Remote
G. Wozniak
@gwozniak@discuss.systems
mastodon 4.7.3Embedded compiler developer. I like debugging stories and computing history.
Posts are removed after a month. Nothing lasts forever.
And cats.
235 Followers
101 Following
50 Posts
Joined December 29, 2022
I maintain that the reason the LLM search stuff gets so much attention in enterprise is that enterprise search has been so bad for so long, just about anything would be better. The crass inefficiency of LLM-based search doesn't matter because it gives back semi-useful results, unlike pretty much *everything* that came before it.
This is not praise for LLM-based search. It's a indictment of previous search setups. (Looking at you, Sharepoint. Confluence. Jira.)
Replying to
@ricci@discuss.systems Cisco Duo absolutely is the best way to share a cat photo. I mean, the best way to communicate with all your friends is to send six digit one-time passwords.
Caya, roughly 12 hours apart. #cats #catsofmastodon
The "too big to fail" argument assumes it is a success.
Replying to
@jab01701mid@mastodon.social @dalias@hachyderm.io I don't work in web stuff at all, but I rail against this mentality all the time. Yes, I have a fast machine. No, I don't think it's a baseline.
I swear it's a mentality thing, where there are those like me who want to make products that allow people to enhance their lives, even if it's not groundbreaking. Then there are other who just want to make products in the belief there is great virtue in that product simply existing.
The latter never care about performance or resource usage.
Replying to
@sundogplanets@mastodon.social My rig probably isn't suitable for this, but I'm getting up and trying anyway.
Replying to
@regehr To me, this is the key point of contention when I talk about these things.
"Contrary to the current narrative that is being espoused, we didn’t necessarily need to write code faster. Especially code we didn’t fully understand, and particularly in huge swaths that we couldn't review in reasonable time frames."
For those of us that have serious problems with LLMs for programming, this part is important.
But for those who push them there is no issue. Generating more code is all that matters. The result can be bad or half-baked. It doesn't matter. All that matters is that something was produced and shipped.
Replying to
@regehr @aburka So my response could be *really* long but I'll keep it relatively brief.
My thesis (please don't read it) was about automatic codegen, leaning in the direction of what LLMs currently do. I thought it was the way to go. I'm much less convinced of that now, and was before the LLM stuff took off.
Whenever I have seen it used, codegen from specs (*) doesn't work out well. And that's with formality behind it. Codegen from informal specs crank up the ease so they have that dopamine hit, but it doesn't last. Everyone intrinsically knows this. Not everyone acknowledges it.
(*) Let's ignore the "code as spec" perspective here, where a compiler counts as an auto codegen tool.
From that perspective, I don't care for LLM autocomplete. It's far too good at disengaging you from the material. Much has been said about this already, and the studies are proving it right.
However, LLMs are an very interesting technology, from a purely technological standpoint. For that, the academic in me says they should be studied. You have to explore the edges. You can predict the broad direction of something pretty well, but the details matter. We don't know the details yet. And that's what academia is for!
And are LLMs good as spitting out a lot of code that kinda works? Yes. Yes they are. It's a reality that cannot be ignored. So that should be studied and taught.
That said, I really hope the teaching involves not just the technology. And this is where I probably diverge from others. Because continued use of the technology as it currently exists is supporting and empowering some of the worst people and systems in the world. If I were researching and teaching this stuff, I'd be very clear about what we know about them aside from their outputs. I think ignoring that aspect is a massive disservice, if not derelict of duty.
I continue to not use them and oppose their usage in most settings because of the massive problems that exist behind them. I'm still not really sure they can even be ethical, if "ethical" means "does not incur undue harm to others". What we have now does not qualify.
When I see or hear of anyone I know use our current crop of LLMs I get a little sad and lose some respect for them, but I also understand that convenience is a thing. I believe we (especially American-style "we") overvalue convenience for individuals, but that is another thread. I get that resisting this stuff is not for everyone and we all have different circumstances. And then there is the relentless pressure from bosses to use it. (Make sure you think about why that is!)
Anyway, my feelings on this are complicated even though I am vehemently against LLM usage. Most of my objection stems from factors that seem outside the technology itself but I feel, however, fundamentally define it. The utility of it (at some level, at least) cannot be denied. It should be taught and studied. But I sincerely hope it is done in a way that does not try to divorce it from these other factors and makes it clear that code is, ultimately, a liability. More of it is not what you want.
And yes, I'm disappointed some people are so excited about it.
Replying to
@regehr@mastodon.social I don't think there's a nice thing I can say about this and if I really let loose even Mastodon might ban me.
I know too many people excited about space stuff that give SpaceX a pass on basically everything.
Replying to
@dan@discuss.systems The political poll you didn't think would ever exist outside of prank calls.
Replying to
@SnoopJ@hachyderm.io @irene@discuss.systems To be fair, we are assuming that Pixel is happy with the configuration.
Replying to
Replying to
@LibraryAndTea@ottawa.place @mayintoronto@beige.party Same. I don't live in Toronto but I live nearby. Downtown is just a dumb place for an airport.
Replying to
@ricci Okay, so I have a program to do this, written in C. I ran it on the first N words on my Ryzen 7 5800. I didn't parallelize this. GCC 14 built at O3.
N = 100, 0.013s
N = 1000, 0.016s
N = 10000, 0.481s
N = 100000, 49.507s
The words file has 1.1M words. I'm going to let this run and see how long it takes.
I'm using mmap and friends and I've only got one call to malloc. Why not optimize a bit?
Replying to
@floe@hci.social @peter_mcmahan@mas.to @davidgerard@circumstances.run I am voting for this one because WTAF
"I tell him: AI didn't take our jobs. Greed did."
https://www.stvn.sh/writing/programming-still-sucks-fqffhyp
Replying to
Replying to
Replying to
Replying to
@davidgerard@circumstances.run Yes, it's funny how "nuance" only ever seems to go one direction.
Replying to
@regehr@mastodon.social @malwareminigun@infosec.exchange @jpetazzo@hachyderm.io @aburka@hachyderm.io True fact: I tried to use it once and I went back to crontab (and I don't know cron that well!)
Replying to
@regehr@mastodon.social @malwareminigun@infosec.exchange @jpetazzo@hachyderm.io @aburka@hachyderm.io Post-it notes work better!
Replying to
@regehr@mastodon.social @malwareminigun@infosec.exchange @jpetazzo@hachyderm.io @aburka@hachyderm.io It's like, hey, I just want to set a calendar reminder to ping me every 90 days to remind me check something. But it turns out that sucks so instead the zeitgeist tells me that I should use a web service for that.
Replying to
@regehr @malwareminigun @jpetazzo @aburka I hate to chirp about "worse is better", but it's hard to say it's not that sometimes.
Replying to
@malwareminigun @jpetazzo @regehr @aburka Just to be clear, I don't think expert level is needed either. It's just that the systems seem to be designed in a way that require it. And, from what I've seen at least, devs have to deal with a lot of systems all the time.
I guess I'm wondering why. I don't know the answer, nor am I trying to be snarky.
To take the crontab example, how did we get to the point where doing the seemingly simple thing of setting a timer requires knowing arcane syntax?
That's a line of questioning I'd be interested in studying, to be honest. It's clear that LLMs are helpful for what exists now. It's also pretty clear that LLMs are an incredibly inefficient way (notwithstanding the ethical considerations) to get there.
Replying to
@malwareminigun @jpetazzo @regehr @aburka Follow-up question to consider: why do devs have to interact with so many systems (at a seemingly expert level)?
Replying to
@palvaro@discuss.systems If what I hear is true, soon he'll forever be fumbling for his cock.
Replying to
@davidgerard @fedward Isn't this what conservative minded economists do for a living? My many decades on this planetary sphere tells me it is.
Replying to
Replying to
@ricci N = 1118213, 6598.81s
I didn't sanitize much, aside from removing basic punctuation, but does this matter, really?
Outsourcing the strcmp call to the langlemangle would do wonders for speed and compute cost.
Replying to
@mayintoronto@beige.party @LibraryAndTea@ottawa.place Yeah, the UP is decent. It's how I get there and I've been happy with it.
And I've heard no pitch about mass transit to get people to this proposed "enhancement" for Billy Bishop.
Replying to
@regehr@mastodon.social @malwareminigun@infosec.exchange @jpetazzo@hachyderm.io @aburka@hachyderm.io Anyway, this is my brain working before bed.
Replying to
@steve@discuss.systems @joe@f.duriansoftware.com I vaguely remember something like this in grad school. I think I messed around with trying to implement it once, but I didn't stick with it. I seem to recall thinking it was easier to write the formula.
Replying to
@regehr@mastodon.social @malwareminigun@infosec.exchange @jpetazzo@hachyderm.io @aburka@hachyderm.io It does. And It's about as easy to use as crontab, if we're being honest.