LLM-based coding is enabling so much! The crazy weekend project now can have compilation to native code and web assembly, allow server-side or client-side rendering, manage multiple types of persistence, include adaptive compression, and do all of this without breaking a sweat.
I've found Claude Code with Opus 4.5+ to be excellent at generating test cases that exercise the different features, and even push into the edge cases. You sometimes need to nudge it into generating more convoluted cases when necessary, but then it is just nudging. I now routinely generate more LOCs of test cases than actual core code, while I used to only write very limited test cases just for the most complex areas amenable to automated testing.
I've been successful at using Claude Code this way:
1. get it to generate code for complex data structures in a separate library project
2. use the code inside a complex existing project (no LLM here)
3. then find a bug in the project, with some fuzzy clues as to causes
4. tell CC about the bug and ask it to generate intensive test cases in the direction of the fuzzy clues
5. get the test cases to reproduce the bug and then CC to fix it by itself
6. take the new code back to the full project and see the issue fixed
All this using C++. I've been a pretty intensive developer for ~35 years. I've done this kind of thing by hand a million times, not any more. We really live in the future now.
Oops, sorry about carelessly throwing the "cringe" label at that. Thanks for the transcript which allowed me to enjoy the content, which I did find very interesting.
I haven't watched the video so I am not sure how he actually talks, but what read cringe to me was things like the following paragraph:
"Stephen Robinson: Yeah. So let’s, let’s break it down into address generation versus execution. So, when you have three load execution ports, you need three load address generators. And so that’s there. On the store side, we have four store address generation units. But we only sustain two stores into the data cache."
Which reads weird. "let's" repeated twice, probably a stutter, could be transcribed just once. The "So" or "And so" the interviewee uses all the time at the start of sentences can also be removed for clearer and easier reading most of the time, without loss of meaning. Some sentences can almost be removed completely as they provide no actual information. The previous paragraph could be transcribed like this:
"Stephen Robinson: Let’s break it down into address generation versus execution. When you have three load execution ports, you need three load address generators. That’s there. On the store side, we have four store address generation units. But we only sustain two stores into the data cache."
I hesitate to remove "That's there." so I left it. But everything else I removed, it makes it clearer, and I think I'm not being unfaithful to the original. Removing the duplicate "let's" is a given as it's normal to stutter when speaking, but you don't really want to transcribe that unless the goal is to transcribe the talking imperfections we all have. And all the other things I removed, "Yeah", "So", "And so", are basically the same type of thing.
I thought this was automated because it had so many of the meaningless go-to words and hesitations from the original. Now that you mention it, automated transcription would probably never have produced something this good. And otherwise we are talking about stylistic preference here, always subjective -- although I'd definitely prefer the style of transcription suggest here.
Thanks again. I read chips and cheese with interest, quite often, and enjoy it quite a lot. Keep up the good work. And sorry for the careless put-down.
He is no Jim Keller, and the mostly[1] automated transcript makes it read cringe, but it is not at all devoid of content.
Some examples of very interesting, non-obvious content:
* Even if store ports are kept fixed (2 in his example), adding store address generators (up to 4 in his example) actually improves performance, because it frees up load port dependencies.
* Within the same core, they use two different styles of load/address address contention mechanisms which he describes as two tables, one with explicit "allows" and the other one with explicit "denies" -- which of course end up converging (I understand it refers to two different encodings which vary in what is stored).
* Between cores, they have completely separate teams which reach different designs for things like this.
* It was interesting to me to discover how isolated the different core design teams work (which makes sense)
* It was interesting to me to picture the load/store address contention subsystem, which must be quite complex and needs to be really fast.
And I stop listing, re different types of workloads, gaming workloads being similar to DB workloads, and even more similar between them than to SPEC benchmarks and so on.
Just go read the interview if you're interested in CPU design!
[1] mostly automated: at least the dialog name labels seem to be hand-edited, as one of them has a typo
Does anyone know about similar programs for non-US residents? I understand the original author is Canadian, but it seems they got granted a "special exception" that is unlikely to be available to, say, people from the EU (my case), or other countries.