adrian_b

14 ms·

151.
by adrian_b 2mo ago
I completely agree with one of the comments from there: > But the fundamental conclusion is: the design was wrong. It should not have used mmapped writes. pwrite would have been far better. It really does not make any sense to use memory
152.
by adrian_b 2mo ago
They are selling it, not running it. It is just a dedicated computer system, which should be managed by its owner, like any other on-prem servers. I doubt that it has a good price/performance ratio, but it is a solution for those who f
153.
by adrian_b 2mo ago
Palantir already offers a "turnkey AI datacenter", i.e. a rack with "NVIDIA Blackwell Ultra systems with eight NVIDIA Blackwell Ultra GPUs and NVIDIA Spectrum-X™ Ethernet networking for AI training and inference". It is
154.
by adrian_b 2mo ago
> The model weights of GLM-5.3 will be publicly available soon in two weeks.
155.
by adrian_b 2mo ago
> The model weights of GLM-5.3 will be publicly available soon in two weeks.
156.
by adrian_b 2mo ago
It was a mediocre design, but it came with superb documentation for every bit of hardware and software in it, including the source code for its BIOS and all the hardware details of the peripheral and expansion interfaces. I wish this had re
157.
by adrian_b 2mo ago
Even the lowliest hardware could run this, but at an unlikely to be useful low speed, e.g. of 3 or 4 tokens per minute (by reading the weights from a couple of 4 TB SSDs for the BF16 model, or from a 4 TB SSD for the FP8 variant). The quest
158.
by adrian_b 2mo ago
What I found the most interesting, and unfortunately not at all surprising, is that the reasoning of the LLMs frequently contained much more useful information than the actual answers, because the answers were censored.
159.
by adrian_b 2mo ago
Yes, there are 5 Nemotron-3.5 models: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-... to be used for further training/fine-tuning. https://huggingface.co/nvidia/NVIDIA-Nemotron
160.
by adrian_b 2mo ago
I agree.
161.
by adrian_b 2mo ago
I doubt that anything could have changed, because OpenBSD has not implemented any new file system and a few years ago they have even removed the support for "UFS with soft updates", which is a file system that is not affected by p
162.
by adrian_b 2mo ago
File system fragility was indeed still an embarassment of many UNIX-derived file systems, up to twenty years ago and perhaps until more recently, but this is no longer true today. Even in the early 2000's, there was one UNIX file syste
163.
by adrian_b 2mo ago
Nothing can be as power efficient as an ASIC, which is designed for a specific purpose, instead of being a programmable device intended to be suitable for a large class of applications. A GPU is much more efficient than an FPGA for what a G
164.
by adrian_b 2mo ago
Not every tail call is for a loop. You can have a set of mutually recursive functions, which tail call each other. In C you can write state machines using "goto" (the implementations with "switch" are typically much more
165.
by adrian_b 2mo ago
That has become no longer true since last year. 20 years is too long in the past, so indeed computers were more expensive then. But if we go only 15 years in the past, to 2011, we find the same price per GB of DRAM as today, so that afforda
166.
by adrian_b 2mo ago
When you look on HuggingFace.co at the files of a model, for each model you will see a file "tokenizer.json". In that file you can see all tokens and their corresponding numeric codes.
167.
by adrian_b 2mo ago
US companies started by publishing their research, but then they have stopped abruptly, seduced by the mirage of heaps of money that could be obtained by having a monopoly on AI. During the last few years, the published research from USA wa
168.
by adrian_b 2mo ago
That is not exactly true, because for a long time tail call optimizations had a lot of restrictions in gcc, so they could be used only seldom. What is said in TFA is correct in the sense that only in recent years the support for tail call o
169.
by adrian_b 2mo ago
As a family name, it seems to be: 後藤 where the first character means "behind" or "back" and the second character refers to wisteria vines. (The first character is familiar to some of the martial arts practitioners, becau
170.
by adrian_b 2mo ago
While the first public release of Fortran was in 1956, its design started in 1954 and the GOTO statement was already included in the preliminary Fortran from 1954. Besides the inventor of parametron, there is also Kazushige Goto, who became
171.
by adrian_b 2mo ago
Elliott 803 used mostly transistors, but it indeed also used magnetic cores for some of its logic gates. The operating principle was different than that of the parametron, but the end effect was about the same. It used highly non-linear mag
172.
by adrian_b 2mo ago
During WWII, Germany used magnetic amplifiers in many different kinds of military equipment, because they had the advantage of being much more rugged and reliable than the electronic devices available at that time. After WWII, the technolog
173.
by adrian_b 2mo ago
Also with liburing you have zero-copy send and receive operations for normal protocols like UDP or TCP, but this requires a NIC that supports scatter/gather DMA (so that the packet headers go to/from kernel buffers, while the data
174.
by adrian_b 2mo ago
You may like more food that is cooked on an open flame, if you like frying or if you like burned crusts, but precision is not the reason. I eat only food that I cook myself from raw ingredients and I have switched years ago from cooking on
175.
by adrian_b 2mo ago
That is correct, DDR5 today is in real US$ slightly cheaper than DRAM was in 2011, 15 years ago. For more than 15 years in the past, the price in real US$ of DRAM per GB was always higher than today. During 2014, there was a brief price pea
176.
by adrian_b 2mo ago
In 1999, AMD Athlon became the first x86 CPU that was able to do both an addition and a multiplication during one clock cycle, for the 80-bit x87 floating-point numbers. The previous Intel CPUs of the P6 family, from Pentium Pro to Pentium
177.
by adrian_b 2mo ago
Engineering never went completely to Windows NT. Between 1997 and 2007, I worked at 3 different companies, in 3 different countries (2 in Europe + Israel) as Design Engineer in electronics. All the serious engineering programs for EDA/
178.
by adrian_b 2mo ago
The fact that sterile mosquito drops were stopped due to the pandemic, was the cause, together with the fact that after the pandemic no measures were taken to compensate this, even if it was obvious that they are imperatively needed. There
179.
by adrian_b 2mo ago
The author might be a moron, but this is not evidence of it. The 44.1 kHz sampling frequency is theoretically able to ensure a 20 kHz pass-band for audio signals. But this does not mean that this is true for an actual recording. A 44.1 kHz
180.
by adrian_b 2mo ago
Virgin vinyls always had a good enough audio quality. Their unavoidable defect was that the quality degraded continuously with the number of times you played them. When played on a cheaper turntable, even playing them only 3 or 4 times coul

More ›