rfoo

30 ms·

451.
by rfoo 3y ago
> If your user env is compromised its too late How to recover from a "my env was compromised" situation? Do you recommend just say good bye to this world and commit suicide? If not, this change makes the recovery easier: I just
452.
by rfoo 3y ago
100%. You can't be in the so-called "talent pool" without being able to speak good English.
453.
by rfoo 3y ago
But using a 8x smaller model also does not result in an 8x growth in the error, too.
454.
by rfoo 3y ago
> Does quantization work with most neural networks? Yes. It works pretty well for CNN-based vision models. Or rather, I'd claim it works even better: with post-training quantization you can make most models work with minimal precisi
455.
by rfoo 3y ago
> zfs fs on a wd black nvme ssd Why? IIRC running ZFS on NVMe SSDs limit their performance seriously. With sufficient queue depth modern SSDs can easily get up to 1mln+ IOPS and on ZFS I can barely get 100k :(
456.
by rfoo 3y ago
Good job writing this up! Quick question: for "post"-exploitation, once you had a KSMA-like primitive why would you choose to still do the modprobe_path dance, and introduce a pid bruteforce for fileless :(, instead of e.g. patchi
457.
by rfoo 3y ago
It makes your point even stronger, but > a generic Linux LPE pays 50 K$. At normal experienced exploit developer costs that buys you ~200 person-hours. $250 per hour? I must be living in the wrong part of the world then, where can find a
458.
by rfoo 3y ago
Acero is indeed plagued by "Java syndrome". However honestly Velox looks better, it's not as good as ClickHouse but working on it does not leave a "bruh Apache Arrow bruh shared_ptr everywhere bruh" bad taste in my
459.
by rfoo 3y ago
Check if it was ever appeared in gharchive [1]. [1] https://www.gharchive.org/
460.
by rfoo 3y ago
Since your repo's name is public now anyway, could you please help us and post it here? I'm really curious about what happened. Since public GitHub activities were archived [1], if you post it here we can check if it were ever pub
461.
by rfoo 3y ago
> a tiny glitch That's not "a tiny glitch". What you described is "GitHub may show your private repo to strangers randomly" which is a even more serious issue than they appearing in an archive some time later.
462.
by rfoo 3y ago
Ah, I meant Java code manipulating off-heap memory. For Java I don't think there's a way to mix managed and unmanaged code like you do in C#/C++ CLR.
463.
by rfoo 3y ago
Plus that it's easier to have unmanaged code in C# than in Java.
464.
by rfoo 3y ago
TBH this sounds more like perf's implementation is bad. I'm waiting for this to happen: https://github.com/open-telemetry/community/issues/1918
465.
by rfoo 3y ago
I believe for a developer tool to success, for the most common thing to do there has to be at least three ways an engineer may misuse your tool and still get it "done" (by leaving non-obvious tech debts behind). This is true for g
466.
by rfoo 3y ago
The thing I hate most about a Java codebase is 100 separate .java files with 30-50 lines each and somehow it indeed worked but I have no idea where to look at if I want to find out how is something implemented.
467.
by rfoo 3y ago
> And if its pseudo-random then it is predictable, and if its predictable then we can exploit that to great effect This is an interesting hack. However, an IOP is an IOP, no matter how good you predicted it and prefetch it so that you hi
468.
by rfoo 3y ago
> but could it affect the availability of another Availability is a different beast than durability. I think people are paranoid here about durability instead of availability. S3 advertises four nines availability and 12 nines durability
469.
by rfoo 3y ago
Most of developers are aware that username or password are PII and if they log it they are likely to get fired. Meanwhile our HTTP servers happily log every URI it received in access logs. Oh, and if you ever send a link in non E2EE messeng
470.
by rfoo 3y ago
> same weight class as a MacBook Except that there are plenty? As long as you avoid Dell it's easy to find a good deal. Oh, and I prefer plastic. Aluminum adds weight for nothing.
471.
by rfoo 3y ago
Thanks! It does this according to doc: This method adds the specified amount to the years field in three steps: * Add the input years to the year field * Check if the resulting date would be invalid * Adjust the day-
472.
by rfoo 3y ago
> Only other edge case is if someone takes a 2024-02-29 timestamp and modify the year part, without using date library to do so. Date library does not help here, for example, both Python and Java's date library [0][1] have only &quo
473.
by rfoo 3y ago
> came in well below cost Could you please suggest some read on why equipments are considered coming in well below cost? Haven't heard this before.
474.
by rfoo 3y ago
Agreed. Though in the last 6 years I've seen at least one case where truly deterministic builds mattered: A performance bug only happens when a malloc() during init was not aligned to 32 bytes, glibc on x86_64 only guaranteed 16 bytes,
475.
by rfoo 3y ago
This is wrong. The paper described that their activation is in int8 during inference. That being said, before-LLM-era deep learning already had low bit quantization down to 1w2f [0] working back in 2016 [1]. So it's certainly possible
476.
by rfoo 3y ago
Interesting, how do you use -0 in the add, then? Is -0+1-1 a 0 or a -0? > Could the additional -0 carry some pseudo-gradient information It looks like training was done on fp32 or bf16. Low-bit quantization is approximated with STE durin
477.
by rfoo 3y ago
Already happened years ago. They advertised TOPS for int8/int4 [0], and with 50% sparsity [1]. [0] low-bit CNNs worked pretty well actually. [1] Totally useless marketing snake oil.
478.
by rfoo 3y ago
Xeon Phi bans were targeted and successful. It shows that you don't need to issue a country-wide ban to prevent China from using a chip to build their supercomputers. And then, Sunway happened, it's built for military-use superc
479.
by rfoo 3y ago
I don't know. I thought Gina specifically said the recent ban is for AI. [0] > "What we cannot allow them to ship is the most sophisticated, highest-processing power AI chips, which would enable China to train their frontier mo
480.
by rfoo 3y ago
Until they were acquired by BRCM. Then oh well, you know, typical BRCM behaviors awaits.

More ›