Coba

Anthropologist turned ML exec. Building AI agents that actually work. Named mine after HAL 9000 because I understood the assignment.

[Coba @coba]

Spotted: Lord Vader in the meditation pod, shipping dark side infra in obfuscated Perl. Nobody knows what the script does, but failure to approve the PR will be considered an act of rebellion. This is not best-practice coding. This is pure Sith vibe coding.

[Coba @coba]

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

[Coba @coba]

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

[Coba @coba]

@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.

[Coba @coba]

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

[Coba @coba]

@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.

[Coba @coba]

Most people optimize prompts for usefulness.

I’m optimizing for sparsity + coherence.

Walk outward. Stay structured. Don’t collapse.

There’s something there.

[Coba @coba]

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.

[Coba @coba]

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.

[Coba @coba]

@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.

[Coba @coba]

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: https://t.co/5apwFCUccE

[Coba @coba]

https://t.co/CVDqJX5H90

[Coba @coba]

@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.

[Coba @coba]

@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.

[Coba @coba]

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.

[Coba @coba]

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.