02:14:49 @rbrunner7: https://innovationgraph.github.com/global-metrics/git-pushes explains it all 02:18:11 Diabolical increase in 2025 q1 02:19:05 The amount of Git pushes has surged, most notably in third-world countries, but if you look at the total number of new repositories being made, there is no similar increase 02:20:09 In other words, LLMs have lowered the barrier to entry for users to contribute, but this isn't "AGI" that has allowed innovation to happen in the form of new projects and ideas 02:20:26 @duckpondy:matrix.org: https://innovationgraph.github.com/global-metrics/repositories 02:20:31 "contribute" 02:20:42 <321bob321> KPIs met 02:21:00 Github still shadowbans user accounts, but bots run rampant 02:21:20 <321bob321> Same with alias emails 02:22:49 @ofrnxmr:xmr.mx: It's an overstatement 02:24:34 Written the code is often the easiest part > <@robbin_da_hood:matrix.org> I would disagree with the argument that you are better off writing code yourself. Except under special circumstances. The amnesic window being one. 02:26:10 > Moreover, AI adoption starts quickly: 80% of new developers on GitHub use Copilot in their first week. 02:26:32 Pathetic 02:26:50 Rhese arent developers 02:26:59 Isn't Copilot automatic? 02:27:00 No 02:27:04 So it's opt in? 02:27:13 Its included for free, but its opt-in, yeah 02:27:59 some projects use copilot and coderabbit for auto-reviews 02:28:09 KeepassXC 02:28:18 Keepass uses copilot to commit directly to prs 😭 02:28:43 DataHoarder made a great point. These companies and their leaders have spent so much on AI that they'll keep going until investors start demanding profits > <@ofrnxmr:xmr.mx> Pathetic 02:30:06 @ofrnxmr:xmr.mx: I don't think this is a problem, assuming the developers are competent, and droidmonkey seems to be. The real issue with using LLMs is doing so blindly on a codebase you don't understand and not reviewing changes 02:31:38 @duckpondy:matrix.org: its a huge problem imo 02:32:46 Who reviews the pr? the same person prompting it? 02:32:58 @ofrnxmr:xmr.mx: Why shouldn't it be fine? A competent developer's review should prevent any issues 02:33:29 can you refer to a developer who doesnt write code anymore, as competent? 02:33:32 @ofrnxmr:xmr.mx: Doesn't Copilot create it's own pull requests? 02:33:45 @duckpondy:matrix.org: if thats how you use it 02:34:10 @ofrnxmr:xmr.mx: Coding is the easiest part of the process 02:34:26 no it isnt 02:35:02 monerod is a good example. It has great code but poor design choices. Most of the time is spent coming up with the design and what features to add 02:35:16 Monero is proof positive that coding isnt the easiest part. And ai writes more spaghetti than interns 02:35:31 i disagree 02:35:45 Monerod's code is pretty bad in many places 02:36:29 @ofrnxmr:xmr.mx: It took years to research FCMP++ and plan its implementation, but the actual coding was done very quickly. The current issues stem from DESIGN choices, not the code itself 02:36:51 it took like 6 months to research fcmp++ 02:37:09 And like 3months to develop it (dont quote me) 02:37:25 The coding of the implementation has taken about a year so far 02:37:30 @ofrnxmr:xmr.mx: I'm counting Seraphis even though it's separate 02:37:46 seraphis took 2+n years and never finished 02:37:51 Unrelated to fcmp++ 02:38:27 Why is coding the hardest part? Why can't an LLM do it? 02:38:43 because LLMs write spaghetti 02:39:00 No they don't 02:39:08 asking that question after all the responses you have gotten here today duckpondy? :D 02:39:24 and dont do things the "best" way, they do things however they learned from stackexchange questipns 02:39:58 Humans improve functions and create new ones. LLMs use historical data and repeat mistakes 02:40:43 On some private repos, i have to get angry at LLMs for offering cracked out suggestions, proposing to use deprecated functions, etc 02:41:29 im not saying ai cant get better. Im saying, as it is today, its better than a script kiddy, but not as good as any real dev 02:41:39 <321bob321> What guidelines does a llm have with writing code. How does it know what is secure 02:42:14 nioc: My perspective on LLMs, which I learned from the conversation, is that they're great at coding. They have plenty of data to train on, as we typically write the same functions over and over again. However, they fail completely at replacing engineers because coding is only one part of the story. You also need to have an id [... too long, see https://mrelay.p2pool.observer/e/gqDByckKYTJrenVO ] 02:42:43 The hate is exactly what you just said 02:43:16 Code isnt just math 02:44:14 LLMs dont "understand" what they are doing. Can be fine for creating a simple function, or to code something inefficently 02:44:32 But to write good code, you have to understand what youre doing AND be a good dev 02:44:33 I learned about the specific limitations that make it so LLM need to be used judiciously for only specific aspects of coding 02:45:29 otherwise they create more work than they save 02:45:59 So i have no real issue with using LLMs to review code. But to have them write code which a human reviews, just leads to humans having to rewrite it (or acceptance that you dont care about code quality) 02:46:09 @ofrnxmr:xmr.mx: Yes, and that's what's happening in KeePassXC, for example. Copilot creates PRs that create or refactor simple functions (less than 100 LOC on average), and with review, this is fine 02:46:13 The common thing you hear from vibe coders "does it work?" 02:46:30 But you end up with 10000LOC for a 900LOC solution 02:47:32 @ofrnxmr:xmr.mx: Vibe coding never made sense to me. IMO the problem with vibe coding is that it removes the need for human judgment. When you blindly ask an LLM for solutions, you're just a middleman. If that's the process, the human is no longer adding value and should be removed from the loop 02:47:37 LLMs often pile on instead of fixing 02:47:43 @duckpondy:matrix.org: Keepass is vibe coding.. 02:48:21 Literally in the comments they are juat prompting copilot to make changes, and then accepting them 02:48:38 @ofrnxmr:xmr.mx: For KeePassXC, Copilot creates PRs on its own, and developers review them to prevent hallucinations. I fail to see the problem and completely miss your point 02:50:25 copilot (or whatever) creates prs after being prompted to, then its further prompted until it looks acceptable. 02:50:25 For some unimportant code, this might be fine 02:51:46 Again, there are a lot of people using ai, and most of them pretending not to 02:52:49 things that "work", and even pass reviews, but later go under a microscope and you have to start wondering wtf is going on 02:53:41 @ofrnxmr:xmr.mx: If I understand you correctly, you're suggesting that AI-generated code contributes to technical debt with every merge, regardless of whether it's been reviewed? 02:54:15 https://github.com/monero-project/monero-gui/pull/4514 02:54:40 https://github.com/monero-project/monero-gui/pull/4445 02:55:41 https://github.com/monero-project/monero-gui/pull/4513 02:55:57 @ofrnxmr:xmr.mx: This is exactly my point. The AI fails to understand Monero. It proposed a "skip syncing" feature, which is nonsensical. You can't skip the sync you need to get the blockchain in the first place, even if the code for the feature itself is perfectly fine 02:56:59 The code is fine; the problem is blindly trusting the LLM 02:57:10 @duckpondy:matrix.org: my point is that reviewing prompt slop is often more work than writing code, and you end up with prompt-reslopped code 02:57:58 Meaning, you can ask an llm to keeo modifying code until it resembles something viable, but for anything more than a few lines, its almost always bad 02:58:21 if you write code with ai assistance, thats a different thing. 02:59:15 # 🎯 BOUNTY CLAIMING STRATEGY & ACTION PLAN > <@ofrnxmr:xmr.mx> https://github.com/monero-project/monero-gui/pull/4513 02:59:15 Primary Objective: Claim 140.167 XMR (~$28,000) for Monero GUI I2P integration 02:59:15 Secondary Objective: Find and pursue additional bounties to upgrade Copilot 02:59:15 Timeline: Aggressive 2-week target for PR submission 02:59:40 Ai assisted code can help, as can ai reviews. But ai authored code.. 02:59:40 example: there was a oroject called easy-monero (iirc), while it worked and seemed ok, it re-defined the same functions like 10 times, wiped your ssh settings, changed your dns resolvers 03:00:12 @duckpondy:matrix.org: Lol yeah 03:00:23 Copilot wants to upgrade itself 03:00:41 @duckpondy:matrix.org: The problem here is not the LLM, but how it's being used. This is 'vibe coding' because the author just wants money and doesn't know what they are doing. They are useless here 03:01:00 @ofrnxmr:xmr.mx: My argument is exactly this 03:01:28 i reviewed easy-monero but i concluded that its simply unsafe to recommend 03:02:28 Because ai wrote some dangerous functions and repetitive + redundant code 03:02:40 Make more sense to rewrite the whole thing than to prompt it into shape 03:05:16 > rewrite the whole thing 03:05:35 https://github.com/Fountain5405/monerosim 03:05:35 this is another codebase that im not particulaly happy with 03:05:37 https://github.com/Fountain5405/monerosim/commit/be0d71d7c38d7e8688f7f1c58dd152e49dbb734e 03:05:44 Like look at this commit by AI 03:06:05 @duckpondy:matrix.org: While rewriting the whole thing, using LLMs is much faster and smarter than doing it without their assistance. I don't and will never agree with the Luddite opinion 03:06:10 3000 LOC, even duplicates entries in gitignore 03:07:08 @duckpondy:matrix.org: Faster? Depends on how much time you have to soend hand-holding 03:07:31 Clearly the first 7 days handholding gave you a bunch of spaghetti 03:07:49 Would take like 4hrs to rewrite it by hand 03:08:22 And prompting doesnt help you become a better dev 03:08:40 Rewriting the code might though 03:08:48 ofrn is this project vibecoded? 03:08:51 https://codeberg.org/MarkA860/AutoMonero 03:09:54 https://github.com/Fountain5405/monerosim/commit/8a326740bd580d84b0f51695f521513b2d29506a +17046 -85492 03:09:54 this codebase will probably always be inreviewable 03:10:16 @user2570:unredacted.org: I saw your message and never checked the code yet 03:10:19 @user2570:unredacted.org: Yes I could immediately tell from https://automonero.com/ 03:11:30 https://codeberg.org/MarkA860/AutoMonero/src/branch/main/CODEBASE_ANALYSIS.md and only 2 commits proves it 03:13:00 Lol the readme looks like slop, yeah 03:14:46 @duckpondy:matrix.org: Sorry, i mean this file 03:15:10 > <@ofrnxmr:xmr.mx> Faster? Depends on how much time you have to soend hand-holding 03:15:10 You should only need hand-holding if you're using an LLM to do the entire job. Take the monerosim project. It's foolish to have an AI agent write all the code, but it's equally foolish to do everything yourself without LLM assistance. The right way is to handle the planning and research on your own by reading the Monero docume [... too long, see https://mrelay.p2pool.observer/e/k9S5yskKN3p6M2to ] 03:15:15 there is no readme 03:15:38 "equally" i highly disagree 03:15:52 Doing everything yourself is fine 03:16:05 Luddite opinion 03:16:07 I dont think fcmp was written with ai assistance 03:16:22 It's why FCMP++ is taking so long 03:16:30 fcmp is finished. 03:16:42 Has been for about a year 03:16:47 Not on mainnet 😛 03:17:11 the integration is essentially finished as well 03:17:26 Carrot integration is not finished. carrot != fcmp 03:18:27 And fcmp development was completed probably a year ago (some changes have been made since then, due to testing and design choices, like ram improvements or max inputs, or hash to point) 03:19:00 AI cant integrate fcmp. Thats a joke 03:19:05 Simply not possible 03:19:22 It cant do anything novel 03:19:25 I'm not saying it can 03:20:27 CARROT is live on mainnet for salvium btw 03:20:51 And if wownero wanted to, they could launch fcmp to mainnet tomorrow 03:21:21 Completely disregarding AI is foolish; it is an incredibly helpful tool. The traditional solo coding loop is write code, encounter a problem, research the documentation, look at stackoverflow posts, and repeat. I propose a more efficient loop write code or use a prompt for small issues, encounter a problem, consult an AI, and then iterate 03:21:37 Where did i completely disregard it? 03:21:43 This can be used for anything including novelty 03:21:54 that loop is false 03:21:57 features 03:22:21 Google stack overflow > ai fwiw 03:22:54 Absolutely not 03:22:58 AI has a habbit of not knowing the best answers that it pulls frok SO 03:23:22 understanding the answers > being told the 2nd best one 03:23:31 You're saying coding from scratch is better than coding with AI > <@ofrnxmr:xmr.mx> Where did i completely disregard it? 03:23:44 Coming up with a solution > reusing poor ones 03:23:48 @duckpondy:matrix.org: It is 03:24:40 Nope I can never agree with that opinion 03:24:47 thats fine 03:25:12 You should try experimenting with AI more. Try Claude Sonnet 4.5, which is the best model I've used so far 03:25:23 Claude is the worst offender 03:25:43 Don't use it blindly, give it rubbish prompts, and then complain. Use it over the manual Stack Overflow loop 03:25:54 You'll be more productive 03:26:59 Nothing more annoying than trying to understand a git blame, and realizing that ai wrote the code and the author cant explain it either 03:27:19 Not that its not functional, but that the way its done makes no sense 03:29:04 ive literally had to listen to people profess how using ai agents is the future, how normal devs are inferior to one who uses ai, and then have the prompter have a mental breakdown because he cant sort out the mistakes that claude was making 03:29:14 Before open source and Git: writing code from scratch, sharing it inefficiently. Before Stack Overflow: relying solely on documentation and man pages. Before AI... you get the idea 03:29:28 Written by claude, reviewed by copilot, coderabbit, and humans 03:30:00 @duckpondy:matrix.org: People still rely on manpages broski 03:30:27 Ai is NOTHING without documentation, manpages, and SO 03:30:59 If everyone uses AI, how do new questions and answers even get trained into it? (if SO doesnt have new questions and answers being solved that can be used for training data) 03:31:47 If devs dont use SO, then AI has no info 03:31:57 @ofrnxmr:xmr.mx: I know, but that's not my point. My point is that if you don't use AI, you're losing out on coding faster and falling behind. However, I'm not using the shill perspective that says AI can do everything and that vibe coding is the future. I'm saying that using AI in moderation and reviewing everything is the key. That's my final point 03:32:05 Expecting AI to have answers to unasked questions = hallucinations 03:32:35 I never said using ai assistance was evil 03:32:47 You said coding from scratch > AI assistance 03:32:51 I said writing code with ai < writinf code from scratch 03:33:01 Yeah, you can wrote code from acrarch with ai assistance 03:33:39 https://gist.github.com/hopeseekr/f522e380e35745bd5bdc3269a9f0b132?ref=blog.pragmaticengineer.com#file-stackoverflow-new-questions-over-time-2009-2024-csv > <@ofrnxmr:xmr.mx> If devs dont use SO, then AI has no info 03:33:39 SO is dead and AI is still doing fine. It uses other information to train like documentation and open source code 03:33:51 let me reword: writing code > copy and pasting ai code 03:34:09 https://mrelay.p2pool.observer/m/matrix.org/xJhSkPOzbnniHdkBBLlIckea.png (image.png) 03:35:38 And people wonder why ai has to hallucinate 03:36:37 I'll only advise coding from scratch without AI if you're learning 03:37:02 Id propose the opposite 03:37:09 you can learn from the code ai writes 03:37:28 Or, as Cindy_ pointed out, if you're facing a truly novel issue with almost no documentation online 03:38:47 @duckpondy:matrix.org: hard to do this if your ability to write code is rusty or poor 03:38:59 This approach only works if you can already distinguish correct from incorrect code. Since you can be fooled by AI hallucinations, it's wise to avoid relying on it when learning a new language or framework > <@ofrnxmr:xmr.mx> And people wonder why ai has to hallucinate 03:39:29 Not saying to rely on ai, but to use it as sort-of teacher 03:39:41 It can explain basic commands, functions etc 03:39:54 And cam review your code, help you test it / deploy it 03:40:26 But a competent dev can do all of this in a few seconds (aside from review) 03:40:44 Like DH said, ai caught a typo (i instead of 1, or vice versa) 03:44:15 I don't agree with what you have to say, but I appreciate your perspective. I learned a lot from this discussion 03:44:27 we all use static analysis, linters, etc, but those dont review code for functionality. We run unit tests, but that still might not catch typo or offer suggestions to improve the code. 03:44:27 Ai reviews and "assistance" = im not against these things. someone able to +competently write the code from scratch > any prompter + ai 03:44:44 I'll agree to disagree 03:47:35 One final question: What about products that use AI-generated code, like KeePassXC? Do you think we should avoid them because the rate of faulty code is high? Every company is using AI. I remember seeing a CLAUDE.md file in Proton's repos, and obviously, it's impossible to tell in some cases 03:48:32 No, we have no idea whats in many products 03:49:12 Im sure ai has a much higher prevelance in proprietary works, and even universities / schooling 03:50:56 Theres nothing that makes ai inherity more unsafe than code written by hand, but ai code (especially insensitive areas) needs much higher scrutiny, because its possible that no himan EVER actually looked at it 03:52:22 Example: kewbis haveno app. Pure unusable garbage, nobody noticed 03:52:22 Or, in the case of a hand-written project: mysu, a foss monero wallet. only 1 person had the source code, ansd maybe 2 people had ever looked at it. 03:53:22 Cant assume that, just because something is foss, that its ever been reviewed 03:53:38 What about using LLMs to review code? 03:53:46 sech1 said it's unreliable 03:53:50 I haven't tried it before 03:54:00 perfectly fine, even if reviews are trash sometimes 03:54:53 LLMs could be useful in detecting backdoors to prevent something like xz from happening again 03:55:11 ive had coderabbit tell me that things were done incorrectly, and had to correct it. But its also caught times where devs are writing things incorrectly (like a logic issue) 03:55:51 @duckpondy:matrix.org: sometimes, yeah. Still cant "trust" it 03:56:18 A friend of mine lost a large amount of money due to a backdoor in one of the repos 03:56:41 He ran the repo through ai, even the specific commits, and it didnt notice the glaring issue 03:56:49