adrian_b

22 ms·

361.
by adrian_b 3mo ago
Any manifesto of this kind unavoidably contains many naive parts. The reason is that it is always easy to notice what is bad around you, but it is very difficult to conceive what should be done about it. The paragraphs of the manifesto abou
362.
by adrian_b 3mo ago
I doubt very much that COVID over-hiring has any relevance today. COVID is now ancient history and there have already been several years of big layoffs when every time I hear the attempt of explaining them through COVID over-hiring. That ma
363.
by adrian_b 3mo ago
Niobium arsenide cannot replace copper in wires or cables, for two reasons, it is many times more expensive and it is fragile, so it cannot bend. Moreover, at normal sizes it is not a better conductor than copper, so it would have no advant
364.
by adrian_b 3mo ago
Before the second WW, mathematics was moving at much faster paces than after WWII, when many theories became more or less settled and only their practical applications continued to develop at a fast pace. The modern variants of some branche
365.
by adrian_b 3mo ago
When I was young, I have read more than half of them. They should better be read in order, starting with the theory of sets. I do not believe that they are useful for learning for the first time some part of mathematics, but they are useful
366.
by adrian_b 3mo ago
Any open-weights model that has ever been published can be run on consumer hardware, even on a mini-PC. The right question is which is the speed that can be achieved on a given hardware and whether it is high enough for the model to be usef
367.
by adrian_b 3mo ago
For USA, indeed there is no end in sight to the memory crisis, as long as its politicians work for companies like Micron, which obtains huge profits from the inflated memory prices. A few days ago, the representatives John Moolenaar (R-MI)
368.
by adrian_b 3mo ago
I assume that this announcement has been prompted by that of Moonshot AI, which has just announced a 2.8T parameter open-weights LLM, Kimi K3, to be published on Huggingface by 27 July. Now the response of Alibaba is that they will also pub
369.
by adrian_b 3mo ago
Easily. With a curve tracer you could measure the breakdown voltage of the transistor when connected with the emitter and the collector reversed and with open base, which is close to the breakdown voltage of the emitter-base junction. Knowi
370.
by adrian_b 3mo ago
> And yep, to preempt the inevitable comment: we used to have even more specialized devices, called "curve tracers", that were designed specifically for making V-I plots. They're more or less extinct now because SMUs can d
371.
by adrian_b 3mo ago
Zilog Z80 was orders of magnitude simpler, but its designers never did something so foolish. At that time, many people still used assembly language and they would have never accepted to write arithmetic expressions in the contorted way forc
372.
by adrian_b 3mo ago
In ISAs where this problem exists, i.e. in x86-64 where the carry flag is updated or not updated regardless of what happens with the other flags, the flags register is split into 2 separate registers, which are renamed independently. The 2
373.
by adrian_b 3mo ago
What you say about partially updated flag registers is right. Because of this, most modern ISAs take care so that the flags register is always updated completely. In legacy ISAs, like in x86-64 where the carry flag is updated or not updated
374.
by adrian_b 3mo ago
In modern OoOE CPUs the flag register is renamed, like all other architectural registers, so it does not hinder in any way the parallel execution. OoOE CPUs need hundreds of registers in order to not hinder the parallel execution, so even t
375.
by adrian_b 3mo ago
Limbs work well only when the operations with them are implemented with vector or matrix instructions. Using limbs is a workaround for the fact that most vector instruction sets also do not implement carry flags (an exception was the discon
376.
by adrian_b 3mo ago
Unfortunately the criticism from that link is absolutely correct. Moreover, at that link it is shown only the ugly way in which RISC-V does multi-word addition. Checking the standard arithmetic operations for overflow is much more horrible
377.
by adrian_b 3mo ago
During the eighties, there were a lot of people, including myself, who could read directly the hex codes of the Z80 instructions as easily as its symbolic assembly language. That was very handy for reverse engineering programs like the Micr
378.
by adrian_b 3mo ago
Making the parity flag not perfectly compatible was a wise choice. In legacy programs the parity flag was seldom tested and practically never after the instructions where in Z80 it behaved differently. This allowed the repurposing of the pa
379.
by adrian_b 3mo ago
You have exactly the same mental image if you start a pipeline in the right way, with a file redirection: "<input_file cmd1 | cmd2 | cmd3 ...". This adds the joy of doing only what is really needed, without extraneous effort. M
380.
by adrian_b 3mo ago
The simpler and better way to write your pipeline is: "< file.txt somecommand | othercommand | anothercommand" There is never any need for "cat". When such a pipeline is used repeatedly in a script, the time for execu
381.
by adrian_b 3mo ago
Yes, but "cat" is superfluous anyway. You can start your pipeline with a file redirection: " <input_file prog1 | prog2 | prog3 | ... | progN >output_file".
382.
by adrian_b 3mo ago
When first looking at the source code, I wondered why one would waste so much time to write 25k lines in raw assembly language, but then I saw that it was generated with Claude, for whom it does not matter much how expanded is the written t
383.
by adrian_b 3mo ago
In many countries, the loser pays all the legal bills. So if you have been wrongly accused, that may cost you nothing.
384.
by adrian_b 3mo ago
If you have monochromatic red and green light sources, like in a laser projector, you can reproduce very well the red-yellow corner, as mentioned by others. However, in CRT displays, the color of the light emitted by the red phosphor was ra
385.
by adrian_b 3mo ago
You can, but the reported speeds for such big models were in the range from a token per a dozen seconds to slightly better than 1 or 2 tokens per second.
386.
by adrian_b 3mo ago
Floating-point operations are stateful in any CPU made during the last 40 years. They depend on a set of configuration flags, which are stored in some internal register, e.g. the MXCSR register of Intel/AMD CPUs. The same program can p
387.
by adrian_b 3mo ago
The standard 6 relational operators that are sufficient for a totally ordered set are: <, >, <=, >=, = and <>. The last is "less or greater", as used in BASIC and SQL, instead of "not equal". In a parti
388.
by adrian_b 3mo ago
There is also NVIDIA Nemotron 3 Ultra, with 561B parameters, which was released a month ago. I do not know yet how smart it is, but the NVIDIA LLMs are very well optimized for fast inference (on their GPUs of course). Previously that was th
389.
by adrian_b 3mo ago
> Since most of the threads are idle waiting, your application needs many threads to optimally utilize the servers resources. I do not think that "many" and "optimally" belong in the same sentence. Instead of having a
390.
by adrian_b 3mo ago
> Annoyingly, though, there is no standard name for this category of control flow constructs. That is wrong. There is a standard name: FORK. In 1963-11, in "A Multiprocessor System Design", Melvin E. Conway introduced the terms

More ›