Yes, it IS confident of its weights & biases... but, if you keep at it, Claude WILL find the smoking gun, eventually. Even a broken watch is correct twice a day ( unless it's a digital watch, without a battery, in which case, it's just broken... )
But seriously -- newer Claude (and OpenAI and Google and ???) models DO find the smoking gun, if you let them keep going until they reveal the weird chain of events that leads to a bug. I was seeing the most obscure UART driver bug, where it would work at 1,500,000 baud (!) but fail by only outputting the 1st char at 230.4k and 460.8k -- and it was due to a very narrow race that would check the buffer, if not full, insert a character, and return BUT sometimes the TX Complete interrupt would happen between the check and the insert, and something else would insert, and then - buf overflow. At 1,500,000 the other process didn't have time to do that phantom insert. ANYWAY, Claude found this and proposed a fix -- simpler: spins on IRQ-protected buffer empty checks.
I'd hate to think how long it would have taken me to find that.
And THAT's the problem -- of course a human CAN find it, with sufficient focus and time; I'm sure you've found a complicated bug pretty easily sometimes, by sheer luck or good engineering instinct.
BUT, it seems to me, as human, we are capable of creating potential execution paths that EXCEED our ability to EVER figure it out -- due to not being smart enough, not enough time on the problem, or something makes it economically unfeasible.
THIS is where LLMs shine -- let 'em bang at the code for as long as it takes.
The recent Mythos bug-finding explosion I think is proof of this conjecture. I think of it like a chessboard, where a machine really can look at all possible execution paths, and locate obscure bugs; a human programer (akin to a chess program) is doing 'alpha-beta pruning' of what's likely, and only after that list is exhausted are the really weird possibilities examined.
LLMs are our friends. And, as for "WTF did the LLM just do" when it generates code? I always include the instruction "For this code you just wrote, use Best Practices to document this code, function by function and class by class, and when necessary, line-by-line, so that a junior SW developer can completely understand how this code works, using the documentation standard we use (e.g. Doxygen)."
I have also used this technique to learn new languages, or explore ones I only know a little -- it has been a godsend for leveling me up on common lisp, for example. "Give detailed comments explaining what the code is doing, assuming the code reader is fluent in C and Python, and use analogs when possible." Stuff like that.
The MOS 6502 was introduced at the WESCON tradeshow in September 1975 and sold for $25 quantity 1. They had a transparent vase full of them, 'proving' they did volume -- but those were all chips that failed post-manufacturing testing, except for the very top layer. Still...
I take the previous point about missing MMX / Vector instructions. Presumably, that's a straightforward optimization, when code is vectorizable? That is, yes, it would be nice to have sections on vector instruction use to demonstrate some appropriate algorithms; point taken.
I think, tho, the fundamental reason is: he specs the machine. It does exactly what he says it does. No edge cases, no surprising behavior that LLVM IR or WASM might have; also, I'm sure he spec'd it to make the algorithms he implements elegant or at least more understandable.
I suspect part our views comes from the purity of Mathematics (Knuth) and the actuality of Computer Science. I'm even sympathetic to your view, because then we could verbatim copy Knuth's algos to WASM etc and run them, and we get a twofer: a deeper understanding of the algo AND a deeper understanding of the underlying V-ISA (PTX, BEAM, WASM, LLVM IR, etc).
I think the learning / understanding of the algos Knuth presents partially comes from us Playing Computer, and manually going through the code. It's that pedagogy comes first, I think.
Next time I run into Don, I'll ask him about this. 'cause you're also correct about L1...Ln caches affecting performance drastically.
(And, yes, I had wondered why Don did MMIX he didn't create a "RISC-V" before there was a RISC-V; that is, why didn't he create the Next Great RISC Machine? He said he talked with John Hennessy and Richard Sites. This was in 1990.)
MMIX uses register windows to make stack frame pointer offsets unnecessary when referring to PUSHed arguments. Don is trying to make the algorithms understandable and correct, and by hiding some details that are handled efficiently by compilers (keeping track of FP and offsets), it benefits the human reader.
This series of books is affectionately dedicated
to the Type 650 computer once installed at
Case Institute of Technology,
in remembrance of many pleasant evenings.
MMIX is for all you youngsters who think RISC is all the rage ;-) and I think he does an admirable job creating a fully-defined machine that does use more modern hardware techniques. The fact that he fully defines his underlying machine is exactly correct, because it lays the foundation for precisely expressing the algorithms, and for giving Time and Space (runtime) estimates.
I believe it's fundamentally incorrect to think of these abstract machines as 'assembly language' but rather, I think, they define a stable foundation onto which accurately described algorithms can be expressed. You're supposed to 'play computer' and follow along -- step by step -- to understand the deep details of the algorithms.
This is exactly correct. Low 10s of kHz is quite functional for machines moving in human space / speed.
If one is trying to do some assembly line (max # of operations per second), the power requirements alone get hard to manage. And then you're managing back EMF, eddy currents, heck, air resistance!
My rule: have dedicated low-level hw provide smooth PID response, mostly on the P term; and have a higher-level control produce the setpoint. Faster response means less need to rely on I or D terms as much (just because delta-T is so relatively small).
Wow, the Python folks do not want to entertain different goals for the language, it would seem. I still remain interested in whether it makes sense to add features to Python anymore, given that LLMs write all the code. Unless, as I said, such features enhance human comprehension (since we have to review the code the LLM produces).
Absolutely serious question: is anybody actually writing code anymore? (Reviewing it, architecting it, sure -- we do that). But writing code? If NOT, then seems to me, what we want in our LLM output is CODE THAT IS TRIVIAL TO UNDERSTAND (I apologize for shouting from the rooftops.)
I'm not sure adding 'features' to Python anymore makes sense - UNLESS those features help humans understand LLM code. Part of the problem is, of course, that LLMs haven't been trained on the latest-and-greatest, so they won't output any of it (even new training events won't capture the latest very much, since there is so little of it relative to what is already out there).
But again -- do new features help HUMANS understand what the LLMs produce? If not, seems to me ... new features merely add complexity for no apparent gain. (Or am I confused about this?) THANKS for any helpful opinions.
If we used it more than just "ok look it works, we'll standardize on this internally" we'd have been happy to buy it, as we all other open source that has a donation option. We were just getting our feet wet, tho, and that's where it seemed burdensome; heck, if it phoned home every time it was used, it would have noticed we only used it about 3 times!
I had heavily used PySimpleGUI in various work projects, and one day, when I had to run some older piece of code I had not run in a while, I get a notice that PySimpleGUI won't work, because was free, but nobody paid, and so, good luck! So I was piping mad, paid the 3-year or whatever the max license fee was, received a code, and THEN I was able to get my stuff to work, like it used to.
LESSON: N E V E R Use code that can "stop working" until you pay ransom. N E V E R.
At this point, it's irrelevant, because the LLMs can replace PySimpleGUI with PyQt etc so --- thanks but no thanks. I did like it because you could throw up something around a CLI and it looked at least presentable. Now, since 2025, nobody codes anymore, so ... seems to me, this PySimpleGUI 6 is just a bit of history.
Excuse my ignorance, but does this functionally mean we can treat this as a 'microkernel' a la minix? I always liked the 'tiny protected subsystem' in Ring 0, then a Ring 1 for Drivers (which are restartable, and dynamically loadable), then one or two rings for User processes (maybe Ring 2 for 'ls' etc and Ring 3 for typical user processes).
I am also curious: What hardware enhancements would benefit 'lightweight, kernel-enforced isolation' ? Do we need memory tags? HW Capability Lists? ?
( I believe we've concentrated far too much in making "damn fast pdp-11s" with our hardware advances, and far less on building Reliable Systems -- even if a few percent of peak possible performance is consumed by extra HW. )
It's possible. I don't know. My tone comes off as support Elon, and I do not, at all. I've seen first-hand almost all of these tactics while I was at <Elon Company>. I'm observing that some people seem to do OK at Elon's companies, and for many years, and never seem to get the boot or be abused in other ways. Therefore, Elon is probably not quite as bad a manager as he is made out to be. This is all I am saying. Since I have firsthand knowledge, I believe my opinion has value. Those that disagree? Show me your Source of Truth. Thank you.
I have always disliked the := as assignment operator convention. In these declarative languages, assignment is done frequently. There is little cognitive load to using '=' as assignment, although perhaps a bit jarring for math folk.
<- is somewhat better, but, again, for such a common operation, a single character is just more convenient. Sure, we could have editors that turn "=" into := or <- but now we're getting too fancy especially for something pedagogical.
I also don't mind the -> for C pointers; and certainly don't mind the <= >= or even == conventions (although at least today's compilers warn when they see "if (a=b) ...".
Ultimately, humans won't be writing code anymore anyway ( ;-) ?) so maybe the issue is entirely moot.
But seriously -- newer Claude (and OpenAI and Google and ???) models DO find the smoking gun, if you let them keep going until they reveal the weird chain of events that leads to a bug. I was seeing the most obscure UART driver bug, where it would work at 1,500,000 baud (!) but fail by only outputting the 1st char at 230.4k and 460.8k -- and it was due to a very narrow race that would check the buffer, if not full, insert a character, and return BUT sometimes the TX Complete interrupt would happen between the check and the insert, and something else would insert, and then - buf overflow. At 1,500,000 the other process didn't have time to do that phantom insert. ANYWAY, Claude found this and proposed a fix -- simpler: spins on IRQ-protected buffer empty checks.
I'd hate to think how long it would have taken me to find that.
And THAT's the problem -- of course a human CAN find it, with sufficient focus and time; I'm sure you've found a complicated bug pretty easily sometimes, by sheer luck or good engineering instinct.
BUT, it seems to me, as human, we are capable of creating potential execution paths that EXCEED our ability to EVER figure it out -- due to not being smart enough, not enough time on the problem, or something makes it economically unfeasible.
THIS is where LLMs shine -- let 'em bang at the code for as long as it takes.
The recent Mythos bug-finding explosion I think is proof of this conjecture. I think of it like a chessboard, where a machine really can look at all possible execution paths, and locate obscure bugs; a human programer (akin to a chess program) is doing 'alpha-beta pruning' of what's likely, and only after that list is exhausted are the really weird possibilities examined.
LLMs are our friends. And, as for "WTF did the LLM just do" when it generates code? I always include the instruction "For this code you just wrote, use Best Practices to document this code, function by function and class by class, and when necessary, line-by-line, so that a junior SW developer can completely understand how this code works, using the documentation standard we use (e.g. Doxygen)."
I have also used this technique to learn new languages, or explore ones I only know a little -- it has been a godsend for leveling me up on common lisp, for example. "Give detailed comments explaining what the code is doing, assuming the code reader is fluent in C and Python, and use analogs when possible." Stuff like that.