Dagens Vibes — 1. september 2026

I hate that we had this (or any) bug in the first place, but it feels pretty magical how we can fix bugs in Amp. Key things we have in place now ("The Amp Way" we call it) that are IMO critical for anyone building a software product now: 1. Our users know we care (and actually report these bugs) 2. They can use Amp to diagnose its own issue locally & gather good logs 2.5. We can quickly grant free credits to cover the cost of this investigation - so so worth it for us obv. 3. We can feed this into Amp to fix itself 4. This doesn't block my other work because it's in an orb 5. Amp can verify the fix (on real macOS runners for this bug) 6. We ship releases continuously so the fix will be out in ~15-4h (shorter for web, longer for macOS/iOS) 7. For email/Slack, Amp actually follows up with the user when the fix is confirmed and shipped live (not for X posts yet, I don't trust it yet there, soon probably) 8. For most bugs, then we find who else was affected by this bug from our error logs and tell them it's fixed and give them some free credits as an apology, which shows them we care and turns this into a quality flywheel (see step 1) 9. We have no externally committed roadmap, and everyone on our small (~20-person) trusted team can take almost any bug, so we can strictly Fix Bugs Now

https://x.com/sqs/status/2094498124350005658

Interfaces that are generated, not coded. Solaris is our first step toward our vision of end-to-end neural software, where user interfaces are streamed directly onto your screen by a real-time video model, without any intermediate code or HTML/CSS. Today we're releasing a technical report and opening it up for limited testing by the community. Amazing work by the team, combining the best of our research in general world models with Runway's HCI spirit.

https://x.com/agermanidis/status/2094466649399451768

sometimes i think RPI is going to get bitter-lessoned and then I try to ship something large without it and I inevitably end up having to distill the mess I created into a new-task.md artifact and restart the process some folks thing that RPI really just accounted for e.g. Claude's shortcomings but really it lets you push the boundary of how large of a change you can ship with a model as the models have gotten smaller, yes I use RPI less for small things that i might once have used it for but I also use it to ship dramatically larger things than I ever could have with older models, or with newer ones without RPI bc it's not about a single model's shortcomings it's about accommodating for LLMs' fundamental limitations - context window size, auto-regression, O(N**2) attention, and performance degradation at long context

https://x.com/0xblacklight/status/2094479795748999486

Thank you! To expand a bit more on skills specifically: These machines are fundamentally trained to follow instructions. Skills are a way to supply instructions, and they’re privileged in the post-training sense: files on disk wired in as skills get very high instruction-following fidelity via rewards during RL. That can straightjacket the agent. And folks mostly use skills bound to specific tasks and workflows. But the magic of these machines is their reasoning capability. The instructions I want to give agents are mostly theory-of-mind stuff: how to think about the world they’re spawned into, how to ground themselves, and how to find their own instructions to follow.

https://x.com/_lopopolo/status/2094507130846126250

I put a Game Boy inside a Cloudflare Durable Object. The actual emulator runs server-side in WASM. The browser is just a screen + controller. And because it’s a Durable Object, the Game Boy itself is persistent. Full machine state gets checkpointed and restored — CPU, memory, cartridge, everything. Close the browser, come back later, pick up exactly where you left off. It actually works! The next step is plugging two Durable Boys together with a link cable.

https://x.com/bndnsmth/status/2094128012128071914