oofbey

14 ms·

271.
by oofbey 11mo ago
Also important to note that this post was authored by the original author of zig, who presumably effectively controls the project.
272.
by oofbey 11mo ago
> The reasons these projects have value isn't necessary user facing The value is what then? Promotion for the tech lead that convinced a bunch of other googlers that they should contribute to this OS project?
273.
by oofbey 11mo ago
Sounds like you have a flaky / damaged device or bad cables. If there really was some kind of conspiratorial timer requiring you to use 1P cables it would certainly be documented. Can’t hide that stuff. Loads of people use Apple device
274.
by oofbey 11mo ago
It is in fact significantly harder for Apple. Because nobody expects random spare time GitHub project to work perfectly. Or even very well. Apple’s reputation, and trillion dollar market value, is based on the idea that their stuff works p
275.
by oofbey 11mo ago
TIL: Heroku is still online. A shadow of its former self but still there.
276.
by oofbey 11mo ago
Go seems to do alright without exceptions. It’s just a convention to return errors, which seems fine as a practice.
277.
by oofbey 11mo ago
Ooh thanks!
278.
by oofbey 11mo ago
I like the idea of having IO integrated. I have trouble remembering the iotop flags and often run it in a parallel window.
279.
by oofbey 11mo ago
Haha. Joke, right? Please?
280.
by oofbey 11mo ago
ML researchers will sometimes vary the size of the training data set to see what happens. It’s not common - except in scaling law research. But it’s never called “ablation”.
281.
by oofbey 11mo ago
That’s backwards. New research and ideas are proven on small models. Lots and lots of ideas are tested that way. Good ideas get scaled up to show they still work on medium sized models. The very best ideas make their way into the code for t
282.
by oofbey 11mo ago
Nothing unique? Totally disagree. Unlimited compute resources aren’t literally unique but there are only a small handful of places in the world that have that. Vast quantities of private data, especially text communications and images. Very
283.
by oofbey 11mo ago
You want more research on small language models? You're confused. There is already WAY more research done on small language models (SLM) than big ones. Why? Because it's easy. It only takes a moderate workstation to train an
284.
by oofbey 11mo ago
Just our understanding of it. That’s flipped multiple times in my lifetime.
285.
by oofbey 11mo ago
That limitation only counts for visiting other galaxies. Travel within the galaxy is always possible, regardless of the universe’s expansion. And Alpha Centauri is super close, even within our galaxy.
286.
by oofbey 1y ago
+1 this. Also, when I read about the language features which make Python intrinsically slow, generally I think "I never use that." e.g. operator overloading meaning you need to do all these pointer dereferences just to add two nu
287.
by oofbey 1y ago
Super inspiring. A lot to read between the lines. Probably fairly introverted - prefers to be by himself than joking with coworkers. But not so much so that he can’t. He’s just really driven to be creative. And found a way, even though life
288.
by oofbey 1y ago
I think we both love Math. But I respectfully disagree. These are games. In a game like Portal, standard intuition about geometry mostly applies, as long as there isn’t a portal in the portion of space you’re considering. The boundaries bet
289.
by oofbey 1y ago
I’ve also had extremely poor luck getting any LLM agent to go through a long list of repetitive tasks. Don’t know why. I’d guess it’s because they’re trained for transactional responses, and thus are horrible at repute anything.
290.
by oofbey 1y ago
By that logic, the game Portal is non-Euclidean. Which might be true in a certain interpretation. But nobody looks at portal and thinks it’s anything but a normal 3D world with portals. Here the geometry is very unusual. But using the exist
291.
by oofbey 1y ago
Sure. But assembly implementations by definition are not portable. And I don’t know what it takes to write a formally verified library line this, but I bet it’s very expensive. In contrast, a rust implementation can be compiled into many ar
292.
by oofbey 1y ago
Latency matters for the autocomplete models. But IMHO those suck and generally just get in the way. For the big agentic tasks or reasoned questions, the many seconds or even minutes of LLM time dwarf RTT even to another continent. Side note
293.
by oofbey 1y ago
I’ve used MBP for many many years, but recently bought an MB Air. I slightly miss the extra ports. I love how much lighter it is. I never notice a speed difference. I’m always ssh’d into a Linux box if crunching any real data, and for UI s
294.
by oofbey 1y ago
Yup. The irony is that the packages which are difficult to build are the ones that most benefit from custom builds.
295.
by oofbey 1y ago
Thanks for that context. Why would I want to use it instead of git?
296.
by oofbey 1y ago
But you know they’re filling patents on that!
297.
by oofbey 1y ago
“Our protection mechanisms, to validate and block any erroneous deployments, failed due to a software defect which allowed the deployment to bypass safety validations.” Very circular way of saying “the validator didn’t do its job”. This is
298.
by oofbey 1y ago
I followed much of this path as well, but with conda (and mamba) in the mix. They all had fairly obvious flaws that uv doesn’t AFAICT. Poetry was the worst for me. It doesn’t even try to manage the Python distribution, so it’s only a partia
299.
by oofbey 1y ago
Have you found it easy to write rust modules with python interfaces? What tools do you recommend?
300.
by oofbey 1y ago
It’s a clear signal the paper is gonna be hard to read. It takes a ton of work to compress complex ideas down to 8 pages for a conference paper. But that work makes it easier to understand. This paper did not do that work. In fact it seems

More ›