Charles Pan retweeted
Charles and Dean answer the call, can’t wait to try this
Introducing Vespin, an inbox for all your coding agents
No more jumping around between 10 different chats, your agents just tell you when they need your attention
vespin.dev
Charles Pan retweeted
SotA forward deployed engineering in 2026: "try this prompt with your guy: ..."
Charles Pan retweeted
I don’t think there’s any benchmarks that accurately reflect the current capabilities of LLMs
Charles Pan retweeted
I think the inevitable ux for this is voice. single voice agent on top managing the other sessions and surfacing info when relevant.
codex got really close with this but the voice model was simultaneously not smart enough and not responsive enough.
This quoted post is unavailable.
Charles Pan retweeted
We just released @c15tdev v2.2.0
c15t is growing faster than ever at 69.7% MoM growth. We've almost hit 5m npm downloads in total, and in the last 30 days hit 1.5m downloads alone.
2.2 adds more of the consent-aware primitives teams need in real apps, and it had the most external contributions we’ve seen so far, which we couldn't be prouder about.
Thanks to everyone using c15t, opening issues and contributing fixes. We're so excited to unveil our next stage of @inth that build on top of everything we've done so far.
check out the changelog here c15t.com/changelog/2.2.0
Charles Pan retweeted
Been reading kimi k3's paper:arxiv.org/abs/2607.24653
They used 3 different runtimes:
1. one to run containers
2. one with gpu support
3. a microvm for isolation guarantees, efficiency, snapshot/fork/fast*.
smol machines is an open source vm designed with all of those properties from the start - to simplify infra that's basically duct taped together.
Charles Pan retweeted
Do you know how to share ONE GPU across many VMs?
I've been working on the answer to this for a while, here's my notes~
1/ Passthrough — 1 GPU wired directly to 1 VM. Native speed. Zero sharing. Whatever you don't use just sits idle.
2/ vGPU / mediated passthrough — the GPU hardware itself splits into isolated slices (NVIDIA vGPU, MIG). Near-native speed, but needs specific datacenter silicon + licensing. Each slice is still its own sealed box — no sharing between slices.
3/ API remoting — the guest's GPU calls get intercepted in software and forwarded to a host process that runs them on the real GPU. Slower (~70-85% native), but now a host sees and can orchestrate every GPU call — the thing hardware-level approaches can never do.
4/ Emulation — no real GPU, pure software. Compatible with anything, unusable for real compute.
I chose api remoting for smol machines.
It's the "slower" option on paper. But it enables me to do forking on vRAM, letting the clone share the parent's already-loaded GPU memory instead of reloading it.
Charles Pan retweeted
Stage is now free for open source!
We’re excited to offer the platform for free for all public repos.
You can even try it out without signing in - just replace any public GitHub url with stagereview.app:
Charles Pan retweeted
There is no way to get the human out of the loop. The Agents simply do not understand obvious user expectations.
Charles Pan retweeted
This. Understand the code you're responsible for.
"My AI agents one-shotted this app overnight and I didn't read any code and just shipped it" is not impressive; it's irresponsible.
You might believe you should spend less time thinking about code because of AI.
I strongly disagree! We’re watching this play out live where tons of AI generated code becomes a liability.
At the end of the day, an engineer needs to be responsible / on call for code that gets shipped to production. If you don’t understand the system you’re trying to debug, you’re probably going to have a bad time.
Yes, AI can help with all of this, if you set up the proper systems. You can have agents triage prod logs, look at errors, etc. You can speed up parts of the investigation, but an engineer needs to make the call. There might be serious customer or financial implications from that change.
I expect the trend continue for trimming dependencies, vendoring code so you can modify it directly, preferring simpler systems with fewer abstractions, and spending waaaay more time thinking about system design and code maintenance.
I’ve said this before, but it’s a great time to get familiar with CS fundamentals and some of the history behind what great software looks like. Many parts will be different in the coming years as AI progresses, but also a lot more than people realize will stay the same.
Charles Pan retweeted
hard to believe code review UI hasn't meaningfully changed in years
@StageReviewApp is finally pushing the space forward
go check it out ↓
Stage is a code review platform designed to help engineers understand AI-generated code.
Your team is drowning in an endless backlog of PRs. @StageReviewApp turns code review into a guided process and enables you to review changes faster than using GitHub.
Congrats on the launch, @ceefryingpan and @DeanStratakos!
ycombinator.com/launches/QQz…
Charles Pan retweeted
Stage is a code review platform designed to help engineers understand AI-generated code.
Your team is drowning in an endless backlog of PRs. @StageReviewApp turns code review into a guided process and enables you to review changes faster than using GitHub.
Congrats on the launch, @ceefryingpan and @DeanStratakos!
ycombinator.com/launches/QQz…
