Understanding Computer Memory Architecture and SSD Internals(codingpirate.com) |
Understanding Computer Memory Architecture and SSD Internals(codingpirate.com) |
It's incorrect, and simply wrong.
Example: When using a precomputed lookup table for the AES S-Box, ROMs are designed to create memory banks for faster arithmetic operations. During the design phase, you must make careful trade-offs regarding how much space to allocate for storing precomputed values. From the start, you need to floorplan how to write these precomputed arrays to achieve realistic space allocation. For this reason, algorithms like Canright's Boolean network are more suitable for compact design and are more efficient at reducing physical memory area.
It was my mistake to write it as general-purpose, and my wording was confusing. Thank you for highlighting this. I will read the article again and improve the writing.
"Occasionally, a designer will allow the CPU address to propagate through the address buffers to the bus at the beginning of every cycle, whether the cycle turns out to be a read hit or a read miss. The cache is designed in such a way as to initiate a main memory access at the same time as the cache access, rather than to wait for a cache miss before commencing the main memory access. This approach can shorten the main memory access time in the event of a cache miss and will serve to improve the operation of singleprocessor, single-tasking systems, but can work to the detriment of multitasking systems and multiprocessing systems, since the cache will allow the CPU to waste considerable amounts of main memory bus bandwidth. Such caches can be referred to as look aside designs. Since the processor has the option of going to either the cache or main memory at all times to request data, the look aside cache can be an add-on to the computer system."
Because the book mainly focused on the older i386 processor memory design, I could have been clearer when describing common characteristics. I updated the common characteristics to include "Higher data-rate processing," which is relevant to the Prefetch architecture and its intent to support more data bits.
(Don't shoot the messenger, just trying out Tangram for the first time.)
The text does not have any of the superfluous words typical for Claude/OpenAI talk.
It is written like an answer to "Explain to me like I am 5", but it contains appropriate information for such a context.
Only towards the end of the article, there are some signs that indicate AI editing, like a too frequent use of the word "simply", e.g.:
> "The one-clock-cycle delay introduced by an RDIMM is not simply a penalty. It results from pipelining the memory interface. "
This would have been better just: "The one-clock-cycle delay introduced by an RDIMM results from pipelining the memory interface."
Overall, it seems that the AI contribution was minimal.
The problem is that for scientific/technical texts it will always be impossible to determine whether they are written by a human or by an AI, in the case when they are correct.
For now, we detect easily the AI generated texts even for scientific/technical information only because they are bad, by including a lot of non-necessary words or by making unexpected mistakes. When the AI-generated text describing some scientific/technical fact will be flawless, it will not be distinguishable from what a competent human will say about the same content.
TFA might have been 100% AI-generated, but in that case it was a better AI than what has generated the immense quantity of slop that I have seen on the Internet lately, which is very easily recognizable as AI-generated, after reading just a few sentences, unlike the article linked by this thread.
What is absolutely certain is that about the first 100 words of TFA it is impossible to distinguish whether they have been written by a human or by a very good AI. If Pangram says that it is 100% confident that it was an AI, that is proof that Pangram is an 100% unreliable tool, so anyone who uses it makes a big mistake.
(Worth re-reading in the future!)
For false negatives, that depends on the type of article at hand. If an error leads to me spending 2 minutes reading an article that ended up being not worth reading I'd be fine with an order of magnitude more than if false positives happen on articles that take 30 minutes to go through. Maybe on the order of 1-3% overall?