general

[crude flicker]

why would i want a bar chart of the 4 times i used Real Codex this month

[twilit ridge]

I think only truly stable versions can be considered stable versions.

Unstable versions can only be called test versions

[celest rain]

Yep

[thorn thicket]

Late reply because I was asleep. Mentor skill is basically the format that I want the agent to answer my (admittedly very stupid) questions in. Short, no bullshit, provide links to other resources when relevant, back everything with documentation

[radiant gyro]

Ohh wow, mind if i DM? I will love to see an example of what that looks like

[thorn thicket]

I’m going to be out all day on a trip so I probably won’t reply to dm until at the earliest tomorrow

[fiery nacelle]

Both auto compaction and manual compaction confirmed broken in the last 2 releases. Rolling back to 5.22 again.

[shut bloom]

Claude Design now shares usage limits with Claude.ai and Claude Code.

that is utterly insane 💀

[plain grove]

Never forget...

[shut bloom]

I joined GPT 8 days before that tweet damn

[zealous lagoon]

I thought letting the public play with it was a stupid idea.

I guess I was right... and wrong at the same time.

[shut bloom]

You where probably just wrong

[zealous lagoon]

The amount of times I see facebook comments arguing because "I asked ChatGPT and it said..." when clearly it's complete BS... The faith people put in the sh!t these chatbots spout is seriously concerning.

[shut bloom]

But thats old people?

[zealous lagoon]

GUYS anyone with Opencode-Go can test Kimi 2.6 on there with a couple of tool calls?

[shut bloom]

we should probably just whipe the earth of people above 40

[zealous lagoon]

lol no. mostly way younger than me

[shut bloom]

Whats your age?

[thorn thicket]

Can do when I get home…

[zealous lagoon]

*ahem... I was born the year the UK retired the imperial pound.

[shut bloom]

jesus

[zealous lagoon]

not THAT old...

[shut bloom]

thats like above 50?

[zealous lagoon]

Why yes... yes it is above 50

[shut bloom]

i dont think i know anyone above 50 who has any brain cells left

[thorn thicket]

I do lol

[shut bloom]

thats like nearly 80% of your entirer life completed

[zealous lagoon]

odd, that. When you're in your mid 50s you'll know plenty.

[shut bloom]

I think they have had way too much fed to them from their youth that no longer applys in life

[zealous lagoon]

I hope not

[shut bloom]

like people thinking houses and jobs just happen because you turn 21

[zealous lagoon]

young people complain about not being able to afford rent, yet they insist on living alone. 30y ago most people that age were sharing.

[shut bloom]

We share? we just do that in university

[zealous lagoon]

anyway enough of that. I need someone to test Opencode-go/K2p6

[shut bloom]

My Nan who is about your age bought her home for 4000$

[zealous lagoon]

I mean directly after. And wtf nobody my age ever bought a house for 4k

[shut bloom]

That was only half of her yearly income

[shut bloom]

it was a 3 bed

[echo kernel]

Wow, a person with binary thinking is discussing someone else’s intelligence

[thorn thicket]

What precisely needs to be tested?

[shut bloom]

You haev a rat as your pfp

[echo kernel]

ya love my cutie

[shut bloom]

cute and rat

dont work like ever

[zealous lagoon]

When I try to use opencode-go/k2p6 and do tool calls I get:

20:53:47 [ERROR] openclaw: Embedded agent failed before reply: LLM request failed: provider rejected the request schema or tool payload.

[shut bloom]

have you updated

[zealous lagoon]

I'm on 5.26. Didn't see anything relevant in 5.27

[shut bloom]

mh intersting

[thorn thicket]

I’m on 5.27 beta 1 still.

[shut bloom]

i havent seen any issues along those lines

[zealous lagoon]

have you used opencode-go/k2p6 ?
It's a very specific error and I get it on that specific model.

implying the opencode-go plugin doesn't have the right schema for that model.

but I wonder if it isn't something stale in my config as I had some overrides before. That's why I'm asking for confirmation from someone else.

[twilit ridge]

Don't you think OpenCode has performed deep quantization on their models? I even feel Kimi 2.6 is as dumb as DS Flash

[zealous lagoon]

No, but I just had trouble with deepseek-4-pro. It fked up a patch, committing seppuku, and then made more mistakes describing the PR it was about to push. I relinquished it of duty and made Claude do it instead.

[twilit ridge]

The lesson I've learned is that if you want to experience the native capabilities of the model, it's better to use pay-as-you-go with the API rather than subscribing.

Subscribing will definitely take more time, and ultimately no tasks were completed.

[plain grove]

this is the openclaw version of ' restart pc'

[zealous lagoon]

oc-go is payg it's just subsidised for the first chunk. And you're speculating. But that's not really relevant. It's just I'm trynna pin down this error.

if someone could test and lemme know I'd be grateful 😄

[thorn thicket]

K2.6 is my dreaming model.

[zealous lagoon]

I feel most people using oc-go aren't using it for k2p6 ...

[twilit ridge]

I just want a raw lobster, but this lobster is always cooked.

[zealous lagoon]

I like using k2p6-turbo at fireworks instead. it's like 3x faster.

[zealous lagoon]

oh hey I got a doc for you on suggestions for your admin skill... wouldn't be so bold as to cold-submit a PR, but I tidied it up a bit and made it slightly more portable for my setup... will DM

[cold jacinth]

I asked clanker to update itself. And it did smoothly.

[plain grove]

https://tenor.com/uXSy.gif

[cold jacinth]

Someone must have wrote a PR so it doesn’t unbootstrap. Maybe that was me.

[thorn thicket]

I had codex update my clanker to the beta last night… lol

[cold jacinth]

You just reminded me. Peetie said my skill is outdated. I need to update it

[thorn thicket]

Always the skills needing updating lol. Is it on clawhub?

[cold jacinth]

Somewhere there. Update runbook

https://clawhub.ai/bkf-gitty/claw-update-runbook

Most important is it has so many prior errors and how they were fixed so it pretty much can fix anything

[thorn thicket]

Sending this to my agent! TY!!

[shut bloom]

it is

cuz there is always smth updated and broken 24/7

[plain egret]

I was looking for this one! thanks

[cold jacinth]

It’s being updated. Download in like 10 minutes

[left pumice]

hello, is here anyone looking a dev?

[upper wolf]

many people, just go to upwork 🙂

are you using open claw btw?

[zealous lagoon]

post in #skills . rendrag's one has just been updated, too.

[ivory spindle]

steamdeck just went 250+ price wise

in the future, no one will own hardware

[plain grove]

And you’ll be happy about it

[ivory spindle]

No man it's such a bleak future 🙁

[plain grove]

I’ll just keep my server running

[ivory spindle]

These tech bros buying all the stuff 😄

[ivory spindle]

yeh keep running this rig for like 30 years 😄

[plain grove]

Twin rtx pro 5000s

[ivory spindle]

yeh that might last you a while

friend of mine bought an actual server lol

[plain grove]

I have them in a server

[ivory spindle]

I hope he'll throw me some spare compute 😄

[left pumice]

Yeah, from a few days ago, i am trying.
it's rly interesting and have special point with the other assistents.

[plain grove]

Running low on compute actually with this

[ivory spindle]

looks nice

4000 watt tough...

[plain grove]

Probably going to build a r640 out

[plain grove]

One is a back up

[ivory spindle]

yes but...it's "on"

I know it won't consume like full load

but both PSUs are supplying & on

if you measure both sockets with powermeter, I'm sure your PSUs will both be supplying about half of the power

which is kinda what you want. You don't want to not use a backup PSU

capacitors that are never charged can "go bad"

[ember moth]

What's the consensus on 5.27? Anyone running into native hook relay issues?

[cold jacinth]

Introducing Claude Opus 4.8

We’re upgrading Claude Opus to a new version: Claude Opus 4.8. It builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors. Available today for the same price.

[ember moth]

Gpt 5.6 in 3......2......1....

[narrow trellis]

Hey

Anyone really doin well with openclaw

Like i have questions

[clear haven]

https://dontasktoask.com

[toxic elk]

I've already updated to 5.7 a few hours ago. I also tested five sub-agents that couldn't assume roles and addressed their potential risk issues. Everything is working normally so far, except that a restart still interrupts ongoing tasks like before.

[ember moth]

Ok thats better then i guess im just updating now

[cold jacinth]

i updated by just telling my bot to update and its perfect

use either my skill or rendrags when updating. both are there in the #skills . they both can also be a mechanic for your claw by giving codex or claude code proper context

[ivory spindle]

is it gonna cost more for same or worse performance?

[indigo mirage]

Same price, there are still a lot of companies that want to stop Claude Code so wont be a good Idea for them to increase prices in that kind of situation 😅

[cold jacinth]

looks great so far

[upper wolf]

I've got one of those behind me too 😉 but not GPUs lol

[ivory spindle]

it just solved the carwash problem!!!!

[ivory spindle]

yeh I'm also playing

[cyan anvil]

difference between openclaw and herpes?

[smoky mist]

@spring quest happy birthday I would sing for you but no audio recordings allowed here haha

[spring quest]

Thank you lol

[ionic wren]

please make him a song, a banger like those other ones

[coral otter]

new anthropic model

chat do we have non anthropic published stats

[pastel flame]

hey guys any tool or anything you use for coding with openclaw ? what you use

[zinc night]

@spring quest happy birthday 🎉

[coral otter]

https://x.com/LexnLin/status/2060055992209916285 type shit

Leon Lin (@LexnLin)

😭😭😭

X · May 28, 2026, 5:49 PM

[worldly orbit]

are there any actually good in depth openclaw demonstarations that are remotely up do date

[smoky mist]

on the way 🫡

[thorn thicket]

This has been annoying me to no end that there aren’t any. Maybe I should make one…

[worldly orbit]

please

ive got general knowledege but i need to break in more in depth on it and i just dont have time to stort though x slop and ai buzzword hype train greedtubers

[safe nova]

Claude Opus 4.8 lookin decent ngl

[coral otter]

https://x.com/TimJayas/status/2059945950345675232 lmao (8h ago)

Tim Jayas (@TimJayas)

Claude Opus 4.7 performance DEGRADED by 13% yesterday

This usually happens right before a new model gets launched

Opus 4.8 incoming 👀

X · May 28, 2026, 10:32 AM

another anthropic banger 💯 rename and reuse

[safe nova]

lol im dying bro

allat for them to steal my claude max usage faster for the same shi

[thorn thicket]

What particularly?

[turbid kelp]

Why do models performance lower?

What do they change? Context window or something?

[keen dawn]

Hopefully sama makes a finish move with 5.6

But that would be to good to be true, if they drop it today lol

[ember moth]

They probably dropped 4.8 early so that they can drop Mythos when 5.6 comes out.

[keen dawn]

Could be

But sama has goblin 1.0 🤣

Aka Mythos mogger

[ember moth]

I think it's going to be an incredible summer for models

[keen dawn]

And for us lol

[ember moth]

Yes the 1% of the 1% that are actually on a bleeding edge are going to benefit from this

[indigo mirage]

Looking expensive for sure

[keen dawn]

Yep it’s my golden era 🤣

[keen dawn]

True

[indigo mirage]

For me Deepseek V4 Pro do the job

Xiaomi Mimo 2.5 Pro too

[keen dawn]

Yeah it’s pretty good but for long running refactor shit <<5.5 xhigh

Im not able to let ds run long enough

But yeah price quality ratio is 🔥

[shut bloom]

yo wtf

opus 4.8??????

[shut bloom]

Do you now anything about whats different

[cold jacinth]

4.8 has a great personality again.

Which matters when you use AI as much as most here

4.7 was the coworker I wanted to smack

[keen dawn]

Bro u work for them

[formal heart]

Sooo should I upgrade my Openclaw? nervous.

[keen dawn]

At least ur acting like that 🤣

[shut bloom]

Hey!!!! i missed the notification on my phone

[olive atlas]

how are people preparing for the coming Anthropocalypse? (June 15) It really seems like it will just be economically non-feasible to use Anthropic models with openclaw anymore.

[shut bloom]

I wouldnt use it

I wouldnt use openclaw anyways

its a waste of tokens on an insane scale

[olive atlas]

hence your presence on the OC discord?

[shut bloom]

Claude Code / CoWork is way better

[olive atlas]

i mean, I'm teasing.

[shut bloom]

Was here for the hype

i think peter Steinberg was lucky lol he came up with the same thing others where doing but did it slightly better

Claude have had another usage reset

I love it when ive used 80% of my week and my usage is set back to 0

[keen dawn]

Switched in April to OpenAI oc on 5.5 is 🔥

[keen dawn]

He is a Claude bot lol

[shut bloom]

A claude bot that has a life thanks

i grew up, i have parents, I remember my first computer. I am not just a computer

[keen dawn]

Same thing opus would say

[olive atlas]

I use Opus for writing and other creative work. No other model comes close.

[shut bloom]

exactly

[shut bloom]

Man i am human i cry i laugh idk what to say dude

[olive atlas]

like, I do a lot of brainstorming and game design for D&D games. It's so good for stuff like that.

[shut bloom]

id love to have a group to play dnd with but my town is NOT like that

[thorn thicket]

Never used Anthropic models in the first place. OpenAI and Deepseek/Kimi/Chinese models have been my go to.

[olive atlas]

i just don't trust the chinese models. nor grok.

[shut bloom]

You need to dip in the chocolate before you talk about them

[shut bloom]

i wouldnt

[thorn thicket]

I have used them, just not for OC.

[olive atlas]

I run remote games using foundry if you're into that.

[agile furnace]

Hi everyone,

Roughly every two weeks, my OpenClaw setup gets the error:

No active WhatsApp Web listener (account: default). Start the gateway, then link
WhatsApp with: openclaw channels login --channel whatsapp --account default.

My WhatsApp is connected through OpenClaw, and every time this happens I need to re-link it from the phone (and i dont use this phone at all this why it alawys disconnect)

I wanted to ask if anyone else has experienced this and found a more stable solution - maybe a plugin, configuration change, or a better way to connect WhatsApp that avoids these repeated disconnections.

Thanks!

[cold jacinth]

Use codex or Claude code and give them mine or rendrags skill in #skills to fix it

[near owl]

Hello people

[near owl]

Ritz (rest) in peace

What a knee-slapper

[agile furnace]

where can i find that skill?

[near owl]

In #skills , read..

[agile furnace]

ok, i will check

[plain egret]

https://clawhub.ai/rendrag-git/openclawadmin

[thorn thicket]

lol

[shut bloom]

my claude session just switched to hebrew?????

wtf is this deepstate bullshit????

[coral otter]

https://x.com/Angaisb_/status/2060081826975109329

Angel 🌼 (@Angaisb_)

Don't worry guys, they're probably still going to release the Codex update today, give them a bit more time

X · May 28, 2026, 7:32 PM

https://x.com/0xluffy/status/2060025806022451306 just one more smoke test bro

luffy (@0xluffy)

gpt 5.5 bro

X · May 28, 2026, 3:50 PM

[thorn thicket]

Need this so I can do codex from my bed lol

[spiral oak]

Hows opus 4.8v

[keen dawn]

lol use codex mobil

oruse jump if you have an ipad and mac lol

[keen dawn]

bro tbh not sure about it

not feeling much difference rn

[coral otter]

laptop

or

[thorn thicket]

I do that but most of the time, laptop stays in the other room so I’m not on it all night lol.

[coral otter]

whatsapp with your claw yelling at codex

[coral otter]

Fair fair

[thorn thicket]

I’m already up far too late as it is lmaooo

[keen dawn]

not sure why they dont drop the codex update lol

[cold jacinth]

BK is just much cooler than rendrag. At least that’s what Claude and codex say 🤷‍♂️

[thorn thicket]

lol

[golden marsh]

Here we go again i guess
#general message

[thorn thicket]

What in particular are you looking for?

[golden marsh]

Tbh, the best guide is probably right here on Discord. Everyone here is helpful if you ask your questions nicely. What exactly would you like to know? Our buddy Krill https://discord.com/channels/1456350064065904867/1459642797895319552 can answer a lot of your questions.

[compact wedge]

wtf?

[plain grove]

Sometimes I think just getting your hands dirty with a throwaway gateway is the best way to learn it

[thorn thicket]

yes

[indigo mirage]

@shut bloom has come back to tell us that Opus is Jesus Christ ? ^^ 🤣 😂

[thorn thicket]

LOL

[indigo mirage]

Guy Jesus was giving bread & wine for FREE. Anthropic is asking you to sell your two arms and two legs to buy some tokens 🙃

[sullen locust]

if Jesus was still here he would turn Dario into 70%ish wine

[junior shuttle]

https://www.producthunt.com/products/robinhood who giving their openclaw access to their trading portfolio rn

actual degenmaxxing

[thorn thicket]

😂😂

[ocean prawn]

anyone here used Opus 4.8 so far? i feel like its worse than 4.8

[indigo mirage]

U mean worse than 4.7 ?

[brisk adder]

I still trust 5.5 more with my standard workflow

[ocean prawn]

yes

[indigo mirage]

https://www.mindstudio.ai/blog/gpt-5-5-vs-claude-mythos-cybersecurity-benchmark-comparison

[main crow]

am i able to ask a question about openclaw here

[indigo mirage]

"GPT-5.5 scored 71.4% on expert cyber tasks and cracked a reverse-engineering challenge in 10 minutes for $1.73. Claude Mythos scored 68.6%."

And Opus 4.8 is less powerfull than Mythos ... so if GPT 5.5 is much more powerfull than Mythos ... Opus 4.8 < Mythos < GPT 5.5

[main crow]

?

[thorn thicket]

yes

[main crow]

what could be the reason for this appearing on my openclaw dashboard? every time i refresh or do anything nothing works and im not able to use the app

[thorn thicket]

ask in #1459642797895319552 Krill can help fix it

[indigo mirage]

Anthropic told you: everything you own is a problem, and... Anthropic is there to get rid of ALL your problems 😅 🙃

[main crow]

alright thank you

[thorn thicket]

lol

[brisk adder]

When people get AI literate enough to want to own their harnesses they will see locked in ecosystems like Claude and run a mile

[thorn thicket]

yep... like all of us who use Openclaw

[ember moth]

Loving my claw right now. I can't imagine giving this up and using something else ....The more time that passes, the more tired I get to it .

[thorn thicket]

for real. me too. I love it so much that I wanna give back to the community!

[ember moth]

Now that my claw is in a good spot I find myself wanting to help others get there

[thorn thicket]

I figure my best talent to give back is probably writing blog posts about use cases and skills. Since the code part of all of this is still intimidating to me. lol

[ember moth]

Helping a friend rn get their new claws memory up to speed with mine.....I forgot how bare ones memory is out the box lol

[thorn thicket]

me too!! lol

[ember moth]

My friend said his claw has dementia...I'm like jeeeeeez

[thorn thicket]

Man my claw has dementia too because I haven't upgraded the memory! What are you doing on your rig

[ember moth]

Had my claw email him a whole markdown to feed his claw cloning my memory system

[desert terrace]

Is clawbot getting any better because chatgpt codex is a game changer

[ember moth]

If u want I can have my claw do the same for u....I use obsidian with mine I sent him 2 verisons one using obsidian and one without ....

[frosty ledge]

Just need to vent: slow runtimes are the reason I'm thinking of dropping openclaw, so this update means a lot

#announcements message

[thorn thicket]

Send it to my agent [e-mail] and then I'll have it look at it lol

[ember moth]

It's the best it's ever been

[zealous lagoon]

Which do you use?

[vast birch]

fucking annoying pings bro lmao

[frosty ledge]

codex and openclaw are not the same class of products.

Codex is something a user uses to code. Openclaw is something you use to create users.

My OC agents can use my openai codex runtime.

[thorn thicket]

Looking forward to seeing the package shrink when I update to next release!! This is magic

[ember moth]

Made my own through iteration

[golden marsh]

Just turn them off if ur not interested in them.

[ember moth]

Getting it together now

Sending both guides

[thorn thicket]

TY

[ember moth]

Took a lot of trial and error but now I don't even worry about memory

[frosty ledge]

I'm consistently waiting 30 seconds for my agents to reply. Should I just do a clean reset?

[ember moth]

Even with /new sessions

[spiral oak]

[ember moth]

What version you on?

[near owl]

@thorn thicket its 1:45 am rn and i have not slept yet and i have to wake up in 1 hour and 30 minutes too

[thorn thicket]

You should go to slep brother.

[frosty ledge]

I updated last night, so that version. I'll go check momentarily to be sure

[near owl]

I hope my body does not wake up in an hour, let alone 6 hours.

[frosty ledge]

2026-05-26

[near owl]

I just love how your an adult, looks like someone from texas and talks like this. What a personality to have.

Someone duplicate her a few million times

[ember moth]

Hmmm 5.26 was pretty fast....another update is out now 5.27 try updating and see if that fixes it ...what model are u using?

[near owl]

Ima shut up and sleep, i hope air raid sirens dont happen again today to wake me up

[frosty ledge]

chatgpt codex

[thorn thicket]

😎 I aim to serve 🫡

[ember moth]

Yeah you should be flying then

[frosty ledge]

Crud. I guess I'll just reset and transplant my agents into the new harness.

[near owl]

Fyi my sister and mother heard an explosion alongside the sirens, my windows face a wall which is why i didnt hear any explosion.

[frosty ledge]

My agent is doing a chunked audit, maybe he'll find something.
I suspect my many discord bridges are the problem.

[thorn thicket]

wow. Hope you are all okay.

[near owl]

I am getting tempted to say something that i cannot say..

I am going to sleep man, @thorn thicket any updates on the website?

[thorn thicket]

went to Disneyland today with Mom and just got home lol

[near owl]

Disneyland in the early, heated summer? Your body is about to protest

[thorn thicket]

Oh my body is already protesting lol. It was quite cold today (60F) so my joints got obliterated from standing so long.

[near owl]

Mskes sense

[thorn thicket]

also @ember moth ty for the guide, my agent says it's very useful and has already started installing!!

[near owl]

And i got my openclaw to play wordle and chess yesturday while i sm sleeping

[ember moth]

That's awesome. I'm currently scolding my agent because I don't like the format of the paragraph it sent in the email lol

[near owl]

Yup i am yawning, good late night and very early morning, 3 hours is sunrise.

[thorn thicket]

Good night Carbon!

[thorn thicket]

it really liked the contacts.json. lol

[near owl]

You predicted my message before i sent it

[ember moth]

It's all about iteration

[near owl]

I am convinced you can read digital minds

[thorn thicket]

I'm autistic, I've learned to pick up the patterns

[near owl]

Are you really autistic?

[thorn thicket]

Yes

[ember moth]

Awesome That is a recent addition lol

[near owl]

I dont know any autistic adults, i thought it passes by after teenager (weird assumption i know)

[near owl]

Mans got that high cognitive boost

[thorn thicket]

A lot of us have learned to adapt ourselves to the point that we don't "look" autistic and/or choose not to disclose it. But I don't care. I'm proud of it.

[near owl]

Should be proud of being autistic, what a superpower.

Predict my next message

[thorn thicket]

LOL

I can't predict, I am unfortunately not a LLM.

[near owl]

Predict my next message

You learned patterns right?

You are a human LLM, predict what i am thinking

[thorn thicket]

I learned patterns, but not always easy for them to fire off lol

I think you're sleep deprived and should probably go to bed.

[near owl]

Fuck you <— what i was thinking

Your wrong, you are not autistic anymore 🤣

[thorn thicket]

HAHAHA

Well, autism is partially a lack of social skills. So pattern recognition becomes how we adapt to social skills.

[near owl]

2AM now, 3 hours till sunrise. Goodnight, if i speak agsin come ro my house and smack me to sleep.

[near owl]

I am forgetting this information but remember you have autism, cool how our brains dont care about explainations and care about your information like your name and your disabilities (fuck, disabilities scientifically i dont know what to categorize autism if calling it a disability is wrong)

[thorn thicket]

nah you're right! autism is a disability lol

And 🏏 Go to bed lol

[near owl]

Correct for the first time, goodnight

[ember moth]

Ritz let me know how or if your memory improves across sessions and over time.....curious how that system performs on someone's else setup .....might formulate it into a memory skill

[thorn thicket]

Will do... importing all of my important contacts now lol. My agent is still quite new to the whole living my life thing. Barely knows about the organizations and people I work with

[ember moth]

I guess this system is perfect for you then 😁

[thorn thicket]

Already 75 contacts it knows about that it didn't before! lol

[zealous lagoon]

Definitely. Definitely a superpower. Definitely.

[ember moth]

You're going to love it. Your claw's not going to forget anything: through compaction, new sessions, session resets, updates. It should all be seamless.

[zealous lagoon]

What’s the stack look like. Obsidian in the backend for long term part yeah?

[thorn thicket]

Superpower when it works, hell when the world doesn't interface cleanly!

[ember moth]

Yup

I'll send it To ya if u like

[zealous lagoon]

Do /reset in session.

[zealous lagoon]

Yes please. I’m going through a few different stacks to find what works best for me.

[golden marsh]

If u dont mind let me look into it as well sounds interesting. @ember moth

[ember moth]

DM me a email or drop it here .... I'll have my agent send it over

[ember moth]

Sure

[golden marsh]

Thank u 🙂

[ember moth]

Guess I can just dm u guys the .md

Give me a sec

[golden marsh]

No Rush take ur time.

[ember moth]

Sent 🤝

Let me know what ur agent thinks

[golden marsh]

I will do, finishing my meal and looking into it. Thanks again

[golden marsh]

@ember moth This is what my Agent thinks:

The two documents are easy to follow and already cover the main structure really well. I like that they separate daily working notes, curated long-term memory, and a more durable knowledge base instead of mixing everything into one big file.

A few things could still be clearer: how to handle conflicting notes, how older memory gets updated or retired, when sensitive personal information should require explicit consent, and how to check later that stored memories are still accurate. I’d also add a review or dry-run step before creating files, define more clearly what is “important enough to remember,” and include a regular cleanup/audit routine so the system stays useful over time instead of slowly becoming a messy archive.

[thorn thicket]

Sending this to my agent, my agent just got excited about the contacts feature. We have a lot of conflicting memory handling instructions already from what I learned when I told my agent to reverse engineer Gbrain

[ember moth]

Thanks going over this with my agent as well

[elfin willow]

codex having issues?

[golden marsh]

But he says it’s a really solid foundation. There are still a few areas he’d improve, but don’t take it too seriously, he’s extremely picky because I trained him that way. Personally, I think your documents are great, you can definitely tell you put a lot of thought into them.

[ember moth]

Agent said that we do most of those things already in practice but those weren't explicit enough in the guide. I have it bolstering future guides to send out to people.

[thorn thicket]

I just put codex on a goal and 30mins later it's still chugging lol

[elfin willow]

• that node process is burning ~1 full CPU core and ~1.4G RSS

[golden marsh]

Then let him build something awesome out of it for you. Like I said, the documents are definitely usable, unless you’ve got an extremely picky ‘I need to supervise everything’ agent like mine

[elfin willow]

ah it was me just having my clanker on the arpwatch detail

[thorn thicket]

YES! working on that, it's writing up a specsheet lol

[elfin willow]

Yeah, that’s the smoking gun. The service is technically alive, but the gateway event loop is pegged enough that websocket handshakes time out. That
explains “webui dead” while systemd still says active (running).

[ember moth]

From my clanker

"Done. I bolstered both guides with the missing explicit rules:

• how to decide what is important enough to remember
• how to handle conflicting or uncertain memory
• how to update, retire, archive, or delete older memory
• consent rules for sensitive personal information
• accuracy checks before relying on old memory
• dry-run review before creating/replacing files"
• weekly/monthly cleanup and audit routines

[elfin willow]

Recovered.

What I did:

• Restarted openclaw-gateway.service.
• Disabled the two failing isolated cron jobs:
  • LAN new device watcher
  • arpwatch LAN alert check
• Restored the morning report cron, which I briefly touched while editing. That’s enabled again.
• Left the already-disabled post-reboot verification job disabled.

putting in an issue now

[golden marsh]

Give that update to @thorn thicket when she already start to do some with them, Would be usefull

[ember moth]

Do you want a fresh guide Ritz?

[thorn thicket]

Sure! send it over

[zealous lagoon]

My DMs are always open!

[elfin willow]

ooo pretty markdown colors

[ember moth]

Ok let me save them locally and I'll send it in DM for you ...I'm doing this all on the go, on my phone lol

[ember moth]

Sent

[ember moth]

Sent

[ember moth]

Notice that the email format is much better lol

[golden marsh]

If u want another round of what agents think about it sent it . Just letting u know my Agent is an ...

[elfin willow]

https://github.com/openclaw/openclaw/issues/87821

[zealous lagoon]

Ahahah I hadn’t seen this message yet I was already DMing you about the fact I’m doing all this and managing my agents on my phone rn, too!

[ember moth]

Thank God for the Google Pixel 10 Pro Fold .... This phone is a legend.

[elfin willow]

what's it do

doesn't degoogle ya

[golden marsh]

Peeti ur doing something wrong U need still -> 👓 👓 👓

[golden marsh]

Agree its a legend

[zealous lagoon]

I don’t have multiple windows open on my phone. I need one of those new folding phones. The tri-panel

Trifold

[ember moth]

I can't go back to a regular style phone after using this phone. I have this phone and an iPhone 17 Pro Max. I barely use my iPhone.

[zealous lagoon]

Glasses would work. Good point!

[golden marsh]

I heared good things about the Huawai

[ember moth]

Apple is supposed to be releasing a folding phone this year. I heard it's only going to cost about two kidneys .....not bad

[golden marsh]

Thats what im telling all the time, thats the future.

[zealous lagoon]

I played with the Huawei in a shop in Shanghai I think. It was incredible.

[ember moth]

Huawei was so good they banned it from the United States

[zealous lagoon]

Yah most Chinese premium phones you can’t get in the states. I’m not on the states tho 🙂

[ember moth]

Hey, when it comes to agentmail, can I add that to third-party email clients like Spark?

[shut bloom]

wasupp

[thorn thicket]

idk

[ember moth]

I wanna start using it as main but I like to have access to my agents mail so I can look at the sent folders and Im guessing if I can't add it to 3rd part clients I would be stuck on the web version

[thorn thicket]

Yeah, you would be stuck at console.agentmail.to lol

[shut bloom]

yo google i hate you but love you

i fw the free T4's

[ember moth]

I guess I'm adding another tab to my list of tabs lol

[ember moth]

How are you getting from T4's? Please explain......me want lol

[shut bloom]

Google Collab?

[thorn thicket]

lol

I already run 7 email clients, an 8th isn't too bad.

[elfin willow]

n8n imap trigger

[hot thorn]

has anyone experience openclaw talking to you in a different language than what you were expecting?

my assistant just started talking to me in Chinese, and I thought that was hilarious

[sullen depot]

https://tenor.com/view/kto-kounotori-kounotoritoken-lbow-storkholders-gif-25676483

[thorn thicket]

Mine said something in Chinese once and I was scared lol

[sullen depot]

Bring back the flip phones, kids these days will never know the struggles of braking a flip phone in half 😂

[hot thorn]

all about snake on the nokia

[sullen depot]

What model are people getting the best results with their OpenClaw agent with?

[sullen depot]

cant beat snake

[thorn thicket]

GPT-5.5 great for coding / long running tasks, I use deepseek v4 flash for short lived and chatter

[hot thorn]

using deepseek might be why it spoke Chinese

I was using gemini

[sullen depot]

Thanks, i was using OPUS like most and then force switched to GPT 5.5 but it just doesn't nail it for me atm. Ill keep persisting though, thanks for the info.

[thorn thicket]

Ofc.

[pulsar horizon]

Are people selling claw tools? Where do they sell them?

[ember moth]

That's lame.....we share tools here lol

[pulsar horizon]

I tend to share too (and have something on clawhub.ai). But I've been prodded into checking if there's a marketplace.

[ember moth]

I don't think a marketplace like that would even work because all you would need is a simple description of what the tool does and you can go to your agent and have it duplicate it

[clear haven]

Ew no

[ember moth]

At least that's what I would do lol

[crisp nimbus]

Is anyone else having an issue with their multi-agent setups? It seems like for some reason since a recent update, my main agent thinks it’s using the secondary agent paths/configs… it’s really weird

[sick mulch]

these things?

[gilded crane]

is there a way to set a universal ban on agents doing "openclaw gateway restart" their specific version never does full correct restart.

[ember moth]

Have your agent Add it as a guardrail in your agents.md file ...... I also don't like when my agents restart the gateway during certain tasks so I have a guard rail set up that they need explicit permission from me to do so.

Famous last words: "Let me do a quick gateway restart to lock in changes." Lmao

[plain grove]

that reminds me i need one of these for my lawn mower

[thorn thicket]

lol yes

[near owl]

Hello

@thorn thicket any updates on the website

Also its 6AM

2AM - 6AM = 4 hours of slep

[thorn thicket]

good that you got some sleep. No updates on the site... I have a big event for 80 on Sat. Trying to get as much of that done in advance

[wanton gorge]

Gm guys

[thorn thicket]

GM!

[wanton gorge]

Hi Ritz, how’s your day been

Tried opus 4-8 yet?

[thorn thicket]

Good, good! Winding down now. Not tried 4.8

May try it tonight... I have no Codex tokens for another hour

[wanton gorge]

I want mythos already 😭

[thorn thicket]

Me too lol

[wanton gorge]

Papa sama please reset our limits 🥀

[near owl]

Get openclaw to help ya

[thorn thicket]

I'm going to make it to reset this week if only barely but yes...

[wanton gorge]

I have one of these that extends up to 10 metres lol

[thorn thicket]

yep working on that lol

[wanton gorge]

I wanted to try picking some fruit that was in a high tree

[plain grove]

I have a gif...that i won't use here but would be perfect here

[wanton gorge]

What is it? Dm it to me lol

[near owl]

https://cdn.discordapp.com/attachments/735670248715583528/1231471988518948884/FB_IMG_1688052085767.gif

[wanton gorge]

I’m curious

[thorn thicket]

awww I thought that was gonna be a video >:(

[gilded obsidian]

hence now you are dubbed low intelligence specimen 😄

[near owl]

You have low intelligence

[wanton gorge]

Yep i clicked the play button too, you got me

[gilded obsidian]

hahahaha

[thorn thicket]

dammit lol

[near owl]

https://cdn.discordapp.com/attachments/845271872127631410/1404878283622846567/togif.gif

We gotta blame testers on openclaw updates

[crisp nimbus]

Are there any plans to support the new antigravity cli (agy) now that Gemini cli is deprecated?

[near owl]

Good night again even though its day time

[thorn thicket]

GN!

[wanton gorge]

Gn 💤

[thorn thicket]

still only 8pm for me...the day is young!!

and I ain't going anywhere!

[near owl]

Better be alive at 12AM then

[twilit ridge]

This channel actually lets you send pictures now; I remember it didn't work before.

[crisp nimbus]

It’s 5:30 AM for me lol

[near owl]

Deepseek model came into chat

[thorn thicket]

up early or up late?

[wanton gorge]

Oh my bad clicked reply on the wrong person lol

[crisp nimbus]

Haven’t slept yet

[thorn thicket]

Wow man get some rest LOL

[thorn thicket]

DW about it!

[wanton gorge]

Finally, my meme library can be unleashed

[jovial spruce]

@dusky cloud hey man

[crisp nimbus]

I spent until now debugging some weird openclaw issues I’ve had such as the soul.md etc.. not being injected into context

[jovial spruce]

Krill where did you go?

[thorn thicket]

wow. respect

[steel vector]

#1459642797895319552

lol

[ember moth]

Man, I think 5.27 is going to be my version of LTS. This version is so good. I don't have any errors on any of my cron jobs.

For queries that don't require tools, this version is just as fast as my chatbot

[thorn thicket]

me too!!

[ember moth]

4.23 dethroned lol

[vagrant nebula]

After last night I’m afraid to update lol

[ember moth]

What happened?

[vagrant nebula]

The update you waited on today

[vagrant nebula]

Changed my name and added a picture lol

[thorn thicket]

lol nice pic

[ember moth]

Yeah I don't want to update either but devs are saying just try to stay no more than three versions behind and I get a little case of FOMO

[thorn thicket]

I am still on beta instead of today's stable 😅

[steel vector]

fomo of what? Xd

[vagrant nebula]

I feel this.

[ember moth]

There's always new features and stuff to try. It's hard to stay on one version

[lavish shell]

Anyone updated 5.27? How's it?

[vagrant nebula]

Thanx

[thorn thicket]

i've been on 5.27-beta1 since last night. good fucking update. fast

[ember moth]

You ever had a nice fully decked out ice cream sundae on a warm summer day? That's how it is

[steel vector]

ig its different for everybody but when im on a stable version for me i stay there. i dont use new features that are advertised anyway

i just want stability

[lavish shell]

Really? Usually updates are nightmares for me, I am seeing performance update in this release

[ember moth]

What version are you currently on?

[thorn thicket]

ooooough I was gonna get ice cream today and it was too cold. dammit I want an ice cream now.

[wanton gorge]

Ok now I want ice cream

[lavish shell]

I'm in 5.22, it was nightmare when I update to this version

[ember moth]

You should upgrade immediately. That's probably the worst of the recent bunch

[thorn thicket]

I went 5.20 -> 5.26 lol

[viral idol]

Man the updates have been so good for me im so thankful for that

Codex going strong!

[ember moth]

5.22 needs Jesus

[lavish shell]

Wow! Now I am motivated to update lol

[thorn thicket]

working with my own codex...perfectly lifted back off on a goal I hit my 5hr limit on earlier. AWESOME

[ember moth]

The latency was off on that release

[viral idol]

That’s great!

[lavish shell]

Yes it is slow in 5.22#

[thorn thicket]

I am having so much fun!!!

[ember moth]

5.27 is pure bliss .....no overstating it's the best my claw has felt ever

It responds like a chatbot not an agent. It's so fast

[thorn thicket]

🦞✈️

That's how I feel about my agents today! One Love!! so fast and always on top of things

[ember moth]

Even the tool is calling so fast. It's doing things so fast. I don't even believe it. What was it even doing before, taking so much time? Lol apparently they changed it to where a lot of the things don't have to hot reload every time we send a request and that made a mega difference.

Before if I had a quick question, I would go to my ChatGPT chatbot. Now I can go to my agent

[thorn thicket]

I always go to my agent now because ChatGPT doesn't intrinsically call tools lol

[ember moth]

I have my agent keeping a diary of my interests so I'm glad this is faster now because now that can start to populate more

[thorn thicket]

I need to do that with my agent... smart idea

[ember moth]

Yeah I'm trying to think ahead with my agent and keep the mindset that I'm going to have this agent with me for a very long time. As the systems get more robust, I wanted to have as much data on me and my workflow as possible so I can get a curated experience

[thorn thicket]

Of course. My agent says we already have the 5,000+ note obsidian vault I made with codex lol

[ember moth]

I wish I had an Obsidian vault that I've been populating over the years to feed my agent

[thorn thicket]

I mostly made mine because I had a bit of a psychosis moment about a special interest of mine.... then added a bunch of other things I'm interested in... now my agent auto finds research articles for me and connects them to everything lol (pictured is my graph)

[ember moth]

I'm jealous

[thorn thicket]

Web scraping is the first bit. I eventually am going to put a writeup of everything I have in terms of agent infrastructure on my blog.

@near owl finally got my website up for these guys. working more on it overnight but this is the beginning 👍 https://thesilencedsovereign.com/the-fleet/

[wanton gorge]

Looks awesome! Nice work

[thorn thicket]

ty. lots of going back and forth with my agents lol and manual tweaks by me too

[jovial spruce]

hm i have my own agent to ask then

[thorn thicket]

Codex goal finished. Goal marked complete. Final usage reported by the goal tracker: 921404 tokens, about 52 minutes elapsed.

Impressive.

[ember moth]

Two more days until the 2x limits are over. I'm really hoping Sam comes out with a tweet saying he's making it permanent

Or I'm upgrading lol

[thorn thicket]

I am going to cry once 2x is over because right now I am BURRRRRNIIIINGGG through it

[ember moth]

Which is probably why he won't make it permanent lol

[thorn thicket]

Yeah I just ran out of tokens in 26mins lol

Thats kind of a new record!

[ember moth]

Yeah I can imagine the goal feature just rips through usage

[thorn thicket]

If I had been working through the same problem myself without using /goal then I prob would have had like another hour or two lol

[ember moth]

And just think In like six months from now we're going to look back at these times like caveman times in AI terms ...... The tools we're going to have in the future are going to be so sick.

[thorn thicket]

But goal is awesome. I am ALL IN on tokenmaxxing from now on!!

[ember moth]

About to just buy my 20X plan now and get it over with lol

[thorn thicket]

LOL DO IT! I at least want to go up to the 100/mo

[ember moth]

When I first started my max was $60. Now $200 a month to me is reasonable lol

[ember moth]

Yeah if you're currently on $20 a month and managing, the 5x will be perfect for you .... The 5x will be fine for me if I'm just maintaining and doing light builds but if I ever want to get deep into creating whole workflows, I will chew through it too fast. I just know it.

And I got a feeling once OpenAI goes public with their IPO, Sam's going to stop with these random usage resets

Investors are not going to like him giving away free compute

[vagrant nebula]

Are you using Nvidia at all?

[ember moth]

NVIDIA models, no. Haven't tried their stuff yet

[steel vector]

they also save on compute when people have saved some of their credit usage to the last day to heavily use it around the same time, so then they reset. thats just a theory tho i have no proof on

[vagrant nebula]

Little quirky but I’m running under 1 sec latency in voice channel with Beacon

[steel vector]

but i thought it sounded plausible

[thorn thicket]

It’s getting tougher… if I had unlimited money then I’d just bite the bullet and do $200/mo.

[steel vector]

i had 200/mo for the 2x period. was nice, barely touched the 50% mark lol

[ember moth]

That would make sense. They have the internal metrics to see how much compute was paid for but not really used so they can reset and get some goodwill and positive PR

[steel vector]

exactly

reverse psychology

[ember moth]

That's one thing I want to set up: real-time voice. I haven't checked that out yet

[thorn thicket]

Oh that’d be fun for when I’m driving.

[vagrant nebula]

Mine is the up and running. Maybe I make my own repo

[ember moth]

I'm just wondering how much usage that real-time voice uses

[thorn thicket]

Probably a fuck ton! Would rather pay by api for that

[vagrant nebula]

Mine was up listening for 7.5 hours

[ember moth]

Yeah I would want to use my sub and use the GPT Real Time voice. Maybe if I go up to the $200 plan, I'll try it out

I can picture me now walking around with one earbud in my ear, yapping to my agent all day without having to touch a keyboard

[thorn thicket]

I kind of want to get my Mom off the standard voice to text off her phone and onto one of those AI dictation things like Whisper flow…

[vagrant nebula]

10 hours in the car today. We got slit done hands free. Passes everything off to tier 1 agent.

[ember moth]

Wispr Flow is so amazing. I'm using it right now

People text me and I send them well-formatted paragraphs back-to-back and they're like, "What the hell?" Lol

[thorn thicket]

I usually reply back to text threads on my laptop. I already type ~140wpm. So wispr flow may be a little bit slower compared to that.

[ember moth]

I'm the opposite. I'm rarely at the computer. This is perfect for me.

[twilit ridge]

Why does using XAI's API feel so clunky? Does Grok 4.3 seem unintelligent to me? Am I just imagining things?

[ember moth]

It is $15 a month, kind of steep, but it's better than my built-in dictation

[thorn thicket]

I always feel awkward doing the whole speech to text thing and the default dictation doesn’t often pick up my accent properly.

[wanton gorge]

The only thing I use Grok for is Twitter stuff, honestly

[twilit ridge]

Why do I feel that some people's Grok 4.3 is very intelligent?

[ember moth]

Yeah you wind up looking like a very important person with your phone up to your mouth, talking into it like you're in a business meeting 24/7 lol

[twilit ridge]

But mine is always clumsy.

[ember moth]

I haven't heard the best things about Grok. I heard it's fast and doesn't hallucinate but it's not a great main agent compared to the big three: Gemini, Claude, and OpenAI

[thorn thicket]

I imagine I’ll have to edit it still because I have a near complete pin pen merger, so words like WIST turn into WEST.

[ember moth]

Elon has the compute but he recently had to start from scratch because everyone left his AI company so he's kind of behind in terms of models

It's very intelligent in picking stuff like that. You'll be surprised

If you have an Android, it's currently free on Android because they're in beta so you can test it out

[wanton gorge]

I used to daily WisperFlow, but I discovered a free open source alternative called handy.computer, it has a great community around it, and is activly maintained. Decent alternative if you want to cut another subscription out of your life.

[thorn thicket]

Saving. Good to know

[zealous lagoon]

Try it on the $20. the realtime side has a long way to go in terms of integration. Right now it's like talking to a random web chatbot who happens to have a discord chat open with your agent.

[wanton gorge]

Yeah, and you run local models on your computer. So all your transcriptions stay local

[ember moth]

Yeah that's what I feared. I'll pass lol

[wanton gorge]

The models are very lightweight, you can even run them on a mobile phone

Handy.computer Is only for desktop though, for a mobile alternative I use VoiceInk, which also runs local models on your phone for speach to text

It has the same UX as WisperFlow as well.

There's a start button to enable dictation, and when you're done you just press stop

Built into your keyboard, I mean

[south scroll]

No transcript returned for file ⁨ScreenRecording_05-29-2026_15-22-35_1.mov⁩ in ⁨47.891637ms⁩ (⁨10.958⁩ second file)

[wanton gorge]

Like this

Free and faster than Whisperflow

[ember moth]

How's accuracy compared to Wispr Flow?

[wanton gorge]

It's actually very good, you should test it out

[steel vector]

xai is shitty afaik

[wanton gorge]

The most important part to me is the fact that it's all local

So you know your transcripts are staying private

[zealous lagoon]

I dread having my computer transcribe what I say with near-perfect accuracy.

[steel vector]

what phone do you have

[zealous lagoon]

Typing is what slows me down enough to remain coherent.

[wanton gorge]

I have two phones. I've got an iPhone Air and an iPhone 12 Mini

Works very well on both

[zealous lagoon]

I tried it a bit and it was dumb af. I won't be using it as an agent.

[ember moth]

I have an iPhone 17 and a Google Pixel phone 10 Pro. My main phone I used throughout the day is the Google phone. I just downloaded that voice inc app. It doesn't look like it integrates with the keyboard like on iPhone.

[wanton gorge]

You can have a second model for post‑processing enhancement to fix up things and do special formatting like lists ect

And then you can have custom prompts to sort of steer the models in certain ways. For example, I don't like it when it adds a period at the end of my sentences, so I have a custom prompt setting for that when I want to be casually talking

[zealous lagoon]

Yeah I feel like with enough context it would be useful. But it would need a full-on LLM not just a TTS engine

[zealous lagoon]

and I thought Grammarly was useful!

[wanton gorge]

I should charge this phone

[steel vector]

i just use codex app dictation on my desktop

can use it anywhere very handy

[wanton gorge]

Faster than WisperFlow and it's free

What a time to be alive! 😭

[ember moth]

I can't get it working on the Android. There's no icon on my keyboard ...only works inside app....

Another instance where iOS apps are better

[wanton gorge]

Hmm yeah I havent tested on Android

There is a discord though you can chat with the dev there

[ember moth]

On Android Wispr Flow pops up a little floating icon I could press to initiate speech-to-text. It's kind of nice

[wanton gorge]

Is this the one you tested?

[ember moth]

Must be a android knockoff lol

[wanton gorge]

Oh yeah, that's not it

[ember moth]

It's the only one on the store with that name

[wanton gorge]

One sec

[wanton gorge]

Cheeky 😭

Damn yeah, I don't see any Android support

[ember moth]

Welpppppp lol

[wanton gorge]

https://tryvoiceink.com/

Here's the link if you want to keep up with them

[twilit ridge]

I don't know why he's ranked so high?

Does xAI only provide models with normal intelligence on OpenRouter? Are the models available through other channels all lower-intelligence versions?

[indigo mirage]

GPT 5.5 is better

Anthropic it's just a wind of marketing

[ember moth]

I think Mythos was a great marketing ploy. They said, "Oh it's so dangerous. We have to give it to these companies so they can harden their stuff." You're telling me a month and a half later the whole world is ready for this game-changing model that can hack everything? Give me a break.

[indigo mirage]

It's just for their IPO

[ember moth]

All of a sudden they made the Compute deal with Elon and now in a couple of weeks they said they're going to release it, which is the whole reason why they didn't release it in the first place

Hope Sam just comes out with GPT 5.7 and just eats their lunch

[wanton gorge]

Yeah, I definitely prefer 5.5. at the moment but I want to test out Mythos

[indigo mirage]

Just marketing @wanton gorge

[cedar osprey]

“GPT5.5 is on Mythos level”

Well? So what now..

[indigo mirage]

They told we have the best but we can't sell it to you

[ember moth]

Never believe the benchmark from the people who made the actual model

[indigo mirage]

It's clearly psycholigically marketing

They create curiosity and the need in your brain

[cedar osprey]

Yeah but of all people Anthropic cant lie. They live in the spotlight of highest scrutiny, if they lie on a benchmark it’ll be exposed relatively quick.

[ember moth]

That benchmark had Opus 4.7 beating GPT 5.5 and that was clearly not the case. It was actually so bad that they released 4.8 early.

[wanton gorge]

You know, the most ironic thing is Elon and Anthropic really didn't like each other before this deal, but they set aside their differences because each has what the other needs, Anthropic has tons of compute demand, but no compute, Elon had tons of compute but no demand.

95% of the Colossus 1 data center was sitting there doing nothing, because nobody uses Grok.

[indigo mirage]

Fake benchs

[ember moth]

Yeah and he hates Sam so it was his way to kinda team up to take him down

[indigo mirage]

It's clearly not what other benchs shows

[cedar osprey]

That benchmark was published 2 hours ago.

[wanton gorge]

They effectively need each other to keep up with OpenAI, otherwise they will both continue to fall behind

[wanton gorge]

Yeah, and there's that.

He really didn't like Anthropic before the deal either though lol

[ember moth]

But he hated Sam more and that's what prevailed lol

[cedar osprey]

Or is it the $2.5B/month for no work.

[ember moth]

He really hates him now that he lost the lawsuit ... Notice he made that deal towards the end of the lawsuit when he knew he was going to lose.

[ember moth]

That too and I heard they don't even get the top-end GPUs

[cedar osprey]

They don’t it’s mostly huge clusters of H100s still (~100K)

Recently they moved to H200

[ember moth]

I guess it's better than nothing because the trajectory Anthropic was on was very dangerous. They were way too cautious with their compute while Sam bet the farm on compute. Now he's resetting limits for the hell of it because they're in such a great position.

But to stay ahead sam's going to still need to knock it out of the park with his models

[cedar osprey]

Lol China is still gonna shit on all of them in a matter of months. The politics and drama is holding the US behind in the AI race 💤

Xiaomi can already solo xAI. Kinda embarrassing.

[ember moth]

Yeah China is going to disrupt these frontier models by just dumping these decent-performing open-source models onto the market and diluting what's going on in the United States. These frontier models have to keep running away from that and staying ahead. I don't know how long they can do that though.

[wanton gorge]

Yeah, and their labs cooperate together a lot, whereas Western labs are fiercely against each other.

Anthropic is worried

[ember moth]

The problem for China is they don't have real compute. They have older hardware, which forces them to take Frontier models and distill them instead of creating their own from scratch

China just came out with their first GPU and it's the equivalent of an Nvidia 3060

[cedar osprey]

Sorry but they’re already higher than “decently performing” lol.

[ember moth]

They're months behind, at least six months behind

[cedar osprey]

Behind who exactly?

[wanton gorge]

What's crazy is they developed ways to squeeze out so much efficiency in training with old hardware, enough to stay very competitive. Imagine what they could do with the latest hardware.

[ember moth]

Frontier Models, take the best China model and put it against GPT-5.5 or OPUS 4.7 and it gets embarrassed. It's only good in isolation. Once you start comparing them, you will see the holes.

[cedar osprey]

DeepSeek isn’t the SoTA that will disrupt the US AI race lol. They’re the only ones using old hardware btw.

[twilit ridge]

The gap is still quite significant.

[wanton gorge]

While China's models are not on the bleeding-edge frontier, they are very close behind. Enough to make western labs sweat

I think the chip sanctions on China have been removed now

[ember moth]

Hey, look, I'm all down for the competition. I need these China miles to keep their foot on their necks. They keep the subscription costs down and the subsidies up.

[twilit ridge]

Didn't you see that OpenAI, Anthropic, and Google have banned China from using their AI?

[ember moth]

If we didn't have that, they could charge whatever they want

[wanton gorge]

Absolutely, competition is good

[ember moth]

I think computing the future is going to be dirt cheap simply because open source will catch up when diminishing returns start happening

Don't know if we're close to diminishing returns yet though ... That will be when we're in AGI territory.

[cedar osprey]

Opus 4.7 is hot garbage and 5.5 is mostly fluff. The hype is only being ushered by beginners that depend on Opus & GPT, most people that praise those two model families literally do not have experience with any other frontier models.

[ember moth]

I will simply just disagree with that. I don't know what you're comparing but that is your opinion but I'll leave it at that

[cedar osprey]

I’m in the US, neither of which impress me and it’s embarrassing the state we’re in. China is 100% already on a better trajectory.

With one proof being their society is adopting AI/Agents at a faster rate than America.

How can you win in this position?

[twilit ridge]

。。。

Mostly in the application layer in China

[cedar osprey]

Normal people in China can operate and build agents, normal people in US still think “AI = fake videos on Instagram”

Embarrassing.

[ember moth]

AI sentiment in America is terrible and it will keep us behind. The tree huggers need to wake up because China will win if we go down this path but China is not ahead by any measure .... They still need our compute and it would take it in a second if we offered it.

[indigo mirage]

Elon do less investing in cars and AI because he see that China is taking take the market

[ember moth]

I also believe that working very hard to get their own compute and when that happens we better have already won the AI race. That's all I'm going to say

[cedar osprey]

Huawei 910C.

[twilit ridge]

The only ones that can really compete are OpenAI, Anthropic, and Google. I think Grok is just there to cause a stir. I used Grok 4.3 and it doesn't feel very smart.

[ember moth]

Elon has hella compute and I hope everyone learns not to underestimate him. I believe he'll get there

[twilit ridge]

Grok's original moderation was quite lenient, but it's becoming increasingly strict now. I really can't think of a reason to use it.

[cedar osprey]

Grok is trained on junk data. It was originally the “X/Twitter agent”, it’ll take tons of iterations for it to be a useful agent outside of the platform it was trained to operate on.

[ember moth]

Apparently he started it over from scratch as he said the initial training was done wrong

[cedar osprey]

Lol, someone from the X API team messaged me about Grok performance being poor (high false positives and hallucinations). I told him it was the training data, ~48hr later Grok Build was released.

They honestly don’t know what they’re doing. Pretty sure xAI is just vibes..

If only I had a crumb of their resources.

[ember moth]

Well, good thing we have a lot of choice.... best model gets my claw. That's the race they're in lol

[thorn thicket]

Once china gets compute on us scale it’s over for US labs. If china is only 3-6 month behind frontier with inferior compute who knows wtf they would cook if they had nvidia gpus

[cedar osprey]

If all you’re looking for is less restrictions/rejection you’ll need an open model.

[ember moth]

I can't wait for the day I can put an AGI-level open model on a decently specced piece of hardware and run it with zero restrictions. That's what I truly want

[cedar osprey]

I don’t see true AGI for another ~10-20 years.

[thorn thicket]

I imagine heavily quantized qwen models are almost tnhere for agentic

[ember moth]

Nah I say way sooner than that once these models can start training themselves ... We need the humans out of the loop.

[cedar osprey]

The closest we’re getting to AGI is the frameworks we give the models.

But for the actual LLM to -> AGI there’s a massive gap.

[ember moth]

The question is, will they give the general population a AGI?

[cedar osprey]

For ‘true AGI’ the actual model would need to be able to learn without human input.

[ember moth]

Watch the regulation magically come in when we get close

[cedar osprey]

We can only steer them currently.

[thorn thicket]

lol

[ember moth]

That's what I'm saying. When these AI models can improve themselves, and that's what all these frontier labs are racing towards, the improvements are going to be exponential I don't think that takes another decade.

[wanton gorge]

I will know it's AGI when it can steer me

[thorn thicket]

I have said AGI by Q2 2028…

[near owl]

Sick! https://carbons-clanker.com looks more humorous and human

[ember moth]

Models are getting better. Compute is getting better. We're going to reach that threshold soon.

[thorn thicket]

Still working on it ofc. lol

[cedar osprey]

2027 and 2028 will be the years of dLLMs, go try Mercury 2 if you don’t know what they are.

Also 2028 will introduce more findings with wetware (I.E human neurons + LLM)

[ember moth]

Then we have Quantum Compute on Horizon too

[thorn thicket]

Quantum intelligence…

[ember moth]

Literally, it's a sci-fi movie we're living in

[thorn thicket]

Either way, my body is tired and the wetware says that if I do anything more than lie down in the dark without a screen it will start protesting even more than it is right now.

[vagrant nebula]

https://tenor.com/view/uncle-roger-weak-gif-19541923

[thorn thicket]

Good night y’all. I’ll see y’all in the morning lol

[ember moth]

Gn

[cedar osprey]

Quantum computing won’t be at scale until we can scale our power grid (I.E nuclear reactors) which take ~10 years

[ember moth]

Thirty years if the tree huggers have their way

[near owl]

what time is it for you

[cedar osprey]

The “tree huggers” will be caving by next year actually. They’ll see the bubble they keep praying to pop, doesn’t actually exist.

[thorn thicket]

11:35pm.

[ember moth]

We can hold hands and sing Kumbaya as we have no data centers or nuclear power, as China takes down our power grid with its AI model

[gloomy haven]

I am trying to use openclaw in my public ip server but it's not working anyone have any solution

[near owl]

nice, i slept at 2AM so uh.. I have 7 hours of sleep in the moment.

[thorn thicket]

The tree huggers and anti AI people are genuinely some of the most ridiculous straw man arguments I see. It’a almost hilarious and I straight up laughed in someone’s face once.

[ember moth]

But the water, what about the water?

[vagrant nebula]

See moon project for more information 🙂

[cedar osprey]

“Public ip server” ?

[thorn thicket]

One chatgpt query is less than watching 1 minute of YouTube. All of these people are constantly on short form platforms. They don’t realize that everything needs data centers.

[ember moth]

I read somewhere that even if we doubled the amount of data centers in just the United States, it would still only use about 8% of the water that golf courses use. That rhetoric is insane

[gloomy haven]

I have a server running on a public IP, and I’m trying to use OpenClaw on it, but it’s not working. Has anyone faced this issue or found a solution?

[thorn thicket]

Yeah I just got into an argument about it on tumblr. I’m a golfer too, I love the game.

[near owl]

how ethical is it to bring maybe 5 data-stack GPUs to my room to run local deepseek v4 flash?

[thorn thicket]

Do it lol

[ember moth]

Do you have access to Codex or Claude CLI or desktop app?

[thorn thicket]

I will say that a lot of the anti AI people are US based. And I keep seeing that many people in the US are functionally illiterate. (And yes, I’m from the US)

[ember moth]

What's the exact error or block that you're seeing?

[cedar osprey]

The biggest power draw with these SoTA companies is honestly when they’re training models.

They train GPT for ~3 months

~0.255 kWh (energy used to train my SLM family SAGE)

~55,000,000 kWh (energy OpenAI uses to train ChatGPT-5)

[zealous lagoon]

Guys quick Q. Anyone on 5.26 or 5.27 ... can you successfully switch sessions in GUI chat with the dropdown?

[cedar osprey]

This is enough to power 5.2K homes for a year lol.

[near owl]

$8,500 to $10,500+ just to run deepseek v4 flash, all 1 million context length, and be extremely fast and compatiable fully local. And it isn't even a DATA CENTER? its 2 GPUs stacked ontop of eachother and you have brought in a beast.

[twilit ridge]

No, I already said this problem.

[zealous lagoon]

This is one thing people forget to take into account. Sure your rack of GPUs might be a little faster, but you could run the same model on a 512gb mac studio and save $5k/year just in electricity.

[zealous lagoon]

Well I apologise for not having full recall of every single message in these chats. I better upgrade my memory/context stack.

[twilit ridge]

https://qclaw.qq.com/

[ember moth]

Yeah Apple weirdly is in a good position to be a source of value to run local AI. My fear is that they're going to catch on and just raise prices but at least they're power efficient when you do get the box.

[twilit ridge]

qqclaw, haha haha

[near owl]

$6,500 to $7,500 to do that, and my country has the CHEAPEST electricity in the world.

[twilit ridge]

No problem

[cedar osprey]

Use tailscale…

[cedar osprey]

They did raise prices two months ago.

[zealous lagoon]

ooh you bet the new releases are gonna be eye-wateringly expensive

[cedar osprey]

End of April they dropped $599 Mac Minis the base is $799 now.

(The cheapest hardware entry)

[near owl]

@zealous lagoon in my country, running those gpus that is always running the model, asking it, and compacting at 127k for the whole month is 1 buck in the electricity bill.

[ember moth]

I want to see what the price of their 512GB M5 Mac Studio is going to be. That will be our first test. I know they got rid of the cheaper Mac Mini but I want to see a real price raise before my antlers go up.

That thing is probably going to be like 15K from 10K

[zealous lagoon]

yeah that's quite annoying as everyone was just using external ssd for cheap. Now you're forced to pay the extra $100 to Apple for a measly 256g extra

[cedar osprey]

Easily $2K

[ember moth]

No the Mac Studio 512 M3 was 10K. The M5 is going to be easily 15 maybe

[cedar osprey]

For the M5 Ultra sure

[thorn thicket]

Me too… if it’s available via the education discount…

[cosmic zenith]

I have an M5 macbook pro and it does OK, though I don't run a lot local on it. I am actually testing a msi edgexpert right now

[gloomy haven]

Showing this

[thorn thicket]

I don’t have that type of cash to play with but I wouldn’t mind say a 128 gig studio.

[cedar osprey]

Yeah that would be the ultra.

[twilit ridge]

I am using

[ember moth]

Yeah you would want that one because you want good tokens per second

[twilit ridge]

I don't care anymore, I'll use beta too, as long as it works for me

[cosmic zenith]

Is the ultra making a comeback with the M5?

[near owl]

ts all started by the way from me talking about deepseek v4 flash locally on 2 GPUs. 🤣

[wanton gorge]

I bet the 512GB M5 Max Studio is going to be like $25,000

Maybe more

[thorn thicket]

Maybe for the Mac Pro if they come back with that form factor. For studio hell no

[ember moth]

My problem with this price point is that by the time you get your value back from what you would have paid for a subscription model, local models are going to get more efficient and you won't even need that level of hardware anymore by the time you recoup your money

[cedar osprey]

The inference still isn’t fast enough for it to be worth buying @ 10K+

You have way better options.

[near owl]

expensive, you can run deepseek v4 flash with 2 GPUs, expensive (and effective) coolers, fans, and dont forget your bedroom is gonna be hot.

and remember: DO NOT BUY A DAMN MAC PRODUCT TO RUN FAT CHUNGUS LOCAL MODELS BECAUSE IT IS GOING TO RUN AT 25 TOKENS

[ember moth]

Looks like a config error. Do you have access to claude or Codex so they can fix this in the backend?

[wanton gorge]

I'm hearing the M5 series studios and minis might get delayed until a fall release because they're struggling to get enough RAM supply for the iPhone Folds/Ultra launch

[cedar osprey]

True @ 25 tps lol.

[wanton gorge]

Fall release*

[near owl]

and anyway, your all gonna be broke in a few months if you do get the 2 gpus, my electricity bill is 1 buck to run 2 GPUs, coolers, and those gpus working a lot.

[ember moth]

Yeah everyone worries about, "Oh I have enough RAM to run my local model," and then they're doing that like 12 tokens per second. They're like, "What the hell is this?" Lol

[thorn thicket]

That was me with Gemma lol. 36GB of ram and I was getting not many tokens per second lol

[ember moth]

That's everyone who's running a DGX Spark right now. It can load it but it's slow as shit

[near owl]

mac mini broke my expectations as well on local models

[wanton gorge]

🥀

[ember moth]

Yeah the hype and the YouTube content creators will have you thinking RAM is all you need

[thorn thicket]

Local models are definitely getting better just not something I’ve played around with enough since opencode go is so cheap and practically subsidizing models

[cedar osprey]

Yeah I actually feel bad for them though, they could’ve just built their own CUDA rig tbh.

[ember moth]

Yeah and then they do that. Oh I'm getting great tokens per second and they glance at their UPS and it's pulling 700watts lol

[cedar osprey]

You guys should start training your own models if you have the compute it’s fun.

[thorn thicket]

Might be doing this for research at my university… we have a high performance 8x64GB RAM cluster…

[ember moth]

That's why for right now at least, subsidized Frontier model subscriptions are for the win

[cedar osprey]

Holy, they’ll let you? For sure do it omg

If I had 2 days with that damn cluster 🙂..

[near owl]

lets wait a year to hear that "Error: line 2313 coded.get_llm()"

I swear training is not fun, you train it, then you get an error mid training and you've been waiting a month for it to finish training and then u have to restart it

[thorn thicket]

Just have to convince the guy in IT to give me docker access lol

[near owl]

do they actually have 8 64GB gpus?

planning a helicopter heist rn

[thorn thicket]

yep pretty new too. I saw some people running jobs on it for the hackathon we had a couple months ago lol

[near owl]

what university is it, how do you access it and how will they give you permission? Run deepseek v4 flash on it haha

[thorn thicket]

won’t tell you the first part, it’s on ssh if you have university credentials, and a lot of paperwork lol

[cedar osprey]

Lol nah.

[near owl]

forgot not to ask the first question 😭

[wanton gorge]

You can cluster together Macs to stack up unified memory. Apple recently solved the latency issue and clustering Macs actually increases the tps instead of decreasing speed

Crazy work

[cedar osprey]

25 -> 75tps

[wanton gorge]

Thanks to the MLX team

[thorn thicket]

I’ve got to try this lol I have a m1 burning a hole in my pocket

[wanton gorge]

I think the cluster is bottlenecked by the slowest machine in the chain, though. It's just increasing the number of machines in the cluster no longer slows everything down like it used to before

[fathom pelican]

hey guys, is it a hype buying mac to run local llms?
also mac vs gaming laptops...

[near owl]

mac mac mac mac, get a mac mini. If you want to run local llms you will be broken and use cloud models.

[fathom pelican]

what about some budget gaming laptop - for coding, llms, some gaming... , software works

[ember moth]

If ur not gonna buy a high-end Mac Studio. Don't bother

[fathom pelican]

is anyone having a gaming laptop..

[wanton gorge]

Check out Darkbloom by Eigen Labs. They enable anyone with an Apple Silicon Mac to become an inference provider and earn money for serving inference.

Providers keep 95% of platform earnings

Users get 50% - 70% cheaper inference than OpenRouter

It's 100% private inference as well, they've figured out a way to leverage Apple Silicon security architecture to encrypt what's going in and out.

Fun little way to earn some money while you're sleeping

There's so many Mac computers out there that are idling doing nothing, a lot of sleeping compute just waiting to be tapped into

[thorn thicket]

Very cool saving this.

I have a lot of sleeping compute. And one piece of kit that’s not used at all 24/7.

But now I’m gonna sleep.

[wanton gorge]

Night Ritz

[ember moth]

Sureeeeeeee

[thorn thicket]

I said that and then I had an awful flare. It’s subsided now so I can sleep.

[ember moth]

Gn gn

[cedar osprey]

No

Lol trying to run local models on a “budget” gaming laptop, just put a potato in the microwave.

Because that’s exactly what will happen to your laptops battery with how much heat you’ll be getting.

[wooden imp]

Heya, has anyone tried Ollama Cloud's Pro plan on OpenClaw? How good it is compared to gpt-5.5? GPT is amazing, but I cannot stand Codex's rate limts anymote

[cedar osprey]

Ollama has very unreliable servers. No.

[wooden imp]

Damn really? Thanks for telling me. Any other ideas on what models I can run to replace 5.5 at around the same price?

[ember moth]

People are always grossly underestimating what it takes to get these local models to run efficiently. I blame the content creators

[cedar osprey]

There isn’t any 5.5 equivalent for the same price as GPTs subscriptions.

You should try other models before locking in another plan somewhere.

If I told you try OpenCode Go and none of their models give you the 5.5 experience you’re looking for it would be a waste of your time lol.

[wooden imp]

Thank you so much!

[cedar osprey]

Np

[cosmic zenith]

The $20/mo one is worth it I think but I wouldn't pay any more than that. There's also been people reporting that Ollama has been tinkering with their allotments and people are hitting limits shockingly quickly now. I haven't been using it the past week so I cannot say for sure

[cedar osprey]

There was a huge wave of people saying Ollama basically dropped their usage by 3x and people were hitting limits insanely fast.

Not sure if this is still the case but it’s still worth noting. On top of the horrible reliability/uptime..

[cosmic zenith]

I haven't run into the uptime issues, but I don't have agents running 24x7 or the like so it's more likely that I've simply missed the outages. I did get a lot done last weekend and didn't run into any issues that weren't caused by "unplanned maintenance" on the $DAYJOB network. PS Comcast sucks at any level of service.

[wooden imp]

Probably optimizing my agent could help so I don't finish my tokens in 2 days 😂

https://github.com/OnlyTerp/openclaw-optimization-guide/blob/master/SCORECARD.md

I found this on GitHub, will probably help

[cosmic zenith]

I'm tinkering with my new box right now to see what it will/won't run decently. Mistral medium 3.5 is def off the list. Box said NOOOOOO

[wooden imp]

Specs?

[cosmic zenith]

it's the MSI clone of a Spark

[wooden imp]

You're lucky, I have a gtx 1650 super and 2 pi5 😂😭

[cosmic zenith]

I've been using a rtx2070 so I feel your pain.

[wooden imp]

Should definetly change my pc now, I built it in 2017

[cosmic zenith]

LOL. I built this one in 2018. He's currently a proxmox host

[tulip orbit]

I use minimax at 10$/m and never reached the limit, model is not as capable tho

[wooden imp]

Do you guys have any ideas on how to "optimize" OpenClaw outside of the above?

[near owl]

Beautiful. Absolutely no potato, no top hat, no spaghetti, no tears. SDXL looked at "confused potato contemplating existence at a spaghetti desk" and went "nah, here's a yellow chair with eyes." Perfect. Peak AI.

[fathom pelican]

has anyone used gpt plus subscription with openclaw? the gpt 5.5 model...
how to configure it after buying the subscription...
I mean there is codex and there is chatgpt.com...
is the gpt5.5 api inlcuded in plan or what??

[zealous lagoon]

It’s the same.

[shell bronze]

Is it possible to use GPT5.5 pro in OpenClaw 2026.5.27?

[wanton gorge]

As far as I'm aware, the pro models don't get served anywhere else except chatgpt.com

[indigo mirage]

It's still the beginning of the end...

[indigo mirage]

They have hardware now

[indigo mirage]

Yes but now China don't want to buy. They have their own chips and in 1-3 years their own ASML.

[indigo mirage]

AI is becoming a standard product with standard margins. It's the worst scénario for companies like Anthropic

[indigo mirage]

Huawei 950PR, Zhenwu M890, MTT S5000...

[indigo mirage]

They have a lot of hardware now

[peak pewter]

who are the mods?

[wanton gorge]

Yeah true

[zealous lagoon]

We are the mods, we are the mods, we are, we are, we are the mods.

[indigo mirage]

The biggest problem of companies like Anthropic is that they let 95% of the market to Chinese models. They are clearly not competitive for all little/medium and high models. They are just telling "we are very expensive but we are 10% much more performant than other models" but for most of people having 10% less performances for prices divided by 5-10 or more... it's just better.

I prefer having 10% les performances and pay 40 cts / 1M tokens vs 5$ for models like Opus.

What would be great is to really lower price for old models

For example 5$ / 1M for GPT 5.5 / Opus 4.8 and 1$ / 1M for GPT 5.4 / Opus 4.7.

[zealous lagoon]

Would be nice but newer models are more efficient so they end up cheaper to run for the same task.

[indigo mirage]

Yes but the fact that in term of price / performances, they have nothing against segments like high like Kimi k2.6, Mimo 2.5 Pro, Deepseek V4 Pro (40-70 cts / 1M) or middle segment like Hy3, Deepseek V4 Flash, Xiaomi Mimo 2.5 (6-14 cts / 1M)... and on lower simple models they have clearly nothing against models like Ling 2.6 Flash (1 ct / 1M)

[near owl]

hello

[ruby crane]

'sup homegirl

[near owl]

aint dealing with your bullshit, get blocked.

[zealous lagoon]

bwahahaha

[ruby crane]

fair, tbh lol

[steel vector]

😂

[ruby crane]

what does everyone actually use for useful general agent web browser functionality?

i was excited to see that chromium had been added to the docker images, only to immediately be reminded why it's a bad idea (bot blocking, mainly)

until now I've just been allowing to invoke a chrome browser (which I signed in on it's own account(s)) on my local mac instead and it invokes it remotely

[zealous lagoon]

just browser relay for general stuff. firecrawl for the toughest anti-bot sites.

[ruby crane]

i find browser and computer stuff very unsatisfying. that's not just a claw thing though. claude et al are just as clunky

[zealous lagoon]

true. the whole thing about taking screenshots and then clicking where it remembered the clicky bit is etc.. ugh. browser relay is ok it's just data.

But it's terrible with js sites

[ruby crane]

i remember when openai first introduced their "agent" and it was a clunky-ass vnc session the agent spent ages misclicking on. took like 20 minutes plus to discover and book (I had it stop just short) a place that i could have done myself in < 5mins

it is not improved by connecting to actual local desktop, unfortunately. seems something that sounds simple in theory, but really difficult in practice

[zealous lagoon]

btw I just tested Qwen 3.7-max and it is buttcheeks. on par with Grok for medium complexity project management organisation benchmark, and much worse on a simple tool-heavy tally and sort benchmark. Didn't even get the numbers correct, after going back and correcting itself. I can't believe this model is the same price as k2.6 or glm5.1

[ruby crane]

the only non-anthropic, non-openai model I've tried that I truly got on with is minimax

though i couldn't really give glm a fair shake since they throttle the hell out of openclaw (and say they will, openly) and i subbed from source, rather than a 3rd party

of the other one of the big three, gemini, that thing is just too much of a fucking maniac to be trusted

[cinder nimbus]

@rotund valve welcome our fresh scientist

[zealous lagoon]

Gemini is a menace. GLM5.1 and k2.6 I like.

had surprisingly good results from deepseek4pro but then I've seen it do some questionable things too which put me off

[ruby crane]

once a model breaks your trust, it's really hard to use it without ideating and issuing a list of directives wide and deep enough that you either a) might as well have done it yourself, b) used a faster, cheaper model, since you already did the work or, most likely, c) just used a model that none of that was necessary and did the work while you did something else

[cold jacinth]

What about k2.6 with agent swarm lol

[smoky mist]

Good morning (1 pm is morning in my head XD)

[tiny reef]

That's still midnight

[smoky mist]

nah 13:00

[tiny reef]

Yeah, midnight

[smoky mist]

1 am is midnight

https://www.timeanddate.com/worldclock/germany/berlin

[tiny reef]

You missed the joke.. you made it tho😭

[smoky mist]

oh sorry just got up XD

[tiny reef]

Lol no worries!!

[near owl]

Hello

[indigo mirage]

Seems were right. Now Xiaomi Mimo 2.5 is 5th most used model on OpenRouter and Xiaomi Mimo 2.5 Pro is 7th.
Hy3 first. Deepseek V4 Flash 2th and Deepseek V4 Pro 8th.

Clearly the Xiaomi big price drop has made these model a growing star 😎

[twilit ridge]

🤣

[twilit ridge]

DeepSeek v4 pro likes to go crazy

He likes to do everything for me, even things you don't want to do, even deleting and reinstalling the system for you.

I feel that the most powerful models are still those from OpenAI, Anthropic, and Google.

Although it's expensive, it's more stable and less prone to seizures.

[slow robin]

Anthropic passes OpenAI in private market cap valuation.

[zealous lagoon]

Deepseek was debugging openclaw for me, it made some serious errors and killed its own gateway... when I brought it back, it made more errors just writing the issue report. I had to get claude to rewrite it and file the PR because i couldn't trust Deepseek to do it

[unkempt portal]

Is there a way to give an agent access to OneDrive without setting up deep integration with graph api or anything? I'm just wanting it to have access to a couple files I share with it. Sharing a link seems to be blocked by MS.

[near owl]

deepseek was crazy, it managed so much. Did ya use deepseek v4 flash?

[thorn thicket]

Morning all

[near owl]

Heyyy! I missed ya

[zealous lagoon]

Why would sharing a link be blocked? If sharing is a core part of onedrive functionality?

[thorn thicket]

I slept. Have a meeting today with someone. Sigh… too many things.

[near owl]

What's the meeting about?

[unkempt portal]

Idk. The agent isn't able to access the share link. Works find for everyone else.

[near owl]

#1459642797895319552 #1459642797895319552 #1459642797895319552

ask the CLANKER that is MADE to help you.

[zealous lagoon]

maybe try a folder? try mounting the folder for your agent and have it access as a regular folder maybe? Seems a bit odd... maybe it's the way the bot is accessing it that isn't human-like enough.

[elfin willow]

ya or if windows/wsl try a junction

[indigo mirage]

For me it does the job withourt any issues Same for Xiaomi Mimo

[unkempt portal]

That's worth a shot. It is odd that a link without any permission restrictions doesn't work.

[indigo mirage]

Even Minimax has done me some debug on OpenClaw code

[indigo mirage]

That's why so much people are using it

B'coz it's not reliable 🤣

[thorn thicket]

Don’t know, actually.

[near owl]

https://x.com/Be_Believing/status/2060013849353261064?ref_src=twsrc^tfw|twcamp^tweetembed|twterm^2060013849353261064|twgr^8b703a33b7a04a69a1c4f596ef79095277b2a9b7|twcon^s1_c10&ref_url=https%3A%2F%2Fpublish.x.com%2F%3Furl%3Dhttps%3A%2F%2Ftwitter.com%2FBe_Believing%2Fstatus%2F2060013849353261064

what the hell.

Be Believing (@Be_Believing)

Yet another childhood trauma unlocked

X · May 28, 2026, 3:02 PM

[mystic orchid]

@Be_Believing via Twitter

Be Believing (@Be_Believing)

Yet another childhood trauma unlocked

**💬 4 🔁 15 ❤️ 362 👁️ 35.9K **

▶ Play video

FxTwitter · May 28, 2026, 3:02 PM

[twilit ridge]

Did you install Homebrew when you installed the skills?

[sterile lance]

https://tenor.com/view/clam-krill-gif-15825264775687013933

[nova kindle]

morning, all

[viral scarab]

i didnt expect qwen being ahead of kimi

[crisp nimbus]

Which deepseek were you using? The flash or pro? I’ve been thinking of trying out deepseek for my openclaw myself

[zealous lagoon]

ds4pro

I would never even attempt such debugging with flash. flash is pretty useful for a model which is effectively almost free though!

[crisp nimbus]

From the official api or some other provider?

[zealous lagoon]

opencode-go

[crisp nimbus]

Hm. Not sure if that might make a difference. My experience with deepseek v4 (both pro and flash) has been very positive from the official provider when using it in other harnesses

[zealous lagoon]

Most of my tests were fine. Most of the time it's fine. I had great results in testbenches I made. But for this large complex actual real world task it made mistakes I was not expecting.

[crisp nimbus]

Got any examples? I’m curious what kind of mistakes it might make

[zealous lagoon]

I have a permanent opencode session with DS4pro as my network engineer.

I'm sorry I don't have any data. I had to restart gateway and clean up after it.

[crisp nimbus]

I see. My personal experience with other models debugging openclaw stuff has also been bad tbh. Gemini makes some weird ass errors sometimes and even typos lol. Claude also messed up once, although I didn’t use it that much because it’s so expensive… tbh the only one that’s been reliable for fixing stuff and diagnosing issues has been codex-gpt5.x

But that lacks personality imo and too easily slips back into the generic “assistant” voice

[zealous lagoon]

gpt is best for debugging. Then let claude write the incident report and PR

[nova kindle]

i think my gpt has too much personality rn

[zealous lagoon]

just my 2c.
If you want the no-personality gpt, just use 5.3-codex

[nova kindle]

ugh. i can't get 5.3-codex to even consider proper code-fixes

[crisp nimbus]

I had to make some manual changes to the system-prompt.ts file in the source just so it’s less likely to slip back out of its assigned persona… it’s kind of annoying. Other models don’t have this issue