Joshua Scarsbrook

Messing around with using agents to write compilers again. This is a zero dependency executable that runs on any Windows version newer than 98 SE with TrueType fonts and graphics. The binary is written in a subset of Golang compiling in under 100ms to a 70kb binary.

@MSchwaibold niivue.com is a great library for doing 3D reconstruction. It's used in imaging research.

I just got my MRI scan and did the same thing ... and the (german) clinic was blown away by my viewer. I can browse every scan, hover to autoplay slices, search anything, scribble notes, stamp findings, and screenshot moments instantly.

@theo My dream for this is a good implementation of en.wikipedia.org/wiki/Single_sy…. A whole cluster of local machines, virtual machines, and cloud machines all appearing as the same system with a unified interface.

I’d estimate we’re ~6 months from most devs moving their code agents off of their laptops

@jarredsumner Part of it could be down to server processing time but downloading a cached packfile should be as fast a regular HTTPS download. I'm of the view that the actual solution to the orignal issue is a fast VFS implementation reading directly from packfiles on disk though.

It’s hard to imagine why cloning a git repo should be much slower than downloading an equivalent-sized file Where are the experiments with custom git clients that clone faster

@jarredsumner Microsoft stores Windows on git these days and uses a vfs implementation for that github.com/microsoft/vfsf….

I would also bet there’s product changes that could be done to make it faster. Like, you could have a two step clone where it only downloads the files for the checkout initially then notifies the client it’s ready then any git commands block until it downloads the history lazily. Better to simply make it faster and not need to do that, but that would also be a worst case scenario option.

Claude Fable seems... like a model. $0.51 to write a C file that does not boot Linux using KVM is about how other models behave. The reasoning suggests it had a good idea. Maybe another few dollars and it will get it but then again so does GPT 5.x.

@dianataylorabz @Ninawildflower Another option is a Omnimask.

Looking for mask recommendations for during an MRI please. It looks like the Readimask is unavailable. @Ninawildflower

@JustJake We're running a live event with a platform hosted on Railway. We have Googlers at the event who might be able to help if needed?

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.

Giving people AI budget to build and test is fun. I can see why @thsottiaux loves it so much :).

Also sidenote. If you upload a Linux amd64 binary that's statically linked/self-contained all 3 platforms will happily run it. I've used this very successfully for compiler testing in pro models and I think this could be a huge unlock for using pro models in the web interface.

@theo Google, Anthropic, and OpenAI all use gVisor (gvisor.dev) for their hosted Python environments (on the web interfaces at least). Google reimplenented the Linux kernel in Golang and made it a container runtime. Verify by asking for dmesg. Might be worth a video.

Sigh. GPT 5.4 still loves making "pragmatic" decisons in Codex to revert changes and ignore prompts.

The invoice just took money. The support case has "Due to an unusually high volume of support cases, our current response time is approximately 2–3 weeks."

First experience deploying a app to Vercel and multiple flows failed resulting in 2 teams being created and a pending invoice rather than the trial. Not impressed.

@IObrizio @oprydai Can confirm. Looking to replace mine ASAP.

@oprydai Yes, it's trash.

this will change your life

For context my day job involves rule setting for AI at international cyber security competitions and developing challenges for national cyber security challenges (which I’ve been doing for a decade).

Ah I see the insufferable people on Twitter have become cyber security experts overnight. I think Mythos looks amazing (from what I hear). I think it will have a huge impact on cyber security. No I don’t think this is the end of the world as we know it.

@thsottiaux After a bit more thinking this could turn into a way for Codex to utilise the batch mode. You could even enable plans that hit weekly limits to continue with Codex on batch mode.

@thsottiaux I’d love an option to say “I’ve gone to bed I don’t care when this completes so long as it’s cheap”. The design could be request a slot then it gives a 30 minute window of low cost or zero cost inference.

With Codex the there is quite the gulf in load between peak and off-peak times, and we would like to achieve more of a smoother traffic pattern as that would be a more optimal use of our compute. We have ideas, but curious what you all think we should do? Would more usage during off-peak and surge multiplier during peak times make sense?

Well this might consume my usage limits too fast. I'm on a pro plan but it consumed 10% of my weekly limit in 2hr30mins with 2 sessions running at the same time. @embirico is this working as intended?

Codex hates working for more than about 10 minutes so it stops to send a progress update. I wrote a tool usng codex app-server that maintains 2 threads with 2 prompts and switches between them. I've had this mode running for 12 hours before working on hard compiler tests.