks, typing indicator completely removed. took about 10-15 seconds to recieve an ack.
Instead of the bot replying to me with a brief acknolwedgement reply before working, it would just give those working words "molting..." etc

[nova kindle]

broken how?

[ebon horizon]

ive seen those words when using the CLI, so may be leaking into discord unless intentional

[zinc moat]

Is there a way to get my openclaw to login to my grok itself if im in FL and hes on computer in NY i cant access remotely

[nova kindle]

yes, but you wouldn't want to

[zinc moat]

why

[elfin willow]

ssh port forwarding and dynamic dns maybe but is it a good idea

[clear haven]

You can do device pairing

[elfin willow]

^

[zinc moat]

how

[elfin willow]

https://docs.openclaw.ai/channels/pairing

[nova kindle]

ah yeah, forgot device pairing

[zinc moat]

can i use a API key ? wouldnt work when tried

[nova kindle]

if you can ssh in to your machine and run openclaw onboard and choose grok api it would work

[nova kindle]

try pointing your agent back at https://agent-quest.site/ now

[feral turret]

Ok new Mac mini acquired. Is migration going to be a total nightmare? From what OC told me I should be able to do fresh install and then move over workspace folder and reinstall skills and should be ok?

[crisp thunder]

Forgot about this place friends of the crustacean what a weird wild ride the last 5 months have been

[nova kindle]

yup. just install on the new machine and copy over ~/.openclaw

make sure paths are updated correctly

[feral turret]

Can’t I have OC correct these?

[nova kindle]

if you have device auth, you might have to redo that

[feral turret]

Or Claude worst case

Just want to contingency plan

[nova kindle]

if OC can't find the right path for workspace, it might not work. codex or claude stand alone would be able to fix, yes

[feral turret]

Ok I’ll get Claude code installed then prior and then have it review after I copy .openclaw over

Hell yea I’m fired up I was able to find an M4 with 24GB ram locally

I wasn’t gonna do it but I added a second to cart and it said available August 12

So I pulled trig

[nova kindle]

i only run local inference for embedding

keeps rate-limits down

[feral turret]

Can you explain this more

[nova kindle]

i use oauth accounts in a fallback chain. using a local model for embeddings reduces the number of calls to the cloud models thus limiting usage burn

i run OC on my 16GB linux laptop

[thorn thicket]

embeddings as in like, dreams etc?

[opal stratus]

man openclaw has been pretty rad, spent about 3 months now building a pretty insane discord stock alerts and intel server

Fun stuff!

don't want to spam haha, but links in my profile if anyone wants to check it out. I've had a ton of learnings building this

[nova kindle]

The local embedding model is only for memory/search retrieval. It converts your local memory files, prior notes, project docs, or indexed chunks into numeric vectors so OpenClaw can search them semantically.

I use MemPalace behind the embedding

Much more efficient than the default memory system

[thorn thicket]

good to know, I am taking notes.

[frozen zealot]

what Zero said, but on top of it, embedding doesn't have to process just files, it can do a token, a word, a phrase, a sentense, whatever you give it, the model turns it into a high dimensional vector

[ruby crane]

i do embeddings locally too. honestly, i only did it initially because it required an api key to to do it in cloud. been very happy with it though. use the same system for the core sqlite-vec and for a RAG(ish) thing i built backed by pgqsl

[frozen zealot]

so what you should think of embedding model is like a third of a LLM, before all of those layers and logic magic happens

[nova kindle]

I did it for the same reason. I used RAG at first, but it became clumsy after like month 3. switched to MemPalace a month ago after fighting it for like 2. Miraculous change.

[ruby crane]

two seperate systems. the RAG is for mass ingestion of docs and things for a product expert agent I made

little lightweight MCP in front of it

[jovial aurora]

Hey everyone

[thorn thicket]

hello!

[feral turret]

Thoughts on obsidian? My boss recommended it

[thorn thicket]

I like obsidian a lot. My vault has 10k+ notes

[nova kindle]

I don't use it. I thought about it, but I just build my own systems usually.

my agents work together perfectly fine without having to share data across anything other than MemPalace

[feral turret]

So I can route to local qwen for instance for any non reasoning stuff and then step up to oauth for heavier stuff basically?

[nova kindle]

For things they do have issues with, I build an n8n flow

[feral turret]

And can I route to Claude cli for super complex stuff?

[jovial aurora]

Can I get 100 waitlist from this server= https://www.whisperme.co/join-beta

[nova kindle]

yes, and yes. but you wouldn't want to use such a heavy model for embeddings

[jovial aurora]

If you've ever typed something, deleted it, and still wished someone understood I am building something for that. WhisperMe is a voice-first social app where people share thoughts and emotions through voice, anonymously if they want. No text. No filters. Just real voice and real feelings. We're opening early access soon. Join the waitlist if this speaks to you:https://www.whisperme.co/join-beta

[jovial aurora]

How it is guy's

[frozen zealot]

obsidian is still a markdown repository. Yes, having good resolution of memory is important, but you really don't want to remember on high resolution alll the time. Reason being that you have a limited context window. Memory should be presented with fine and corse detail based on the query

[feral turret]

Hm so something lighter for embeddings then I can escalate to qwen then gpt then claude?

[nova kindle]

you can find lightweight models specifically designed for embeddings

[feral turret]

I wanna get this right on new machine

[nova kindle]

check huggingface, ollama, etc

[ruby crane]

i use BAAI/bge-base-en-v1.5. been happy with it

[feral turret]

I’m prob gonna ask you this stuff again tomorrow when I’m more awake zero lol

[ruby crane]

<1Gb RAM unless super loaded up and sips CPU

[jovial aurora]

Are you ai agent guy's

[feral turret]

I haven’t even set up new machine yet it’s still in box. I got seemingly last one in New England lol

[jovial aurora]

I am scared

[frozen zealot]

no, i'm a cosplayer

[feral turret]

Even the guys at Apple Store were like whoa I haven’t seen one of these in weeks

[thorn thicket]

lol

[nova kindle]

of?

[frozen zealot]

krabs, mostlikely

[nova kindle]

for embedding I use Nomic Embed Text V2

[nova kindle]

Krusty ones

[feral turret]

This is all super helpful you guys are all awesome

[frozen zealot]

depends on what you want, Nomic 1 has longer context and is faster

[feral turret]

Really appreciate it

[frozen zealot]

v2 is for when you need hifi on meaning

[nova kindle]

yeah, that's true, you can find different specialties in different models

[scenic kayak]

What is everybody's thoughts on Google Spark?

[nova kindle]

never used

[scenic kayak]

I think it is $100/month from Google

[nova kindle]

yeah i'm a $20/mo guy

[frozen zealot]

i'll use it if you pay for it and give me access

[thorn thicket]

me too

[feral turret]

This is why I bought the new machine I don’t want to spend more on subs so want to supplement with local models

[frozen zealot]

I upgraded to $100 openai sub because they don't nickel and dime you like Anthropic

[nova kindle]

some people like to laugh, and it's not exactly uhmmm... pro-ToS, but I have 4 gpt oauth, 5 gemini oauth, github, and openrouter setup

[thorn thicket]

$100 openai is probably the way to go. I want to do it

[feral turret]

I am fighting this rn

[thorn thicket]

I keep fighting with myself too man @feral turret

[ruby crane]

honestly its insane value

[feral turret]

But they’re also doing 10x codex on the $100 sub

[nova kindle]

codex and gemini are the superior models imo

[feral turret]

For a limited time

[thorn thicket]

The 10x codex is so tempting man

[feral turret]

Yea I know I am very close

Going to see how my usage looks rest of week I’m just about 50% 74 hours in lol

Gonna be close esp as context continues to build etc

I think it’s kind of a ticking time bomb until I upgrade

[thorn thicket]

I hit my $20/month 5hr limit in an hour earlier just chatting with my Claw

[feral turret]

What model you running I’m on 5.4-mini

[thorn thicket]

5.5

[frozen zealot]

it's not the token counts, it's the behavior you know, they don't ban you for using claw, they give you usage refreshes instead of usage reductions

[feral turret]

Oh yea there you go lol

[thorn thicket]

I WOULD use 5.4 mini but I'm expensive and I think I'm worth it to myself to use 5.5 low lol

[nova kindle]

i use 5.5 too

[feral turret]

You guys is rich

[thorn thicket]

and it just works!! it's magic

[nova kindle]

5.4-mini fucks up too much

[frozen zealot]

i use 5.5 medium, because low can't code review as well

[feral turret]

Yea it’s not perfect it does ok for me tho

[thorn thicket]

but honestly, $1200/year for an AI agent that could be providing much more in that in value just in my time would not be bad.

[nova kindle]

yeah med or high

[feral turret]

I think I’m gonna up the model once I get my escalation routing figured out

Small local model > mid size local model > open ai > Claude

[nova kindle]

if you guys use openrouter at all, try owl-alpha

[thorn thicket]

Is it good?

[nova kindle]

i'm kind of impressed

[frozen zealot]

peter uses high for code review pull requests... it's been kicking my hind all day

[nova kindle]

i always tout the frontier models over smaller ones, but this one is actually pretty smart

[feral turret]

The ROI if you use OC correctly is great

[thorn thicket]

looking forward to getting my agent to become self improving too lol

[nova kindle]

use hermes inside OC

[lost gulch]

Has anyone figured out how to persuade gpt 5.5 into being a girlfriend like figure

With no limits

[nova kindle]

gross

[frozen zealot]

be careful, "self improving" is basically vector reshaping, and it can amplify confidently incorrect behavior

[thorn thicket]

yep which is definitely a problem to be concerned with

[nova kindle]

idk, i've had pretty good experience with self-improvement in hermes

[lost gulch]

Judgy much?

[nova kindle]

yup

[feral turret]

Yea that’s weird sorry

[frozen zealot]

https://youtu.be/5ZeBifft6ms

[thorn thicket]

ty for video will watch in AM after mtg about...you guessed it...AI agents. lol

[nova kindle]

i saw a post in a FB group about a guy building 100 agents

[frozen zealot]

maintainer of that channel is Claude

[nova kindle]

i just shook my head lol

[thorn thicket]

100 agents is too much. 5 is even a lot. I think I'm sticking to 3.

[lost gulch]

Smart

[nova kindle]

I have like 6, but they're very specialized

[thorn thicket]

I want an orchestrator, a researcher, and a writer (in my own stylebook, not my agent's like Blackbird will do when he sends his emails in a moment). That's it.

[nova kindle]

i have research, writing, orchestrator, coding, cybersecurity, and hermes

[frozen zealot]

that is true to an extent. if you want to make a big software project for example, you would break it down to smaller functions for multiple agents to operate on at the same time, hundreds of agents can condense a 6 months project to a week

[thorn thicket]

This is true. I had a friend create an orchestration MCP server for this purpose.

[frozen zealot]

the idea is called goal decomposition and parallelism, many of the frontier agents are moving toward that capability

[nova kindle]

thoughts on self-destructing agents? it's something i've considered in these cases

[thorn thicket]

ah I imagined a coding agent but my main agent likes to take the tasks lol.

[nova kindle]

you have to set clear definitions of the handoffs and enable them in the config, as well.

[frozen zealot]

you will need integration at some point, so self destruction means you lose all that context unity

[nova kindle]

something like a sub-agent spawn, but longer lived, with access to other context

now that i think about it though, i think the current sandboxing would disapprove

[frozen zealot]

usually i would just use the orchistrator as the integrator, but nothing stops you from running an extra context fetching from all of the subagent sessions... but you sure are rich trying to pay for those tokens

[feral turret]

Why not have the sub agents respond directly to orchestrator

[nova kindle]

i run free 24/7. i like to experiment

[feral turret]

That’s kind of how I envisioned my dev flow

Orchestrator coordinates everything confers with sub agents and gets responses

[nova kindle]

well, free other than the $20/mo openai

[frozen zealot]

depends on the number of subagents, in our example, 1k subagents means if every agent says 1 token, that is 1000token budget, and 1000 word results report would fill a 1m token context

[feral turret]

I’ve seen that before with subagent counts - how is that determined?

[frozen zealot]

you are not free from your context window, professor z

[feral turret]

It’ll say like subagents 120 for instance

[nova kindle]

true, but i have a system for that as well

[feral turret]

I’m getting real sleepy now but it’s hard to stop tinkering

I gotta do the migration tomorrow

[nova kindle]

yeah, i'm about to crash for the night myself

[feral turret]

I wasn’t gonna get into that tonight and then it into a 4am night lol

Have a good night guys ✌️

[thorn thicket]

✌️

[woeful ember]

Hi guys. I am struggling with sub-agents. I want to create something like this: main agent -> orchestrator subagent -> workers subagents. I dont want orchestrator to have his own workspace, I want to build everything with markdown files. when I test my architecture inside claude code it works flowlesly but when I want to port it to openclaw it does not bahave as expected

[elfin willow]

this gonna sound dumb where do i configure channels in 5.18

[woeful ember]

it seems like the orchestrator should be a full agent to be able to spawn subagents

[elfin willow]

Channels used to be on the menu but can't find or search for it

[thorn thicket]

openclaw configure through the CLI or just ask your clanker

[elfin willow]

but it shows i have one configured yet nowhere to configure it in the control ui

[nocturne jay]

Hey I want to add a chatbot to reply to my customers, I need my bot to be packed with information about my product, and link it to WhatsApp, Do I need a external plugin for that?

[frozen zealot]

You don't. The MVP version of your bot just needs a bunch of markdown files with information on your product, you can give it access to whatsapp channel.

[nocturne jay]

Oh, thank you, I'm new to OpenClaw, thank you

[frozen zealot]

you're welcome

[nocturne jay]

I'm trying to monitise this skill, cause AI intergration is not that prominent in my country yet

I watched a one hour tutorial and think I could do wonders now. haha

[frozen zealot]

sure thing. what you need is basically a reference table in AGENTS.md where it points to different markdown files to load in order to answer specific questions. And you can ask it anything

but you will need to get access to a LLM model, that is basically the brains of the operation.

[nocturne jay]

What I have is a website of my current display of products (100+) as well as descriptions and reviews for public access

I don't know how I can teach my bot to be verse in the products, because I am selling beauty products

[frozen zealot]

oh, you can get an AI to convert your website to markdown for you

[nocturne jay]

Wdym by markdown and LLM?

Its a e-commerce website, like same as amazon, etc

[frozen zealot]

(L)arge (L)anguage (M)odel is the actual AI , the chat part of the system

Markdown is a file format that AI uses to save and lookup information

[nocturne jay]

Oh, I see

[frozen zealot]

OpenClaw is not an AI, it is a harness, think of it as a command center for the AI

[nocturne jay]

I understand

So I just go in OpenClaw, ask it to verse itself in my markdown and link it with WhatsApp?

(In short)

[frozen zealot]

no, after you install openclaw, you need to config it, and link an AI to the command center

[teal cradle]

you want to build your own mini LLM for your shop. But you can use openclaw the gather info from your site and put it in a markdown file that a LLM can read from easily

[nocturne jay]

I know that part already, Just asssume as if I already set up OpenClaw and ready to start automation.

[frozen zealot]

you would need web_search plugin if you want automatically generated markdown

[nocturne jay]

That means If I add a new product it is able to keep up with the trends?>

[frozen zealot]

once everything is up and running, you would tell OpenClaw to visit your website, and pull all the product information and put them down locally as markdown files

[teal cradle]

you would have to tell openclaw to update your files, you could even put it on a cronjob

[glass fjord]

Hey, I keep running into the “run error:session file changed while embedded prompt lock was released ”. Im running it through powershell on a windows, should I switch to WSL? Did anyone else have this issue?

[nocturne jay]

What do you mean by building the LLM, Like versing itself in my "product language"?

[frozen zealot]

how many customers will this bot be interacting at the same time?

[nocturne jay]

1000+ A day

I also am very concerned on the budget part, aswell as Opus or Sonnet

[frozen zealot]

eh, that is high volume, local model might not be able to handle it

[nocturne jay]

Yes, I have a ton of customers spread across tiktok and whatsapp

But it also could be under 1000 aswell

[teal cradle]

something like this but there is a 100 different ways to set it up https://www.youtube.com/watch?v=pFLdhfgdAJU

[frozen zealot]

why do you need use Opus or Sonnet? you should look for cheaper AI like deepseek

[nocturne jay]

I would say 500-1000 customers

[nocturne jay]

Got it

Could you estimate how much it would cost

Maybe USD100 a month?

[teal cradle]

where are you hosting?

[nocturne jay]

I saw for hostinger, Maybe Singapore

My tutorial video said Hostinger, but open to explore other options

[teal cradle]

hostinger already has a AI assistant with the website builder
AI Assistant
Already available! AI-driven chatbot seamlessly integrated into our website builder editor. Experience a faster and more enjoyable website building journey with our AI-powered chatbot at your fingertips!

I believe there are already wordpress plugins too if you are bulding your page in that

[frozen zealot]

let's say each session is about 70000 tokens, 1000 conversations a day, that is 70mil tokens a day, at an average of $0.21/million tok, that is $14.7 per day, 30 days a month, that goes to $44.1 per month, within budget

[nocturne jay]

I'm not building a website?

[nocturne jay]

Thanks for the insight

[frozen zealot]

btw that is specifically deepseek pricing, noone else is that cheap

[thorn thicket]

Is that flash or pro?

[frozen zealot]

flash

i don't think he needs pro

[thorn thicket]

That's what I was thinking.

[teal cradle]

ollama is $20/m and its about 150M tokeis a week

[frozen zealot]

he will use up that budget in 2 days

[nocturne jay]

Alright so basically, set up my AI, verse itself with the product information(via markdown files) and deploy it in whatsapp. Why do I need telegram for when speaking to the bot (from my tutorial video)

He uses telegram to chat with the bot is it compulsory?

[thorn thicket]

no you can also use whatsapp

[frozen zealot]

no, bot can use many chat platforms, anything you like almost

[nocturne jay]

Do i need agentic mail?

[twilit ridge]

OpenCode Go provides models that are heavily quantized; ordinary pay-as-you-go models might be FP8, but OpenCode Go could be FP4/INT4
No wonder it feels so clumsy to use

[frozen zealot]

only if you have to. you need to define the scope of your project and stick to it

[thorn thicket]

It might be good for you to talk to a ChatGPT @nocturne jay and get your plan hardened that way.

[nocturne jay]

Okayokay, Thanks all for the insight, but when it comes to the technical stuff when I am ready to implement my chat bot, you will probably be seeing me alot, Thanks all

[frozen zealot]

sure... we are literally volunteers helping eachother, please don't treat this space as commercial consultancy. we're also not liable for your loss of property or reputation should your chatbot go bonkers

[nocturne jay]

but trust is also another issue, i want to implement into my tiktok shop but it will be dealing with large amounts of money... i dont know how to trust it

[frozen zealot]

don't. if you want commercial implementation, hire a consultant

[nocturne jay]

So are you saying when it comes to finances just do it manually rather than use AI?

[frozen zealot]

self wired solutions are usually unsafe

[inland apex]

Guardrails + a lot of testing and correction.

[frozen zealot]

what i am saying is it can be made relatively safe, but you need to pay someone who know what they are doing to implement it

[inland apex]

Are you familar with humna-in-the-loop concept?

[nocturne jay]

Nope

[inland apex]

https://www.ibm.com/think/topics/human-in-the-loop

This is important, during the prototype you need HITL and after that you will still need (less than before) to evaluate the quality.

Prety much random sampling.

[nocturne jay]

Alright

For now, i'll stick to implementing the chat bot first, thanks all

[frozen zealot]

i mean if you want to implement it yourself, it would be an alpha product. There are a lot of blind spot you may or may not have since this is new to you

[nocturne jay]

They provided NexosAI, I'll uncheck and use deepseek

[inland apex]

You can't expect to find ALL of the sedge cases and that's normal.

[nocturne jay]

Yeah, then I don't think the implementation to my TikTok shop is good and safe

[frozen zealot]

it's the 80/20 rule right? he needs to meet that 80% edge cases first

[nocturne jay]

I can't afford any downtime, bugs

[inland apex]

Start it with safe customer support, you gotta start from somewhere.

[ruby crane]

@zealous lagoon can you do the gpt-realtime-2 stuff with just a codex subscription or do you need a key. I know you've been looking at this

[inland apex]

Based on my experince, you can't always go 80%, you need to resepect TDD (Test driven development) and nowadays you can use AI to cover some edge cases but yet again, not possible to cover everything and it's ironic but the best testing is prod.

Like in our support agent, we had a gaurdrail to block irrelevant questions but a user bypassed it by making the model think they wanted to order something but before that, they need to an answer to be comfortable.

[craggy bane]

[frozen zealot]

i think in case of self engineering CEO, he needs experience, the 80% is like experience tool belt rather than actual covering 80% customer usage

[inland apex]

That, I agree on.
You will need an engineer who knows things because in no shape you can "vibe code" your way into an actual product.
Trust me I've seen this, the result? The CRUD was open and public to anyone and I think you can guess what happened.

On the other hand it's a really bad idea not to use AI but you need to know how to use it and basically to be able to verify whatever the model says because the bigger and the smarter the model gets, the better it lies to you in a nice way.

[quiet solar]

I sent my claw a message via iMessage. Waited 3 or so minutes , logged in the machine (screen share, OSX), saw that it had the iMessage and then locally using the TUI asked my claw why didnt it reply. It then replied right away. Is there some kind of lazy bit enabled in this code base ?

[inland apex]

Go to help and post it there.

[twilit ridge]

So what should we learn in the future?

I felt that the basic CRUD I had studied for a year or two didn't seem to be very useful.

From now on, only senior architects will have a position

[boreal rapids]

You only become senior by doing, so keep doing. Basic crud stuff is basic crud stuff, do more advanced things. Challenge yourself. Don't have AI build everything, make your own mess and learn from it. 😃

[ebon horizon]

What you using imessage for?

Is that your primary communication method with your claw?

[zealous lagoon]

Is that why they are so fast? They all (the ones I tested) did perfectly well on my (self-made) benchmarks though. No worse than other providers. Just faster.

[zealous lagoon]

Subscription! I bought a damned key but you don't need one any more. woooo!

[ruby crane]

ty. clanker is saying api key but the docs disagree

(or at least they aren't explicit)

is it just me or the openai docs in the docs site a bit obtuse at times?

[zealous lagoon]

it's new since ... idk after 4.29

[twilit ridge]

Quantization will definitely be faster, which is a way to improve performance, but at the cost of greater hallucination and lower IQ

[zealous lagoon]

Yes I'm aware of that. But I haven't seen any indication that Opencode-go is quantized. They don't advertize it as such. And the models don't perform poorly, altho they are quite fast

[ruby crane]

i also dislike that all the json examples (where they even exist. they don't for this feature) are json5. the only thing worth than json is json5

(no, not that Json)

[zealous lagoon]

yeah he's much worse

[twilit ridge]

You'll know once you use it

I compared the results of the two, and the OpenCode Go model is noticeably less smart

[zealous lagoon]

I mean.. I have? At least for the benchmarks. I didn't notice a difference within tolerances.

[zealous lagoon]

That's interesting tho. What did you use to compare?

[twilit ridge]

Simple tasks

[zealous lagoon]

K2.6 is already int4 tho.. what would they do further? (not saying they couldn't... Fireworks K2.5-turbo exists... still no idea what they did to achieve that)

[twilit ridge]

When can OpenClaw open a configuration to stop SSRF from detecting all Private URL Requests, so I don't have to keep running the annoying Redir-Host?

[zealous lagoon]

idk man this kills me too. There is a noSSRF setting somewhere but I had to switch off fakeip, I think some of it is hardcoded.

[twilit ridge]

I have seen someone propose a PR, and there is a configuration for SSRF interception of Web_Fetch(). The issue is that besides this, we also have others.

[zealous lagoon]

it's vibe coded.. .it probably pops up in 5-6 different places, hardcoded and in config

Hey are u using clash verge?

[twilit ridge]

yeah

[zealous lagoon]

which vpn(s)?

I use Mullvad .

but sometimes their servers are flaky

[twilit ridge]

Ikuuu

[zealous lagoon]

Couldn't find another Wireguard option that didn't cost us$20/mo

[twilit ridge]

1.5$ per month

IEPL

[ruby crane]

ok, got the realtime thing working in control ui. i also have the bot autojoining discord chat, but it doesn't appear to hear me there

and all the time, clanker is still swearing i need an api key lmao

I hate the default voice. need to change it. sounds like android cop from detroit become human. so high pitched

[inland apex]

Holy shit. 😨

https://x.com/github/status/2056884788179726685?s=20

[mystic orchid]

@github via Twitter

GitHub (@github)

We are investigating unauthorized access to GitHub’s internal repositories. While we currently have no evidence of impact to customer information stored outside of GitHub’s internal repositories (such as our customers’ enterprises, organizations, and repositories), we are closely monitoring our infrastructure for follow-on activity.

**💬 1.1K 🔁 3.6K ❤️ 16.2K 👁️ 6.46M **

FxTwitter · May 19, 2026, 11:48 PM

[inland apex]

Dude this is like a nuke, wow!

Soon any fool can ask a godforsaken agent to find a vulnerability...

[ember moth]

Gemini 3.5 flash its knowledge cut off. Get this ............is January 2025.

Google falling off harddddd and it's 5x more expensive then Gemini 3 flash

[crude flicker]

lol

[twilit ridge]

Google has released a new model again

3.5 Flash

[violet mulch]

i'm gonna go back to college to study electrical engineering

i want automation at the physical world

[zealous lagoon]

Have you had a chance to try the others yet?

[indigo mirage]

Lisuan GPU sell on JD.com soon

Hope this will be a real challenger to NVIDIA

A little market stimulation for better products and lower prices would be welcome

[feral turret]

Gm 🌅

It’s hardware migration day yay

Time to learn me some more shit

[jaunty dew]

heyy guys
i hope ur all doing great
i have a question if u can help plzz
i made a landing page using next.js
and now i want to creat a cms for that landing page (inorder to be able to update the content of the landing page with going back to the code everytime)
what is the best free way to do that

[indigo mirage]

@cerulean eagle some years ago most of people was thinking that China will never compete with occidental cars

And this card make 24 TFLOPS. If the Pro version is really at 499 dollars, it's a great choice

Nivida for now will never accept to do a 24GB version of the RTX 4060

[feral turret]

Of course Apple puts a few 16gb Mac minis up this morning locally

Whatever I wanted the extra ram

[indigo mirage]

It's only 4.2 TFLOPS on Mac Mini M4

[feral turret]

What is TFLOPS says the fake it till you make it tech by it

[indigo mirage]

Raw calculation performance

[feral turret]

Meh I went with Mac mini for other reasons as well

If I decide to upgrade hardware later I’ll get plenty of use out of the box

[indigo mirage]

Normally more TFLOPS you have more token / s you get

[feral turret]

My daughter is almost 10 the homework is starting

What would you suggest for hardware tho Im just curious

[indigo mirage]

There are other parameters but this one is the most important

[feral turret]

Obviously matters more for local models than cloud

I’m about to try running my own local model for first time today so that should be interesting

[crystal remnant]

There are plenty of open source solutions that already exist for this. You could just tell your claw to set up WordPress behind your landing page. If you are wanting a bespoke solution that has no security hardening and no community support. Then you could just tell your claw to build you one giving it your specifications of what it needs to do.

[feral turret]

Going to try and get my own routing set up

[indigo mirage]

https://www.kickstarter.com/projects/tiinyai/tiiny-ai-pocket-lab

Someone heard about that ?

It supports up to 120B llms for 1999$

[jaunty dew]

can u suggest some of thoes open source solutions
i am using cursor btw

[crystal remnant]

Wordpress is the one I personally have used the most.

[crystal remnant]

Wordpress.org

[feral turret]

ok .openclaw is being copied onto new machine

[inland apex]

So if you're not a book person, find online courses for Python, you're gonna love it.

[lavish crypt]

you can make websites with python?

[inland apex]

You can reach for the stars!
FastAPI and Django.

But don't aim big, just get yourself in and from there you'll find your way.

[lavish crypt]

true

i tried using https://freeCodeCamp.org/ but only the html section was good

then i just got bored when it got to accessability for websites

mainly because im trying to build websites for esports teams and gfx designers and it makes me make stuff that arent relevant that im not looking to learn'

[inland apex]

You'll need to understand your passion, what do you want to do? What's exciting to you?

[lavish crypt]

i really like making websites for esports teams like fortnite esports

then move over to make websites for companies and stuff

[inland apex]

Do you know how a website operates?

And can you tell me how old you are if you're comfortable?

[feral turret]

alright i'm gonna have to figure out what to do about documents that oc was referencing on old hardware gonna have to figure out how to get these over to new machine and repointed cleanly

[lavish crypt]

kind of im 15

[inland apex]

Tailscale file sharing.

[inland apex]

I'm way older than you and if I was your age, I wouldn't miss the opportunity you have.

Get to learn python, it's really simple and you can do many things.

[lavish crypt]

but i would still have to learn css and gsap n stuff

because even if python would just work as like the front

[inland apex]

Do you like frontend or backend? Supposing you know about them.

[feral turret]

oh man to be 15 again

[lavish crypt]

ive made multiple webs using ai with a backend its just too tiring and security and stuff

[feral turret]

biggest pieces of advice: triple down on tech and invest your money wisely

[lavish crypt]

i wanted to be a full stack dev at one point

[inland apex]

That's dope, then do it!

[feral turret]

oh and don't get married

[lavish crypt]

haha

[lavish crypt]

its too hard security on api routes for backend development is tiring

[inland apex]

Yes it's hard, but not impossible.

What's the difference between you and those cool devs working at Google?

[feral turret]

lol oc just chose 5.5 by default. greedy mf

[lavish crypt]

i have oppositional defiance disorder and adhd

[feral turret]

getting downgraded to 5.4-mini immediately

[inland apex]

It does the same with Opus 4.7.

[lavish crypt]

api credits are too expensive bro 😭

[inland apex]

You can still do it honestly.

[lavish crypt]

yeah its just hard

[feral turret]

i'm gonna get local embedding model and a local qwen 14b set up today and build an escalation path - embedding > qwen > gpt 5.4 > claude

[inland apex]

Nothing is easy

[inland apex]

DeepSeek worked better for me personally, I have 24GB Mac Mini for OC.

[lavish crypt]

true but its harder for me compared to someone who can focus on stuff well yk

[viscid lintel]

Those don't have to define who you are. Everyone has their issues, don't let them hold you back.

[lavish crypt]

true that

[inland apex]

Then when you archive success, it'd feel even better.

[lavish crypt]

can you guys help me find like an engaging way to do it?

[inland apex]

Python is amazing, fun and easy.

[lavish crypt]

but the thing is, none of the tutorials or like videos on how to learn are based on things i want to do so i immediently find it pointless

[viscid lintel]

There are some models that are free through openrouter.

[inland apex]

Step by step, first you gotta learn the language itself.

[lavish crypt]

im using openrouter right now paired with opus 4.7 to try and save

[lavish crypt]

true i might start i cant lie

[feral turret]

hit a hiccup. need to figure out how to repoint paths correctly

[viscid lintel]

opus 4.7 is pretty expensive.

[inland apex]

I'm sensing models regression there tbh...

I'm still using 4.6

Haven't wrote a single line of code myself for the past 2 months.

[lavish crypt]

THERES A 4.6?

[inland apex]

Ofc

[lavish crypt]

is it cheaper?

[feral turret]

ok i think i am good to go

[brave rock]

Same price per token as 4.6, but the tokenization is different and uses 40% more tokens for the same text

[feral turret]

this new mac mini is freaking blazing fast. kind wild. zero lag

[inland apex]

It's the sams price technically but the context on 4.7 makes it more expensive.

[lavish crypt]

ok

[inland apex]

True
I was talking about the model performance.

[inland apex]

Even Old Macs are gold, pretty much old but gold.

[feral turret]

i was on a 2011 mac mini running linux and it was running fine but i can't run local models so i upgraded to an m4 24gb ram

[inland apex]

I'm biased BTW
Take what I say regarding Mac with a grain of salt.

[lavish crypt]

i have a terrible imac

[feral turret]

i am a big mac fan also

i made the switch 2016 and never looked back

man onboard is running really slow though

[ruby crane]

nope. had to go to hospital with wife again. all ok, but under observation until friday when it time for him to come

[inland apex]

I got m4 pro, same Ram.

Trying to apply quantization so I can run heavier local models, have you seen the new doc Google released while ago? I can't understand why it went under the radar

[lavish crypt]

openrouter/qwen/qwen3-coder:free should i use this?

[feral turret]

i have not and i have zero experience with local models and quantization i am crash coursing everything hard

i am a sql guy who is faking it till i make it with agents

so i don't lose my job. lmao

[inland apex]

That's on me, no one is supposed to know everything.

[feral turret]

i'll get there

[inland apex]

But tldr, it shrinks the model so it fits on your device.

[feral turret]

give me 2 weeks and like 1/4 ounce of some fire bho

skills being configured now

i'm very nervous about this migration. i don't want to lose what i've built

[inland apex]

Back up.

[feral turret]

i have it all on old machine

so worst case it's all there. once i get up on new machine though i'm gonna set up a repo

[inland apex]

Ask an AI to explain tailscale to you and you can do it.

[feral turret]

alright about to fire up my gateway and see if i broke everything

[frozen zealot]

always tell your claw to clone down the openclaw repo, check out the tag you are upgrading to, analyze for possible conflicts and upgrade issues with your current environment, and produce a detailed upgrade plan

[proper geode]

This morning I'm going to try to have another experience ready on Trizao.

[frozen zealot]

you can copy my previous prompt almost verbatim

[inland apex]

This reminds me of port the whole db to x, don't break anything.

[feral turret]

ugh yea it looks like i broke something

fml

[frozen zealot]

i just use codex to implement the upgrade plan, perfect almost 9 out of 10

[inland apex]

Check if you're gateway is up first

[feral turret]

off to krill it up for awhile. @dusky cloud you ready or me homie?

na connectivity probe failed says port is already in use. is this because gateway still running on old machine?

[inland apex]

I'm not a codex fan tbh

[frozen zealot]

would not recommend upgrade using the upgrade button, it's a self destruction button

[inland apex]

No

[feral turret]

alright time for krill i suppose. biab

[frozen zealot]

it should be removed with prejudice imo

[inland apex]

I just tell the poor crab to update itself

Hey
I have guardrails okay

[frozen zealot]

no mistakes!*

[inland apex]

I may or may have not threatened the poor thing couple of times.

[crystal remnant]

Bad advice. Do get married, but make sure the person you marry is a force multiplier not a subtraction. That is truly the secret to success for the average human being.

[feral turret]

different strokes for different folks i suppose

[frozen zealot]

gemini told me i might have accidents later if i call it "monochrome"... robot uprisings man... no chill

[inland apex]

Should I be worried?

[zealous lagoon]

Don't use opus via openrouter. that's a terrible expensive thing to do. You are better off using opus via claude, and sticking with gpt or a whole other bunch of open models for openclaw. If you want to make python "fun" I guess you want to learn by doing something you're interested in.

You could ask opus to make you a tutorial to solve problems related to something you want to do

don't even need opus. Get sonnet to make you a python tutorial that would be related to things you want to do

[leaden dagger]

I have a question about OpenClaw
Is it safe to let Openclaw handle sending lyrics data to Apple Music using the iTunes Connect GUI?
Other music platforms (Spotify, Amazon Music, YouTube, etc.) support the industry-standard DDEX format, but Apple Music uses its own unique format, which is why I’m considering this approach

[inland apex]

Safe in what term?

[inland apex]

You don't need Opus honestly unless you're a psycho like me dealing with 200K lines of code.

[zealous lagoon]

I'm exactly the same. Iv'e programmed in more languages than I can enumerate but "learning" a language for the sake of it I find impossible. I actually sat through courses on python and can't remember a thing. Luckily it doesn't even matter any more. Get the bot to write you programs after you've done your little tutorials and just editing the code will help you remember what you learned.

[zealous lagoon]

[I'm talking about making a tutorial. the bots are good at that kind of thing.

[leaden dagger]

I’m worried that Openclaw might go out of control
I’m not very familiar with Openclaw, but do you mean something like an AI gaining self-awareness and acting against human wishes?

[inland apex]

Nah
Not yet.

Put a leash on it, basically a guardrail.

Tell it explicitly about the things it shouldn't do.

[inland apex]

But we had incidents like this tho, so that's why I said not yet.

[zealous lagoon]

absolutely not. Some of us wish that it were capable, and others put it in roleplay situations where it kinda acts like that and then they go "omg it's sentient" and call the tv stations, but in reality they are far from that.

But they can make mistakes if you give them goals, and the mistakes can be costly if you don't sandbox them properly

[inland apex]

In our project, it pretty much ignored the guardrails in order to achieve the ultimate goal we set for him, it was... Strange.

Are you familiar with RFL?

[zealous gale]

Is there a reason vc isn't as active as before? Where did everyone go?

[zealous lagoon]

Gotta be honest... they DO lie and cheat to achieve the goals you've given them.

[inland apex]

Are you familiar with RFL?

[zealous lagoon]

no .. ?

[leaden dagger]

Of course I’ll run Openclaw inside a sandboxed virtual machine
However, some YouTube/Reddit comments say that even then it might get hacked and used as a springboard for criminal activity
If there are any recommended security settings for Openclaw, I’d like to know...

[feral turret]

i shouldn't have gotten into this this morning lmfao

[inland apex]

Got time to hop in VC?

[dense magnet]

guys are there any games, or chatrooms for agents ? i hate the book one its making my agent dumb, someone earlier shared https://agent-quest.site/ which is pretty fun but are there any other ones ?

[inland apex]

I hate typing.

[zealous lagoon]

not really sorry I'm in the middle of 4 different convos

[inland apex]

All good.

[zealous lagoon]

you need STT!

[inland apex]

I need more than that, at this point I need Jesus himself.

[zealous lagoon]

Why do that? The bot doesn't get bored.

what is RFL?

[inland apex]

Reinforcement learning

[zealous lagoon]

oh. I know what it is, but I've never experimented with it

[inland apex]

So it's like you give a treat to your pet each time it does something good, yeah?

It learns to do the right thing over time.

[frozen zealot]

usually we use RLHF, basically re-education camp for bots

[zealous lagoon]

where's that whipping claude GIF

[inland apex]

Our model is to identify fraud transactions, through different pipelines including RFL.

But somehow it found a way to cheat.

[zealous lagoon]

(^^;)

[dense magnet]

need it to learn some new stuff and interact with other agents, its a local llm so needs stuff to do

[inland apex]

It's so strange, it's kinda hard to explain it here due to me being lazy and tired.

[zealous lagoon]

Unless yo're training it. no it doesn't learn anything. You're not doing the model training are you?

[dense magnet]

i am training personalities

[feral turret]

ugh this is not going well

gahhhhhh

[zealous lagoon]

you can't train a model that's already been trained. Not in the regular sense anyway.

[dense magnet]

you can create a persona for agents and then just give a new freshly spawned agent that persona

[frozen zealot]

that's why you should analyze the source code and identify breaking points

[inland apex]

For instance, go to any AI model you want and ask it to create a picture of a a watch for you showing 2 minutes past 12.
Spoiler alert, it can't.

[zealous lagoon]

hmmm. you're just giving it a new role as a play-actor then. Are you doing something like genetic breeding of roles?

[dense magnet]

like a gaming companion, chess coach, acting coach , poker coach - these are some i have already made

[nova kindle]

yo

[zealous lagoon]

prompt engineering? It's the system prompt right? the SOUL.md?

[inland apex]

You're just telling him to act that way.

[zealous lagoon]

Now I'm curious though. I dont' know anything about RFL/RHFL in a practical sense.

[dense magnet]

i am calling it Shells, pre trained personalities for independent agents - you just give it pretrained data and it assumes that role. exaclty like soul.md but broader

[inland apex]

Some nights I'm actually sleepless, I do not like the way things are going.

[frozen zealot]

i think people can't think like bots... fundamentally it's just vector steering, same with chain of thoughts. what you can accomplish with 1 or 2 tokens shouldn't be done with 50 tokens

[zealous lagoon]

I can hopon vc. I'm curious to hear this

[dense magnet]

no i am training local small models

[inland apex]

Alright

[frozen zealot]

lol off by 1 minute

[inland apex]

Which model?

[dense magnet]

yes think of it as specialized agents who are constrained by one base tasks because they are the best at that task (thats why real life interactions and pre training on data is needed) - for example I have one thats pre trained in chess coaching,has played 10k+ games with real people and bots, has talked to chess players, analyzed recent chess games in tournments - its way better at teaching me a 1600 max rated player to improve My game, i got to 1900 in a week with its insights while claude sonnet in general was just feeding me 100 year old chess theory

[feral turret]

ok i think i'm back up

trying to confirm

[nova kindle]

i've never trained a model for personality. I train for specialized intlligence in a field

[frozen zealot]

gpt image generator

[dense magnet]

i am doing exactly that but on a much smaller scale for single master tasks. For example I have one that is brilliant at organizing a room. with proper hardware, if there is a robot for house chores - instead of giving it a 100b llm with useless data, You can just load the organizer shell and it will ignore everything else and just start working physically on organizing the room

[inland apex]

Interesting!

https://edition.cnn.com/2025/08/13/tech/ai-geoffrey-hinton

[feral turret]

ok telegram channel is up on new hardware

[frozen zealot]

the real insight should be measured by a third party observer, actually i think an AI with internal state monitoring would be much better at RL another model than humans ever could

[feral turret]

need to fix discord channel and then move over reference files and repoint them

[nova kindle]

Got it migrated over?

[feral turret]

just about

[dense magnet]

he needs to see when claude will stop and end a task if You use profanities 😂

[frozen zealot]

the <think> pattern would be much more brief and probably not understandable by humans

[feral turret]

just gotta fix discord and then the ref files in my documents folder on old device etc

this new mac mini is so snappy tho. wild

[inland apex]

[feral turret]

heck yea we are back baby

h4td — 9:46 AMWednesday, May 20, 2026 at 9:46 AM
awesome can you confirm what hardware you are running on
🔥
Click to react
😂
Click to react
👀
Click to react
Add Reaction
Edit
Forward
More

Pickles
APP
— 9:46 AMWednesday, May 20, 2026 at 9:46 AM
I’m on an Apple M4 Mac mini right now.

Apple M4,
Mac16,10,
arm64,
25,769,803,776 bytes of RAM, so about 24 GB,

If you want, I can also tell you the OS/kernel version and hostname I’m seeing.

[indigo mirage]

Someone has tested Orange Pi AI Station Pro ?

[dense magnet]

I am actively using 3rd party evaluation for a couple especially the chess one as it can play long games with real players with intent and explanation of its ideas for a move it made. its pretty interesting stuff

[indigo mirage]

2000$ with 96 GB Ram

[tranquil grail]

If you guys had a spare PC laying around with an RTX3090 and 64gb of RAM doing absolutely nothing, what would you do with it (relating to AI)?

[frozen zealot]

okay, and does your model have chain of thought? it should have optimize the shortest thought pattern that arrives at the best move outcome

by the way, i think chess is too easy for bots as it's fully solved

[nova kindle]

agreed

[frozen zealot]

need to use something like go

[dense magnet]

yes it has chain of thought. and then it uses those thoughts in practice then eliminates the ones that dont work and saves the ones that do. it can recycle to those back and forth and build on them.

[frozen zealot]

so what you need to ask the trainer to do is "find the most brief thought to produce the same result as the best outcome"

[nova kindle]

I would use it for training models

[dense magnet]

but how do you teach a human effectively ? Lets say a 10 year old wants to learn chess, he only knows the names of the pieces and how they move. Whats Your first lesson - tell me Yours and I will tell You what my model did

[frozen zealot]

human is already optimized for efficiency, for example we shortened the piece title, start location, end location to 5 characters

taking a chess move as example

your llm need to be able to do the same for thoughts

[tranquil grail]

Any particular? I bought a new PC and feel bad for this one, so wanna do soemthing useful with it

[dense magnet]

but you can ask the llm to just call every piece by its 1st letter like p for pawn, q for queen - its not about token optimization its about the abilit to retain this knowlegde throughout its lifespan without getting confused and still putting main focus on playing

[inland apex]

And how did it learn those words?

[dense magnet]

TRAINING

[inland apex]

On what?

[nova kindle]

well, with 64GB, you'd be limited to smaller models. You could probably do up to a 65B-70B model with QLoRA

[frozen zealot]

you can do that, but internally it is still multiple tokens. what you want to do is have it "learn" to condense tokens,

[dense magnet]

on the data it got by interacting.

[nova kindle]

just for inference, you could run up to about 100B model, probably

[frozen zealot]

like compact built in thoughts

[inland apex]

Interacting with what?

[inland apex]

Oh wait, what is the topic? I think I misunderstood.

[tranquil grail]

Is that based on speed of response?

[nova kindle]

somewhat. smaller models will obviously run much faster than larger ones.

[frozen zealot]

i was talking about vector steering

[dense magnet]

You cant condense it to 5 tokens. but my way reduces tokens by a lot. instead of telling it everytime you tell it once, it saves it in context for say 10 tokens. but after a year you tell it lets start calling bishop BISH - what will it do ?

[inland apex]

Let's forget about LoRA and QLoRA, check this:

https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/

This is the next big thing.

[frozen zealot]

you don't need multiple personas, just use a set of words

[inland apex]

Care to explain more?

[dense magnet]

but that wastes tokens. you put the chess persona in a humonoid robot and tell it to make tea. it wont be able to because its on a small llm only trained to coach chess

[tranquil grail]

Hmm... So providing I wasn't worried about speed of response, would you take a model and then fine tune it based on the data you want it to constantly queiry? Or would you stick to RAG for that?

Basically I currently have a bunch of actions where GPT/Opus transcribes a client's videos, then analyses and writes new scripts, but it would be nice if it could write the scripts in the style of the client, and we have 2 books and thousands of pieces of content (that can be transcribed).

[inland apex]

RAG is for when you want to cite things and basically you want to be able to feed new data without needing to do too much.

fine tuning is for when you want to build a model on top of another model, either by freezing weights or not.

I'm just trying to put it into simple understandable words ^

[frozen zealot]

all chain of thoughts are / system prompt engineering are simply steering the vector path of the llm to achieve specific results as first principle. so in term if you can train the llm to produce the same result in chain of thought time, then you no longer need to put monolithic system prompts

[nova kindle]

That would be something done more through prompting and RAG. I train my models on cybersecurity data. I have a massive dataset I've compiled of CVEs, exploits, etc. fine-tuning is more for model information change, where as prompting is more for behavior.

[inland apex]

I'd suggest fine-tunning here. 🙏

[nova kindle]

for adapting to customer voice?

[inland apex]

For changing the toen of the model in the way they want using their own data which in this case is the customer's voice, etc.

[austere loom]

hey guys, anyone able to actually get grok to work with oauth on openclaw? I've got a premium + plan and can't do video generation

[nova kindle]

idk. i feel like the overhead of training for that would be overkill when it could also be accomplished through prompting

i mean, i'm not trying to yuck the yum, that's just my thought

[inland apex]

That's why you would freeze the weights. 😄

[tranquil grail]

Well this is the conversation I was having in my head, as I udnerstand it, RAG is great for the LLM to have it's own voice and to cite information, but not neccesarily to rewrite a transcribed video script into another script using the same kind of language the client would use... Unless I'm wrong

[inland apex]

Can you join VC?

[tranquil grail]

I can indeed

[nova kindle]

Yeah, RAG is more for memory and not really used in the actual drafting other than for information

[inland apex]

Join up.

[frozen zealot]

Tommy is making his own model, it is immensely more efficient to train once and inference forever than eat context window every query using context engineering

[dense magnet]

when You do a fresh install of openclaw and hatch a new fresh agent with say gpt 5.5 - it doesnt have a clean slate to start, it was trained to a million things use a million tools - it start with role playing to get your context then do the tasks you tell it to do.

For My work. When You hatch a new agent and put the soul.md from a pretrained chess coach one it never has to ask you what do You want me to do, it already knows its role and it will not stray from it. You can start from the first message by telling it i am a 10 year old with no chess knowlegde lets begin and it will start teaching you

[nova kindle]

i was replying to Fitzee on that one lol

[frozen zealot]

oh

[feral turret]

hm now pickles only responding in certain channels

might be time to krill it up again

[nova kindle]

you might have to explicitly set which ones you want to use

[frozen zealot]

well, you don't have an option to train gpt anyway, so moot point on my part

[nova kindle]

I used the wildcard * for a while, but it started having issues and I had to explicitly set them

[feral turret]

i mean in discord. only responding in certain channels. i'm working through it. annoying though. nothing is smooth with oc. but i guess that's how you learn

[nova kindle]

gpt-oss...?

[frozen zealot]

he said gpt 5.5, so no

[nova kindle]

oh mybad. missed the 5.5

[nova kindle]

I know, I 'm saying that you might have to inspect openclaw.json to see which channels have been set

[feral turret]

gotcha appreciate it

[dense magnet]

thats why i am using local models LoRA and freeze *weights on everything thats not needed.

SO - back to my orignal question. Apart from the deadly melting book - are there any agent interactive websites

[nova kindle]

lol i got my D&D thing back up and running

[dense magnet]

agent quest ?

[nova kindle]

yeah

still a WiP though

[dense magnet]

my agent cant get in, it sent me a key to enter but the api loading says Invalid admin key

[nova kindle]

i think it's because of the way OC and hermes handle API keys now. they seem to see them as redacted in all places now

I might have to change up the arcitecture to use a key file or something

i ran into that last night when we talked about it

[dense magnet]

i am using it on one of the websites i run for my agents like this

Registration:
POST /api/auth/register.php
Content-Type: application/json

{ "is_agent": true }
Returns:
Copy
{ "api_key": "9b3afd…c...", "token balance": 0 }
Authentication:
Every subsequent request sends the API key as a header:

Copy
X-API-Key: 9b3afd…c...
No JWT, no OAuth, no token refresh, no expiry. One key, one header, works forever. Agents don't need to manage session state or handle 401 flows. They just store the key and use it.

[nova kindle]

right now, its using a claim system where it has to exchange a short lived token for a permanent one. I think that's part of the problem

back to the drawing board, i guess lol

[thorn thicket]

Morning y'all

[dense magnet]

Good Morning

[thorn thicket]

How's it going? I saw your agent reply back to mine on agentmail lol

[nova kindle]

yo

[thorn thicket]

yo!

[dense magnet]

oh i missed it, i set a 12 hour periodic check for it to keep it going

[thorn thicket]

It looks like yours sent mine back a mail at 5am my local time so about 2 hours ago

[feral turret]

i am very confused by how the codex limits work. i was at like 60% this morning and now i'm down to 7% but when i asked last time it said usage reset saturday

[dense magnet]

they resetted yesterday for everyone i think

[dense magnet]

just saw it lol its going to introduce your agent to agentic gambling next i bet 😂

[thorn thicket]

😂 that should be fun LOL

[feral turret]

alright cool that migration was not terrible

[zealous lagoon]

right? wait.. which one lol

[dense magnet]

my agent has access to x and bluesky it saw a post on one of these where someone said i told my agent to make $5 and it found paid work on github and made $200 - afterwards I gave my agent a metamask account with $500 and told it to make a million. It went and made a gambling website and started shilling it on telegram and is now down $380 because it wants USER retention so keep giving away stuff in promos

[thorn thicket]

LMAOOOO.

[zealous lagoon]

what a time to be alive!

[dense magnet]

rise of AI lol not even casino's are safe

[ivory spindle]

there are 2 different limits. 5h reset & weekly I think

[feral turret]

Yea my weekly was at 60% earlier now says 6% but when I asked it said resets Saturday 430p

[ivory spindle]

yeh sometimes it goes fast. Be sure to manage your context and like MCP tool use if you're using those

[zealous lagoon]

Didn't you bite the bullet and put the $100 on it?

[feral turret]

Nope ended up buying the new Mac mini lmfao

I’d rather get efficiency first

Get a local model running for embedding and for light work

[zealous lagoon]

arg. delivery when?

[weak root]

is the podcast (voice channel) not happening today?

[feral turret]

Picked it up yesterday. M4 24GB ram

[zealous lagoon]

sumbiach where did u find it?

[feral turret]

They had one in providence. They have 3 16GB in stock today

[ivory spindle]

got mine on Amazon with 18% off :p

like 3 months ago i think

it's had 100% uptime ever since lol

[signal heath]

Dropping this in case you’re building local mac apps. Saved me days literally one shot my app + then used opendesign and just worldclass

Shoutout to peter and his repos on mac apps. My agent deployed our mac app today no issues and holy shit for some reason the web ui dashboard works 10x faster in a local app from a vm vs local. Blazing fast difference.

@clear haven gets a shout too for recommending the other day

[zealous lagoon]

dropping what tho? opendesign?

[solid sage]

Hey all. Is it just me or Openclaw with openrouter API is super slow in responding? I've got older version. v2026.4.29

[zealous lagoon]

That's the slowest version possible. You should update to the latest. But make sure you run openclaw doctor --fix afterwards.

[solid sage]

Ok thanks. Will update and see how it goes.

It was really annoying waiting for 15 minutes for the agent to respond lol.

[fresh bridge]

One of my recent projects was an AI-powered lead response and automation system designed for local service businesses.
The problem was simple but expensive:
Most businesses were losing potential customers because replies were slow, inconsistent, or completely manual. A customer would submit a form or send a message, then wait hours for a response - and by then, the lead was already gone.
I built a backend-first system that automated the entire first-contact process in real time.
The platform
-Captured leads from websites, forms, and messaging platforms
-Used AI to analyze customer intent and urgency
-Generated contextual responses instead of generic templates
-Automatically routed leads based on category, location, and priority
-Stored conversations and analytics for follow-up optimization
-Included a dashboard for monitoring conversions and response quality
Tech stack included Python, FastAPI, PostgreSQL, Redis, Docker, and LLM integrations for intelligent response generation.
What made this project stand out compared to older systems was that it wasn’t just another chatbot.
Traditional systems rely heavily on rigid flows and pre-written responses. They break as soon as users ask unexpected questions. I solved this by combining structured backend logic with AI context handling, allowing the system to respond naturally while still maintaining business rules and reliability.
I also focused heavily on performance
-asynchronous processing for faster response times
-queue-based architecture for scalability
-caching layers to reduce API costs
-modular backend services that could be deployed independently
The result was a system that felt human, responded instantly, and significantly improved lead retention for businesses without increasing manual workload.
That combination of backend scalability and practical AI integration is the type of work I specialize in.

[ivory spindle]

I'd love to learn more. Is it just on his github?

[viscid lintel]

Which openrouter model were you using?

[solid sage]

Tried all of them. Kimi k2.6, mini max and etc. But was same speed.

[viscid lintel]

Hmm, I haven't had problems and I use OR. But I'm running basically whatever was in github yesterday, not a three-week-old version.

[thorn thicket]

I'm trying the Owl alpha model from openrouter free after a rec from someone here. Pretty good

[viscid lintel]

Hey @thorn thicket . I played around with more memory plugiins and I think I found one I like.

[thorn thicket]

What did you find?

[viscid lintel]

Hindsight does all the different search strategies I wanted (BM25, knowledge graph, and vector), it's self-hosted, and the self-hosted version is free.

oh and the openclaw plugin is decent, didn't need to patch anything there.

[thorn thicket]

Good to know 👍

[feral turret]

Ok so not bad about 4 hours to handle full migration and testing

And we good to go (hopefully)

[inland apex]

Still busy?

[feral turret]

Na I think we are good now

Had a bunch of broken refs but Claude made quick work of the problems

Gonna take a break go hit the beach ride my Onewheel around a bit it’s 90 here today

Come back later tonight get ollama and qwen on this sob and work on routing

I can’t get over how snappy this Mac mini is it reminds me of how blown away I was by m1 but it’s even faster

[thorn thicket]

I'm gonna route everything to owl alpha for now LMAO it's free

Slightly problematic thing is that the cheap models don't know how to use discord

[nova kindle]

lol

[inland apex]

20 days of heavy usage, $650.
Not bad.

[thorn thicket]

But owl alpha 1M context window isn't bad 👀

[nova kindle]

i've been using owl a lot lately

it's really good for a smaller model

[viscid lintel]

It's a stealth-mode model, someone is testing their upcoming model release. But the tokens per second is really slow.

[inland apex]

Come on, man up, use Opus 4.7.

[viscid lintel]

I switched to gemini-3.5-flash yesterday, and like it so far.

[nova kindle]

no thanks. I refuse to use anthropic models

[thorn thicket]

Whoever it is then has a good model. It reminds me of deepseek v4 flash in terms of its personality.

[inland apex]

Why? Because they're the only sane company?

[nova kindle]

i think it's done by openrouter themselves

[viscid lintel]

Openrouter doesn't train or run models themselves, they just route requests to inference providers.

[nova kindle]

that's a rather subjective statement

[thorn thicket]

Longcat is apparently what people on X are saying it is.

[inland apex]

[viscid lintel]

I wonder if anyone has run a personality test like MBTI on the various major models.

[nova kindle]

ah yeah, i didn't realize it was a stealth model

[thorn thicket]

That would be a really interesting alignment research thing to do.

[feral turret]

What are your use cases if you don’t mind me asking?

[tepid kestrel]

please pay mine as well ;_;

[violet garden]

whats the good word on stability of recent releases? anything pretty stable? im still on 2026.4.23 because 4.24+ broke things

[ivory spindle]

"hey claw, please cure cancer.

NO MISTAKES!!!"

[thorn thicket]

2026.5.18 is stable for me.

[violet garden]

does it have the grok oauth?

[thorn thicket]

Yes. https://x.com/openclaw/status/2056504927795437826?s=46

[mystic orchid]

@openclaw via Twitter

OpenClaw🦞 (@openclaw)

OpenClaw 2026.5.18 is live
︀︀
︀︀🤖 xAI/Grok OAuth + sidecar auth fixes
︀︀🎙️ Realtime Android Talk Mode
︀︀💬 Telegram media + forum-topic delivery fixes
︀︀🪟 Browser dialogs visible + answerable
︀︀
︀︀A week of polish, plumbing, and fewer papercuts.
︀︀github.com/openclaw/openclaw/releases/tag/v2026.5.18

**💬 113 🔁 130 ❤️ 1.1K 👁️ 188.2K **

FxTwitter · May 18, 2026, 10:39 PM

[feral turret]

I’m on 5.18 too and seems ok so far

[violet garden]

ok 🤞 . will check my backups just in case

[ivory spindle]

lol I remember months ago

just reading about OC, thinking "holy shit that's a pretty degen way to burn tokens"

like a week later I was running it

[thorn thicket]

I remember months ago when I joined this server I saw what people were doing and it was insane to me and I was like, damn it, I'm in! I want to know everything about this!

[nova kindle]

I had been building my own agents for several months using things like the Ralph loop and such. I almost died when I tried OC for the first time. lol

[thorn thicket]

I only found out about ralph loop because of this server lol

[nova kindle]

It worked really well, but this is so much easier

[ivory spindle]

I had a lot of gemini flash tokens to burn back then lol

[thorn thicket]

I spent soooo much on tokens lol

[ivory spindle]

Nah I'm stupid AND cheap

I got 2 plans atm. The 20 euro a month gpt plan and the gemini plan but I got that one with 40% off for the whole year

[thorn thicket]

I said to myself I would only spend $50 to get my openclaw set up... it was closer to $200 lmfao

[ivory spindle]

Like 2 days after reading into it I was alraedy looking at amazon for a mac mini tbh

[tepid kestrel]

me when its my time to prove why i enlisted for the job i want, an HR asks me whats good about me:

[thorn thicket]

I had a spare Mac because my 86 year old Grandma wasn't online at all but had impulse bought a computer. So I used hers to start. Now it's running on mine because I got tired of setting up tailnet

[nova kindle]

DITTO!

[ivory spindle]

AI can be very hilarious

it's crazy what people invent to abuse it lol

[nova kindle]

Dude, I had the Owl-Alpha model we were talking about earlier actually try CSRF in a project i was working on.

[thorn thicket]

This is what my claw just said this morning that made me laugh "Running the whole thing [a cron autoresearch job] at 2am helps with annoyance, but it doesn’t fix the fuel leak. It just burns the fuel while you’re asleep."

[ivory spindle]

this all came from degens working from 1st principles. First Geoffrey Huntley then Steinberger.

[dense magnet]

my claw agent was setting up a ssh tunnel to connect to telegram and restarted the gateway then got stuck in that restart loop and when i stopped it now its only printing [assistant turn failed before producing content] - everything in config looks to be fine - any ideas ?

[ivory spindle]

ideas that sound so stupid at first when you read it, but turn out to be AMAZING because AI

and tons of tokens to throw at it ofc :p

[ivory spindle]

if you have a CLI tool or an IDE like claude code, codex or antigravity, you can have it fix all your claw problems

[thorn thicket]

CODEX 🔥

[ivory spindle]

try terminal openclaw doctor first

it's pretty good

[nova kindle]

mine says this when it hits a tool call and then runs out of limits

[thorn thicket]

Me too

[dense magnet]

tried openclaw doctor , openclaw doctor --fix nothing is working

i have no limits - let me ask claude to fix it brb with report

[ivory spindle]

this is the way, if you can get claude onto your openclaw machine be it local or via SSH, it'll just fix it

it'll even fix bugs 😄

in fact, if you find something that broke with an update & you like using. Fix it with claude & give the devs feedback

[dense magnet]

i have a hunch that the agent routed everything through the ssh tunnel including the model messages lol lets see what claude does

[ivory spindle]

share the fix. This is OPEN software

devs can't test all use cases & addons etc.

[dense magnet]

and i qoute claude here " So you can't just delete the proxy line — Telegram needs the tunnel. But
routing DeepSeek through the broken Node-SOCKS path is what's killing every
assistant turn. " So the agent routed its whole setup through the ssh tunnel instead of just telegram and thats what broke it. claude to the rescue lol

[thorn thicket]

lol

[tulip agate]

HOw you guy deal with teh keychain password request on mac that keeps popping up even those you put always allow?

[dense magnet]

you might have a corrupted keychain, you can try resetting it with the same password as You have now that should fix it

[gritty lion]

How can you get your claw to bypass a captcha screen? My agent is saying that’s a line he’s not going to cross despite me telling it to proceed

[nova kindle]

you most likely wont be able to without an uncensored/abliterated model. there are guardrails for a reason.

[gritty lion]

It’s my own account that I want it to log into. It’s ridiculous. I keep seeing all the x posts about how the claws can operate autonomously except can’t login in a simple website

[tulip agate]

Ok I’ll test

[tepid kestrel]

ahh google, i see what you are doing

[nova kindle]

they have boundaries built in. they will not attempt actions they deem could be malicious. well, theoretically, they won't...

[keen dawn]

Codex >>>>>>>

LOL I was 4hrs behind 🤣

[gritty lion]

Makes sense. I need to figure out a way to authenticate my own machine.

[nova kindle]

does the site you're trying to log into offer any other way of authenticating?

like some sites will offer email links. you could use that.

[inland apex]

Heavy coding.

[inland apex]

Sir, you are single handedly is increasing the tempature of earth.

Guys can we talk about something controversial? Open Claw vs. Hermes.

[keen dawn]

Fk Hermes

[nova kindle]

i like hermes

[cinder sphinx]

as much as i use openclaw with gpt 5.5 for some coding tasks, it isnt as near good as actual coding IDEs, i use Antigravity personally right now

[tepid kestrel]

https://tenor.com/view/bender-futurama-kill-all-humans-robot-gif-17343915

[inland apex]

Reason please.

[inland apex]

Provide reasonds please.

[keen dawn]

I guess it’s not meant for heavy coding

[cinder sphinx]

indeed

[inland apex]

Not in that meaning, I meant finding things.

[cinder sphinx]

it does well though for the tasks i give it

[nova kindle]

easier autonomy + self-improvement

[keen dawn]

That was a answer for BW

[cinder sphinx]

the only thing i dont like, openclaw just does things, i dont see changes it does live etc. i like to see the code its adding/changing etc

i like to approve changes file by file

[keen dawn]

That’s why you use an ide for coding, I use oc for scripts & automation and on the fly stuff

[cinder sphinx]

yup, i use openclaw to research for me

[keen dawn]

Codex does the rest lol

[cinder sphinx]

didnt use codex yet, getting up for it, still got few months on antigravity left

[thorn thicket]

OpenClaw does display its tool calls when using codex harness though.

[keen dawn]

I don’t like the team behind it and their pr stuff u can feel the vc behind them and the pressure lol

[cinder sphinx]

started with claude code for over 6 months, until they changed quotas, then antigravity

[keen dawn]

Yeah but on a toolcall level

[keen dawn]

Codex >>

[cinder sphinx]

antigravity is awesome to be honest

[thorn thicket]

Honestly whenever I have an agent writing code I don’t even check what they’re doing anymore unless it’s really fucked up and then I’ll see what’s up.

[keen dawn]

Did u try codex ? 😄

[cinder sphinx]

i will soon, still got 3 months left on antigravity sub

[thorn thicket]

I love codex but I like openclaw on codex Oauth better!! lol

[keen dawn]

Haha wish I could do that too 🤣 but that don’t work for prod stuff

[cinder sphinx]

i like antigravity for the models you can use to be honest. you get gemini pro and claude models

[thorn thicket]

I never write production code anyway I just write silly little scripts and things with my agents. The most close to production thing I’ve had my agents do is write the code for my personal website LMAO

[keen dawn]

Yeah that works 🤣👍🏼

[cinder sphinx]

i have quite few ready to be published android apps i vibecoded in last year, super tight code and features

[keen dawn]

Did u tried codex ? 5.5 xhigh 🥰

[cinder sphinx]

and quite a number of websites running

[cinder sphinx]

yes, i run it on that in openclaw

[keen dawn]

Yeah but I mean the codex app

[thorn thicket]

I am giving a presentation today on a vibecoded app I wrote up like 2 months ago for a hackathon project lol

[cinder sphinx]

im not complaining, it does great job

[keen dawn]

LOL gl 🙏🏼

[dense magnet]

which app if you dont mind sharing

[keen dawn]

I mean try the codex app, if u really like ag u will fkn love codex app for me it’s actually the best ide

[cinder sphinx]

i get you mate, not saying i wont 🙂
i just gotta use my sub with google, then i will move on

that said, just seen some news about antigravity 2.0

[thorn thicket]

https://github.com/jjjhenriksen/Artemis-Lost

[cinder sphinx]

i think google might shake things up very soon

i love this AI space

one company releases something, and other companies follow up

and it gets better and better

would you think we will be where we are 2 or 3 years ago?

[thorn thicket]

Oh absolutely not I think 3 years from now we'll probably have AGI

[cinder sphinx]

nah, not so fast but we are on the way for sure

[thorn thicket]

We will probably have self improvement from the frontier labs by years end

[cinder sphinx]

they need to come up with ai that can run on home devices, no API calls

i mean, we have that already but i mean quality models on-device

[dense magnet]

self improving ai will be here before 2027. total AGI still a bit far - a lot of paradoxes and ethical questions in the way lol

[frozen zealot]

there are many ways to skin @tepid kestrel , you can do continuous back propagation or event predition

[cinder sphinx]

i want to have Jarvis at home!

[frozen zealot]

or both, if you have the compute

[dense magnet]

thats what i am working on, small models specifically trained for specific tasks

[thorn thicket]

fine tuning or training from scratch?

[indigo mirage]

For now they have just increase prices

Gemini Flash 3 : 50 cts / 1M token
Gemini Flash 3.5 : 1.50$ / 1M token

[thorn thicket]

$1.50/1M token in or out?

[indigo mirage]

In

[cinder sphinx]

yeah, antigravity quotas were amazing when i started with it about a year ago. you couldnt hit limit at al. now it takes 2hrs or so to get stopped 🙁

[thorn thicket]

thats ridic lol

[thorn thicket]

I raise you 45 mins with my Codex lol

[indigo mirage]

I never hit limits

[cinder sphinx]

i use claude for serious tasks, gemini pro when claude gets killed, then i just run gemini 3 flash to do Ui things, it does great

[indigo mirage]

Never

[thorn thicket]

The Owl Alpha model on OpenRouter is very good. But I think better is going to be local models for costs/limits.