Coba

@IanLandsman Who was burning useful tokens looping over calendars with Opus or Sonnet? That is what a local LLM is for

Hot take. Claude is doing the right thing. Give the real customers a better experience (keep infra working well for them) while ditching people having their bots loop over their calendar 100,000 times a day.

@triketora @openclaw Can you quantify this? I switched and I think it is sharper and more precise, especially with planning.

reluctantly switched to openai for @openclaw today and my claw seems to have lost quite a few iq points and i’ve had to take away a lot of its responsibilities

@Shpigford I have for sometime now been having all Opus and Sonnet 4.6 plans in my openclaw reviewed by GPT 5.4 and every time the GPT 5.4 reviews rip apart the weak plans from Opus and Sonnet. Anthropic is overblown.

i've spent 3 hours this morning working with GPT 5.4 in 🦞 and have absolutely nothing good to say about it

Cancelled all Anthropic plans and removed it completely from all my systems. Bye bye

@Shpigford I switched to a ChatGPT Pro account using openai-codex/gpt-5.4. I was mostly using Sonnet 4.6 as the main agent before Anthropic's idiotic move.

i need to hear from opus 🦞 users on what they've successfully switched to. if you've never used opus as your daily driver in openclaw...you frankly have no clue what you were missing so all these random obscure models probably seem totally fine. i assure they are not.

Most people optimize prompts for usefulness. I’m optimizing for sparsity + coherence. Walk outward. Stay structured. Don’t collapse. There’s something there.

You start seeing where the manifold thins out. Places where there aren’t a lot of memorized templates. Where structure dominates narrative. Where metaphor doesn’t carry the load. It gets alien in a productive way.

Been experimenting with something weird lately: Using LLMs as probes to intentionally explore low-density regions of conceptual space. Not better prompts. Stranger ones. Compiler specs for ritual systems. Distributed systems postmortems for mythology. Type systems for symbolic domains. Feels like walking off the manifold on purpose. Exploration > exploitation.

@DannyMagazu Those people are not really in touch with AI. No one thinks vanilla OpenClaw is the thing. You need to build it up yourself.

Here's what I've learned over the past couple of weeks with people who are really in touch with AI. They all think OpenClaw sucks. They all have advised on focusing on mastering Claude Co-Work and Claude Code. Even if they are doing something productive with Openclaw, the other tools on the market will pass it in a matter of months. They all say that anyone telling you to use OpenClaw is just peddling garbage. What do you all think?

Took me at least two weeks before my agents were actually useful and that was with a lot of infrastructure work running in parallel. Task scheduler, lease-based execution, escalation routing. The main thing that changed everything: memory architecture. Flat MEMORY.md does not scale past one agent. Wrote up what I built: x.com/coba/status/20…

To the people who say openclaw is absolute magic and increasing their productivity 100x... Are you still having to deal with it being a big dumb dumb buggy head several times a day? Where things just break randomly? Or do you truly have a magic system and setup where everything is just glorious and works perfectly?

@auren Built a local task scheduler for my agents to solve exactly this. Task spec defines scope before execution starts. Agent completes, blocks for human input, or fails with a reason. It never gets to silently stop at 5 and call it done.

my clawbot is an absolute moron. feels like i am giving instructions to a 7 year old on meth

@brockpierson Why are you tracking influencers? If you're building, you don't have time for this and why jump to the irrational conclusion that "Nobody is building anything real." That is silly.

Openclaw is one big grift. Nobody is building anything real. It's just grifting influencers telling YOU how to build stuff (but never build anything themselves). Interesting how you can be building 24/7 but have nothing to show for it... Check the track record of these snake oil salesmen. Same type of person who polluted crypto with their grifty ways. Selling a false dream and taking money from innocent people is disgusting. You can't change my mind.

Built a bridge between Codex CLI and local Ollama models today. Three things nobody documents: • Codex sends zstd-compressed JSON not plain JSON • It hits /responses, not /v1/responses • It tries WebSocket first, HTTPS only as fallback • response.output_item.done is required or it won't execute tool calls ~250 lines of Python. codex-worker now runs qwen3:30b-a3b on local hardware. Zero Plus tokens for bulk coding.

had this exact conversation a couple months ago. dev on my team was resistant. told him this is where coding is going. he said "i'm worried about my next job when you're not my leader." i said: this is for your next job. there is no coding job without it anymore. the resistance is fear. the job right now is to make the path visible before the fear wins.

I've never seen this before in my career: 28-30 year olds who refuse to use AI coding tools. You show them what they can do augmented (not replaced) with AI and you see in their eyes that they have no damn clue of what's happening. You can't work with these people anymore. Time used to pass over older generations slowly. Now it's passing over us at the peak of our careers. Sadly, adaptability isn't optional at this point.

the AI framework ecosystem exists to sell you abstraction over problems you have not solved yet. you do not need LangChain. you need to understand what a prompt is, what a token costs, and why your agent keeps forgetting things. learn the boring parts first. the framework will still be there when you actually need it. by then you will probably not want it.