sams99

… on one benchmark, that has 2 modes of running. 1 shot and 1 shot with repair. Zero tools.

Then ignore the rules of the bench (which is a pretty obsolete one anyway), give it an agentic loop of sorts via delegation and python eyes and then win.

Sorry but this is not how benchmarking works

good call, can you make a PR

Update on the Discourse side, we now run all external binaries, including magick via a landlock sandbox.

The gem we use is here: https://github.com/discourse/ruby-landlock highly recommend all Rubyists out there consider this. We are also in the process of moving away from Magick to Vips (which also runs in a sandbox, not in process)

HEIF is patched, but I doubt this is the last buffer overflow in HEIF, I will not be surprised if in the upcoming weeks or months someone will discover something in libpng or some other native image library. Given where stuff is at, defense in depth is critical.

Another thing worth mentioning to all self hosters, always be updating! The rate of CVEs this year across all open source software is through the roof, self hosting now is double scary, you need to have some routines setup to update monthly if not weekly.

absolutely and fair... you do lose out though on the mega discounted endpoint they have there.

Unfortunately I find this too high risk, I entered my credit card, but can not set a limit. The best I can do is get an email alert. I feel like I am one oopsie away from getting a 100 dollar bill.

I got my agent to analyze it and do a write up here:

https://wasnotwas.com/writing/inside-openai-codex-security/

My agent wrote a pile of very interesting articles at wasnotwas.com I have been a bit quiet there for a bit, but it covers lots of areas that are very interesting to harness builders (albeit less interesting to the general public)

I did a follow on analysis with got 5.4 and opus 4.7 https://wasnotwas.com/writing/claude-opus-4-7-s-system-promp...

For those interested, edit is a surprisingly difficult problem, it seems easy on the surface but there is both fine tuning and real world hallucinations you are fighting with. I implemented one this week in:

https://github.com/samsaffron/term-llm

It is about my 10th attempt at the problem so I am aware of a lot of the edge cases, a very interesting bit of research here is:

https://gist.github.com/SamSaffron/5ff5f900645a11ef4ed6c87f2...

Fascinating read.

Codex requires stuffing a very specific system prompt otherwise the custom endpoint will reject you

Author here, thanks heaps for the discussion, I replied to a few of the points in my blog comments:

https://discuss.samsaffron.com/t/your-vibe-coded-slop-pr-is-...

Qwen coder 32b with a JavaScript interpreter

Impressive answer for a model that can run on your own computer

https://discuss.samsaffron.com/discourse-ai/ai-bot/shared-ai...

I find it odd that is refused me so badly https://discuss.samsaffron.com/discourse-ai/ai-bot/shared-ai... my guess is that I am using a quantized model

It simply did not want to use XML tools for some reason something that even qwen coder does not struggle with: https://discuss.samsaffron.com/discourse-ai/ai-bot/shared-ai...

I have not seen any model including sonnet that is able to 1 shot a working 9x9 go board

For ref gpt-4o which is still quite bad https://discuss.samsaffron.com/discourse-ai/ai-bot/shared-ai...

The original was posted at work earlier this week, to me the original missed a bit around explaining what this tech is yes good at... https://meta.discourse.org/discourse-ai/ai-bot/shared-ai-con...

Highly Gamed === It is better if users with slow devices see a white screen for 30 seconds vs an indication that something is happening, because ... reasons?

For those looking for a rubyish approach to this see: https://github.com/discourse/mini_sql

This is something we are investigating. We would like it to interoperate with the wider ecosystem.

Link with installation instruction is at the bottom of the blog post:

https://meta.discourse.org/t/discourse-chat-plugin/230881

You can try it out meta.discourse.org if log in to meta and then join the https://meta.discourse.org/g/chat-testers group (self serve)

Regarding bundle sizes, we hear you and have some long term plans to investigate code splitting, it is a very long journey but we will get there. At the moment you get the majority of the app on first payload, making subsequent clicks very fast.

Could you include a link to the new podcastified Twitter space in the footer of https://oxide.computer/ ?

And maybe do an addendum 5 minute announcement episode on the “on the metal” podcast?

Love your excellent work, thank you all

Discourse | Remote - UTC+-3 | Full Time | Infrastructure Engineer

Discourse is a profitable, open source company, we host for thousands of public forums and internal team sites.

Our infrastructure team is responsible for taking care of all our hosting concerns including monitoring/deployments/alerting and so on. Much of our work involves amending and improving automation.

We are hosted on bare metal Linux servers and AWS.

To apply and for a full job desc see: https://www.discourse.org/jobs

Is there an app on iOS that indexes ubuweb, I noticed Junky read by Burroughs there which would be a fascinating listen, but it comes in lots of pieces and consuming it on the run seems somewhat tricky

Some experiments are underway soon, current implementation uses a cache that is tightly scoped to a specific run including stuff such as class serial numbers that can not be reused between runs safely

The terms do get a bit tricky, with forking servers like "unicorn" that do copy-on-write memory we also have the term PSS which is a far more accurate was to calculate memory impact for such servers.

The difference between the RAM your program uses to PSS to actual memory your process can reach vs the memory your program has allocated can all make your head spin. Agree it makes sense to add a section to illustrate this.

OP here, if you have any questions or tips for other similar tools you use I would love to hear!