New article! A user is reporting full system freezes while using Superluminal on Linux. What do you do? Cry? Well, we did a little bit.
But we also dove into the kernel...again, fixing several issues in eBPF's spinlock implementation. Read all about it:
https://t.co/HHAyGdnZYZ
Have you tried the Pi agent? My usage of LLMs when implementing something tends to be full of sidequests that eat up a lot of context, which sticks around forever in normal agents.
With Pi I aggressively use its /tree command to prune context when I’m done with a sidequest, keeping the context focused on the main task. I tend to never go above ~40% context usage on a 200k context with this.
It’s basically manual memory management, but I quite like the control it gives you. I have sessions that I’ve been working on and off in for weeks that still have a working context, which I could never manage to do with the big lab agents.
Neither imo. I'm not super familiar with TS, but it looks to me like you're iterating over the messages array twice; once to copy to the newMessages array, then again to prune. It probably won't matter perf-wise, but it seems like a waste.
I'd start with an empty array (can you reserve in TS?), go through the old array, for non-image-bearing-messages, just copy 1:1, for image-bearing-messages, produce a new message with the images stripped. I'm guessing non-image-bearing-messages are far more common than image messages, so the slow path is only hit infrequently.
@olson_dan My pet theory of AI use is that each developer goes through the hype cycle individually, at a much faster rate than the industry does as a whole. I’ve certainly gone through it myself.
Due to popular demand, you can now browse all my recommended reading/viewing suggestions here:
https://t.co/V3X0F18XuE
Sadly, Space Karen's hard core engineered Twitter API is absolutey garbage, and I can only get tweets back to April. Will backfile with an export.
Enjoy.
Me: we’re running into an issue on Linux with dbus, we think it’s related to <this> github issue
Claude: <long, detailed, plausible explanation of why it’s a OOM issue>
Me: what? did you even read the issue?
Claude: You're right, I didn't read it. Let me actually look at it and revise.
Claude: You're right and I owe you a correction. I didn't fetch the issue and made up an explanation that sounded plausible. Now that I've actually read it:
agi is here folks, fear for your jobs
Today is a beautiful day for updates!
On the Stable channel: rolling up a large number of features, perf optimizations, and QoL changes made over the past months
On the Insider channel: support for our instrumentation API has been implemented on PS4 & 5!
Go check it out!
I’m so absolutely freaking tired of waking up to a freshly rebooted machine because Windows, not me, decided installing updates I absolutely do not care about was more important than preserving my debugging state. Whoever is in charge of this at MS should be fired.
@nicbarkeragain Your stuff is always very good and in depth, and I’m trying to do my part with my recently started blog (be the change you want to see in the world and all that), but it’s so easy to get drowned out by an endless amount of low effort, content free spam.
@nicbarkeragain I feel like this is the result of technical writing turning from being about sharing knowledge to “content marketing” and “building an audience”.
It’s the worst. I miss the days of having multiple good, in depth, technical blogs to read from on a semi regular basis.
@badlogicgames The upside is that showing up with a good bug report makes you stand out.
I recently submitted a GPU driver bug to an IHV with a small, self contained, 100% repro app and they were so surprised they fixed the issue then and there and shipped it in the next driver update :D
@badlogicgames Tbf, it has been like this for a long time.
The number of bug reports we’ve received over the years that round to “it doesn’t work” from a group of SWEs that I know is otherwise excellent is… disconcerting.