Ray Chen

🏗️ @railway | [email protected]

[Ray Chen @raychen]

@shahaayush1999 @thisismahmoud @JustJake So sorry about the poor experience. "the vendor that burns me the most" is not what we want to be known for :-(

We're taking this very seriously and working on improving reliability overall -- I hope we can re-earn your trust again

[Ray Chen @raychen]

@theo Slight correction: they suspended our account

But still, yeah

[Ray Chen @raychen]

@ngeloxyz @pugson Yup exactly. We consult our network's control plane to know how to route traffic, and the control plane is down right now.

[Jake @JustJake]

It appears Google Cloud has blocked our account, and so some services are unavailable

We’ve escalated this to Google and will keep people posted. Deepest apologies.

[Ray Chen @raychen]

@boristane railway - probably every hour

[Ray Chen @raychen]

Pi gang -- what am I doing wrong?! I really want to use Pi because of its customizability, but the harness itself seems subpar for me. I've noticed that Pi+Opus4.7 requires more steering as well.

[Ray Chen @raychen]

I've been test-driving Claude Code versus Pi+GPT5.5 the past week. Here are the results.

159 Pi sessions vs 60 CC sessions. On the same project, Pi needed 2x the corrections and got cursed at 3x as often.

Opus 4.7 has been getting anecdotally getting dumber, but it looks like I'm keeping CC as my daily driver after all.

Prompt: "Stalk all my Claude Code (~/.claude/projects/) and Pi (~/.pi/agent/sessions/) sessions the past week, and tell me which one required less steering, and pisses me off the least."

[Ray Chen @raychen]

On a related note, Pi+GPT5.5 seems to be amazing at Rust. @thisismahmoud maybe that's why you prefer it 😛

[Ray Chen @raychen]

I've been test-driving Claude Code versus Pi+GPT5.5 the past week. Here are the results.

159 Pi sessions vs 60 CC sessions. On the same project, Pi needed 2x the corrections and got cursed at 3x as often.

Opus 4.7 has been getting anecdotally getting dumber, but it looks like I'm keeping CC as my daily driver after all.

Prompt: "Stalk all my Claude Code (~/.claude/projects/) and Pi (~/.pi/agent/sessions/) sessions the past week, and tell me which one required less steering, and pisses me off the least."

[Ray Chen @raychen]

@dominikkoch (unless your agent harness exposes cost-per-turn natively like Pi does)

[Ray Chen @raychen]

Share how much you've spent on agentic coding!

curl -fsSL "https://t.co/QTKSemJzdS" | bash -s

Works with Claude, Codex, & Pi. Most of it is estimated from consumed tokens according to your sessions, except when using Pi.

[Ray Chen @raychen]

@dominikkoch hmm yeah, might be inaccurate cos it goes off your history and estimates according to token rates

[Ray Chen @raychen]

@mattvagni what if LLMs were actually ghosts

[Ray Chen @raychen]

@ricardomacario @JustJake Looking into this right now - https://t.co/6bqL0HZpHr

[Ray Chen @raychen]

If you're not using https://t.co/Dxhxxzx860 for design, you're missing out

It works amazingly well with Opus 4.7 when you let it go wild with vague and not-too-descriptive prompts

[Ray Chen @raychen]

@varungoti @Railway Responded in your support thread :-)

[Ray Chen @raychen]

Information architecture is about to become VERY important. In a world where LLMs are the retrieval layer, your schema becomes your moat.

Garbage graph in, garbage answers out.

[Ray Chen @raychen]

@TimHaines Very much so. TanStack is amazing.

[Ray Chen @raychen]

@DhravyaShah Agent inside sandbox that can orchestrate its own sandboxes ;-)

[Ray Chen @raychen]

🫂 @AnthropicAI @cursor_ai