←
Ray Chen715 posts
Ray Chen avatar

Ray Chen

@raychen

meow@ray.cat

◷ Joined December 2016▧ 140 Photos and videos
Ray Chen avatar
Ray Chen@raychen·Oct 23

All deployments on Railway is now enabled. Please feel free to shout here if it doesn't work.

Enabled for Pro users now! Working on enabling Hobby & Free/Trial next.

Ray Chen avatar
Ray Chen@raychen·Oct 22

Enabled for Pro users now! Working on enabling Hobby & Free/Trial next.

Builds and deployments have been re-enabled for @Railway Enterprise customers. We’re working on enabling builds for all tiers, starting with Pro next. Apologies for the inconvenience folks! Live updates available at status.railway.com/incident/…

Ray Chen avatar
Ray Chen@raychen·Oct 22

Builds and deployments have been re-enabled for @Railway Enterprise customers. We’re working on enabling builds for all tiers, starting with Pro next. Apologies for the inconvenience folks! Live updates available at status.railway.com/incident/…

Ray Chen avatar
Ray Chen@raychen·Sep 14

Customer support in 2026 is just two LLMs talking to each other with a meat proxy on each end.

Ray Chen avatar
Ray Chen@raychen·Sep 4

Culture is the compression of a thousand decisions you no longer have to explain.

Ray Chen avatar
Ray Chen@raychen·Sep 4

A community member in our Discord asked "what causes @Railway to go down". I ended up writing more than I planned. Sharing because the answer is less obvious than people think! (Caveat: this is the high-level version. Infra failure modes as a platform are long tail and I left plenty out, but this covers the shape of it.)

Post attachment
Ray Chen avatar
Ray Chen@raychen·Aug 25

Model-specific ASIC like Taalas is the way forward IMO. TPUs are still bottlenecked by its HBM for model weights. Taalas etches the model on the silicon so weights are directly fabricated into the silicon itself. Taalas has accomplished 17k TPS on Llama 3.1 8B at ~250W power draw which is astonishing. Imagine a future where models are cartridges you can plug-and-play like components of a Framework laptop.

Noah@itsnoahdAug 25

OpenAI's Jalapeno chip is already performing at 2x performance. I am very confident that we created GPUs for processing pixels concurrently as the CPU fell behind. We are now using GPUs to handle tokens and I just dont think its sustainable. The TPU is the way forward for driving down prices.

Post attachment
Ray Chen avatar
Ray Chen@raychen·Aug 24

The best interview question is a design prompt with no right answer, because the job has no right answers either.

Ray Chen avatar
Ray Chen@raychen·Aug 21

It’s 100% a skill issue. I’ve been fully remote in globally distributed orgs for more than a decade, and I’ve built successful remote teams and processes spanning 15-hour timezones. A good remote team ships *more* efficiently than an in-person team. The key is autonomy, ownership, & clear written comms as defaults. Writing is also why remote orgs compound. Every decision leaves an artifact. In-person orgs evaporate their context daily. “Remote doesn’t work” almost always decodes to “I can’t tell who’s working/what they’re working on without watching them” which is less about remote and more a confession about your org chart.

I've seen leaders shitting on remote work quite a bit lately. I can confidently say it's a skill issue x100 Remote work isn't perfect but neither is in person. A good remote team can ship just as efficiently as a good in person team. It is harder to build a good remote team though as most people don't have self management experience. Build a team of high ownership people and you won't lose.

Ray Chen avatar
Ray Chen@raychen·Aug 21

Churn dashboards are a lagging indicator. Your support queue is the leading one. Guess which one gets exec attention? ;-)

Ray Chen avatar
Ray Chen@raychen·Aug 21

"We're data-driven" usually means "we're waiting for numbers to give us permission to do the obvious thing"