f_devd

49 ms·

271.
by f_devd 4y ago
I agree, although power is also quite well managed in USB PD since both devices know what is supposed to come in, the cheapest chip that I found from TI[0] has: > Integrated reverse current protection, undervoltage protection, overvoltag
272.
by f_devd 4y ago
Do you suspect it's primarily the 'prestige' that was missing or the analysis (or maybe even something like layout), which made the initial paper submissions fail? If papers are rejected purely on prestige rather than merit t
273.
by f_devd 4y ago
Using any current architecture it is infeasible to do backprop (training) due to the massive communication requirements. Inference is possible to do in sharded way but still not as practical as just loading the model weight that are needed
274.
by f_devd 4y ago
Prolonged unsupervised physical access is usually already seen as a compromise. Regardless although there is a lot more publicity on Apple's "online attack" difficulty, it's mostly the same story on any semi-recent versi
275.
by f_devd 4y ago
Encrypted at rest does not actually do anything against interoperability of protocols (as I put in my comment), a secure element/coprocessor is nice but still does nothing against compatibility. Even if the entire protocol is somehow i
276.
by f_devd 4y ago
What security guarantees do Apple currently provide? I can't imagine much more than E2EE & maybe encrypted-at-rest (which is not a protocol-level feature anyway).
277.
by f_devd 4y ago
Could also be Sirens[1] which was all about sin as activation, while sinusoidal positional encoding are (iirc) just from the original transformer paper. [1]: https://arxiv.org/abs/2006.09661
278.
by f_devd 4y ago
> Toyota Corolla is related to a BMW car. The analogy is somewhat accurate, but also moot, since within the ML community "ChatGPT" can be used either as the product or the method (more specifically called RLHF) somewhat interch
279.
by f_devd 4y ago
I wonder if you could get a better pixels/bit ratio when using DCT/2DFFT based encoding since you'd still encode lower frequency data but it would be in a format that compression algorithms would also try to maintain.
280.
by f_devd 4y ago
Depends on how you link rust, using PyO3 makes it arguably easier to do link python code to rust than any C construction I could think off. Linking a single C file into python is quite difficult (if not using JIT like cppyy) because you nee
281.
by f_devd 4y ago
> Doing several tests a day on someone could create a lot of false positives! Only if they don't adjust the threshold/filter, properly calibrated you'd expect false positives to be reduced.
282.
by f_devd 4y ago
Interesting that it's both patented and under LGPLv3/GPLv3 (unclear which applies), seems like a way to avoid non-FOSS/proprietary implementations although I would think releasing it as regular GPLv3 would already serve as pr
283.
by f_devd 4y ago
Seems like primitives other than square & hexagon don't really work (not even triangle)
284.
by f_devd 4y ago
Learnable activation functions are a thing famously Swish[0] is is a trainable SiLU which was found through symbolic search/optimization [1], but as it turns out that doesn't magically make make neural networks orders better. [0]:
285.
by f_devd 4y ago
Although the article is recent the paper from the article has been available on preprint/arxiv since June 2021[1], implementations for pytorch & tensorflow are also available[2] for those interested. [1]: https://arxiv.o
286.
by f_devd 4y ago
They do have a IPv6 address under Linux & Others if that's what you mean
287.
by f_devd 4y ago
It depends on your definition of cheating, but this is definitely not "using humans to answer all questions one could ask". Rather it's a way to tune language models to be more like assistants rather than "most likely co
288.
by f_devd 4y ago
Generally it's without weights, but MusicLM is also a WIP. More mature implementations have descriptions on how to train them and follow ups on small scale/crowd-sourced experiments & research[1]. [1]: https://githu
289.
by f_devd 4y ago
I've actually been working on a Unicode only code prediction LM, and it already works pretty well the main issues with Unicode models as also found in the article is the large sequence length required compared a sentencePiece or BPE to
290.
by f_devd 4y ago
Very cool project but knowing how much access adb gives to your phone I wouldn't trust it unless you're self-hosting.
291.
by f_devd 4y ago
Certainly is a possible outcome although there are a few problems with the current paper as I see it: * Quite slow to execute (somewhat inherit to diffusion models) * Requires a lot of human data which increases dev time, since it needs to
292.
by f_devd 4y ago
I had exactly the same thought on the DAG structure and tried to create it in Obsidian using some linking magic but found it really clunky, currently working on making something myself using CRDTs, probably not for a wider audience though.
293.
by f_devd 4y ago
You might be right, although something to note is that xdelta 1.x is referenced while xdelta 3.1 is what gdelta claims to beat.
294.
by f_devd 4y ago
I've worked on (alleged) sota delta compressors before and was suprised how much code is needed here to enable OMP parallelism & LZMA compression. It could also just be the difference in delta technique, I will definitely look thro
295.
by f_devd 4y ago
A similar paper showed this for Language modeling and vision back in 2021: https://arxiv.org/abs/2105.08050
296.
by f_devd 4y ago
I think it depends on the structure, I find the structure of GP always very engaging: X (technical or not) can be represented as structure Y (math/cs). Side note: I suppose this would be the informal version of type/category theor
297.
by f_devd 4y ago
AMD had already done that with their 3D V-Cache, and to your point of SRAM & Logic not scaling on the same process they have also been working on that with the just released RDNA3 GPUs (The I/O-die is 7nm and the Logic die is 5nm I
298.
by f_devd 4y ago
Maybe in your case self-hosted/on-premise OnlyOffice is an option[0][1], but as I implied earlier the main issue currently isn't that there aren't alternatives for each individual service but that often a combined package is
299.
by f_devd 4y ago
I'm not sure why there is such a doomer sentiment (mostly from the US community but also some EU) about stepping away from Office 365. There are already existing replacements which do comply with GDPR for all of their service (modulo a
300.
by f_devd 4y ago
It seems about as accurate as GPT-3. Electromagnetic flux: > Album by OMD from 1980 Real band, fake album. I'm guessing it gets confused since I'm combining 2 topics, but it prefers making something up rather than combining the

More ›