gcr

18 ms·

271.
by gcr 10mo ago
What do you suggest instead? Handrolled code with “import openai”? BAML?
272.
by gcr 10mo ago
google research already does this by separating corp and prod quota scheduling, i'd be shocked if other big labs don't do this also
273.
by gcr 10mo ago
Cloud Run tries to be this but every service like this has quirks. For example, GCR doesn’t let you deploy to high-CPU/MEM instances, has lower performance due to multi-tenant hosts, etc
274.
by gcr 10mo ago
Note the timeline. Nothing's changing soon. > Next year, you’ll be able to opt-in to 45 day certificates for early adopters and testing via the tlsserver profile. In 2027, we’ll lower the default certificate lifetime to 64 days, and
275.
by gcr 10mo ago
Links to some articles are broken, eg. https://www.mikkelaaland.com/sweat-bathing-and-the-body.html gives a cert error on iOS
276.
by gcr 10mo ago
Oh. TIL, sorry for pushing back
277.
by gcr 10mo ago
I have shitty eyesight and old is the only version that’s usable to me because text zooming works far better in the old layout. Are you using a screen reader by any chance?
278.
by gcr 10mo ago
Academic papers should be neutral when dissing previous work imo. "[Previous work] is just bad [previous other work]" isn't a professional way to talk about the merits and drawbacks of competing approaches, but boy does it ge
279.
by gcr 10mo ago
I had the misfortune of visiting an Amazon Go store. They charged me for items that I never picked. No option to contest the receipt....until the "would you recommend a friend visit amazon Go" survey popped up. I responded negativ
280.
by gcr 10mo ago
lots of googlers use BYOD iPhones and the corp suite for this use case is fairly well-supported
281.
by gcr 10mo ago
i basically can't use the ChatGPT app on the subway for these reasons. the moment the websocket connection drops, i have to edit my last message and resubmit it unchanged. it's like the client, not the server, is responsible for w
282.
by gcr 10mo ago
Look, this is an earnest new author who isn’t from academia. Dogpiles are only useful if they include useful feedback. A professor once extended a similar kindness to me on a particularly rough draft of my own (very early) work, and it was
283.
by gcr 10mo ago
you’ll get there! :)
284.
by gcr 11mo ago
actually, here's a broader thought. since this approach only works for classification, why not make that the whole story and spin it as a positive? Call your approach a "classification foundation model" (for example) and say
285.
by gcr 11mo ago
agreed re: title/abstract stretching. good work stands on its own without needing hype. "we found a nifty way to distill llama-70b using a much smaller student transformer model; the key is using intermediate activation layers in
286.
by gcr 11mo ago
thanks for sharing! If I understand correctly, you're training a smaller model to approximate concatenate(layer[1], layer[5], layer[10], ...), using a loss function that combines reconstruction error w/ end-to-end accuracy. then,
287.
by gcr 11mo ago
do you mean latency on a color screen? (my experience with color eInk is that it adds quite a lot of latency)
288.
by gcr 11mo ago
Yes, that's what I'm saying. I'm agreeing with your response to amluto.
289.
by gcr 11mo ago
Yeah I was confused about that bit too. AWS and GCP's metadata servers know which instances were deployed, so they presumably have some way of verifying the instance's identity out-of-band, such as being tagged by an internal job
290.
by gcr 11mo ago
See there are users who like Liquid Glass, just as there are users who like TouchID. A lot of Apple’s best work turned out to be quite polarizing at the time. iOS 7’s design language was almost universally panned, but if it were “the wrong
291.
by gcr 11mo ago
Your workflow is fascinating! What languages do you work in, if you don’t mind me asking? Both git and jj have sparse checkouts these days, it sounds like you’d be into that Do you vendor the libraries you use? Python packages typically don
292.
by gcr 11mo ago
Keep in mind that you still need to occasionally run jj commands, even in collocated repositories Example: jj snapshots the working copy on any jj command but not any git command It’s a leaky abstraction
293.
by gcr 11mo ago
What if the sandbox needs access to the history? Imagine asking for help bisecting something, for example
294.
by gcr 11mo ago
Don’t you just want a clone in that case?
295.
by gcr 11mo ago
isn’t it strictly better to use your editor’s JJ support and ask Claude to use jj? (I personally think jj shouldn’t support collocated repositories, but happy to learn what others see in them…)
296.
by gcr 11mo ago
jj may use git as (one of) its backing stores, and its collocation offers some compatibility at the cost of important tradeoffs, but it isn’t intended to be a git frontend.
297.
by gcr 11mo ago
Why are you using git tools within collocated jj repositories? I never understood that use case tbh
298.
by gcr 11mo ago
The secret is held by the metadata server that the CI instance has access to Or: the deployment service knows the identity of the instance, so its secret is its private key Or, how PyPI does it: the deployment service coordinates with the t
299.
by gcr 11mo ago
And how is that exposed to the CI/CD? An environment variable? Some bespoke tool that the CI runs to read the secret from the Secure Enclave? Your approach boils down to “lets give each step its own access to its own hardware-protected
300.
by gcr 11mo ago
“Good CI systems shouldn’t support secrets, at most there should be [the most complicated secret support ever]” Let’s just call it secret support. I agree with your suggestion that capabilities-based APIs are better, but CI/CD needs to

More ›