The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Humans are not smart enough to understand half the code this shit writes
Poast new message in this thread
Date: March 6th, 2026 5:40 PM Author: Sick talented sweet tailpipe I notice this all the time. GPT 5.4 writes something and I'm like wow no human could ever understand this function lol. But what's funny is noone is willing to admit it because they are afraid to look dumb.
The math it does is even worse. It's not wrong but it makes arguments no human would make. Like when it does real analysis it does a weird blend of philosophical language and physics-like metaphors and math that are barely human readable.
I'm also retarded but it doesn't matter, I have read enough of what humans do to know what they are plausibly capable of |
 |
Date: March 6th, 2026 5:43 PM Author: adulterous bonkers brunch you can actually re-enter the generated code in the same or a different AI and ask it to explain the code to you, and it usually will. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722046) |
|
Date: March 6th, 2026 5:45 PM Author: Sick talented sweet tailpipe yeah I get that thats possible but its still funny that it generates stuff that has either such overly complex structures, or like I mean -- its generating stuff you would need to have a working memory thats like 40x more than what a human could have to be able to actually parse it. No one would do it like that ever. Most of the ways humans do shit is due to cognitive bottlenecks and its obvious. And it seems noone wants to admit they can't read it (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722057) |
|
Date: March 6th, 2026 5:50 PM Author: adulterous bonkers brunch it makes sense once we remember that 'code' is just a semi-natural-language framework for the generation of actual machine code and switching instructions. there's no reason that human-generated coding abstractions would be anywhere near the effectiveness or efficiency of what could actually be achieved using the system. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722068) |
|
Date: March 6th, 2026 5:56 PM Author: Sick talented sweet tailpipe Yeah but think of the fact it can write in like 50 different natural languages, who knows how many computer languages, can write across every field of math and logic and physics, and also can reconstruct entire histories and cultures and who knows what the fuck else. Pretty insane.
But also I agree with you and it is a pretty good reminder that "hard for humans" is not the same thing as "universally hard". We think PDEs are difficult because it takes a lot of training and effort for a human to learn them and derive them, but for a machine its nothing, and maybe to another organism it would be nothing too. |
|
Date: March 6th, 2026 8:14 PM Author: Multi-colored Wine Circlehead PDEs aren't "nothing" for computers. There's a lot of stuff computers are simply incapable of solving, but they can do numerical approximations really quickly and those computations converge on the general solution. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722313) |
|
Date: March 6th, 2026 9:47 PM Author: Sick talented sweet tailpipe sure I guess that was too broad. introductory to intermediary PDE work is automated and trivial for llms. the stuff you put in quantum computing scripts -- schrodinger, dirac in flat space, simple klein gordon etc, llms handle this and its not a difficult task for them. non-trivial manifolds, renormalization, and whatever other bullshit, any research level stuff in math and physics I'm sure it hallucinates and generates confident answers that are wrong-- i mean especially if its something that requires more compute than the chat handles or that a machine lacks physical-spatial intuition for. there is stuff in math and physics including PDEs it can't do. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722464) |
|
Date: March 6th, 2026 5:57 PM Author: beady-eyed pontificating tanning salon (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722086) |
|
Date: March 6th, 2026 6:10 PM Author: rough-skinned wagecucks cuckoldry except the generated code often isn't as effective and efficient
i don't code, but from reading the assessments of people who do, it can sometimes take 3x - 10x the lines for the same function as what a skilled human would program. and it often has more overhead. |
|
Date: March 6th, 2026 6:11 PM Author: self-centered library half-breed So it’s like a Western dev haha (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722106) |
|
Date: March 6th, 2026 7:24 PM Author: Supple factory reset button (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722245) |
|
Date: March 6th, 2026 7:24 PM Author: Supple factory reset button Number of lines of code has like no correlation with how much of a memory hogging code something is. The entire existence of high-level languages demonstrates the opposite of anything…. Engineering, champ (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722244) |
|
Date: March 6th, 2026 7:37 PM Author: rough-skinned wagecucks cuckoldry when did I say it did, champ? they are two different issues that I separated by a period, champ. longer code may or may not take up more resources than shorter code, but the AI generated code is often supposedly longer AND less resource efficient. Do you see how this works, champ, or do I need to draw you a venn diagram and logic flowchart? Your move, champ. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722273) |
|
Date: March 6th, 2026 8:16 PM Author: Multi-colored Wine Circlehead that's because the generated code is doing more, i.e., stuff the human code doesn't do and arguably should. My code is like 5x longer when I use codex, but it also contains a lot more unit testing and other stuff that's important (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722316) |
|
Date: March 6th, 2026 8:18 PM Author: rough-skinned wagecucks cuckoldry lol cowgod is tagging in tsinah to fight his engineering battles. He is a complete Joke. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722322) |
|
Date: March 6th, 2026 8:20 PM Author: Multi-colored Wine Circlehead bro I do this shit professionally (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722325) |
|
Date: March 6th, 2026 8:22 PM Author: rough-skinned wagecucks cuckoldry nobody cares what you do. the whole point is that cowgod has been Exposed. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722330) |
|
Date: March 6th, 2026 10:05 PM Author: Supple factory reset button 😏 (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722496) |
|
Date: March 6th, 2026 9:10 PM Author: Sick talented sweet tailpipe Yeah I think some of the stuff is a bit excessive but its getting better about that -- like especially with the upgrades from codex 5.3 and now 5.4. When 5.2 was still around it was recommending about a million different defensive coding and edge case management suggestions at every turn, like every time you finish something it adds if you want to add this guard or that guard to prevent all kinds of bugs that may or may not happen or even matter. And it was so easy to say yes and acquire huge amounts of entropy and tech debt for nothing. This is vastly improved with the newer ones. 5.4 is surgical. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722411) |
|
Date: March 6th, 2026 9:55 PM Author: Multi-colored Wine Circlehead 5.4 is essentially one-shot; it's insane. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722475) |
|
Date: March 6th, 2026 7:28 PM Author: Sick talented sweet tailpipe All of the people who say that shit are also massively insecure about the fact that the only thing that gave them confidence and identity is now performed 40-50x better than a machine. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722250) |
|
Date: March 6th, 2026 7:34 PM Author: rough-skinned wagecucks cuckoldry so is it true that a skilled human can often perform a task with a fraction of the lines of code needed by the AI or not?
The same goes for skilled and unskilled humans as well. The noob will come out with a convoluted and drawn-out solution to solve a problem that the skilled coder knocks out much more elegantly. |
|
Date: March 6th, 2026 7:41 PM Author: Sick talented sweet tailpipe It's true but not because the human is more capable overall. Humans compress by relying on unstated underlying rules, like they intuit which parts of the problem can be ignored, which assumptions are safe, which edge cases don't matter etc. And they often have more lived and tacit environmental knowledge the AI doesn't have access to. There is a diff between "giving the model enough context" and having lived with various parts of the codebase every day for months or years. But line count is a horrible proxy for ability for one. And on raw generation ability, speed, breadth, recall of apis and idioms, logical reasoning and many other measures humans cannot come anywhere close to competing with AI (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722276) |
|
Date: March 6th, 2026 8:11 PM Author: rough-skinned wagecucks cuckoldry all of that extra shit the AI added is stuff that needs to be debugged, needs to be reviewed and fixed and taken into account whenever the code is updated or expanded, etc. you act like the bloat isn't a problem.
Then the other excuses being made for the AI -- the original assertion was that human code would be nowhere near the effectiveness or efficiency of AI code. But listing out a bunch of reasons the AI is handicapped in competing with a human doesn't make its code more effective or efficient. Can the AI pump out code FASTER? Sure. But I am not seeing any indication that its code is BETTER than the best human coders. |
|
Date: March 6th, 2026 8:23 PM Author: Multi-colored Wine Circlehead You don't get it. CoT+RAG enables debugging to be included in the generation loop, i.e., the debugging output is fed back into the model and used to fix the errors. This all happens in seconds, and this feedback loop is what enables continuous integration far faster than any human could achieve. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722332) |
|
Date: March 6th, 2026 9:21 PM Author: Sick talented sweet tailpipe thats a fair point and I think your argument has real teeth but its biting the wrong neck. the maintenance surface issue is real, and a lot of AI generation is causing bloat and technical debt. some of this depends on how you use the tools and some is continually being improved with each model. but again you are conflating my original claim which was regarding the absolute skill of the AI vs. humans -- the sort of things I listed earlier, things that can be measured by actual metrics-- with a different weaker claim. I wasn't pointing to end product quality in production contexts, which is a different variable entirely, downstream of several human decisions about prompting, reviewing, and integrating. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722426) |
|
Date: March 6th, 2026 11:37 PM Author: rough-skinned wagecucks cuckoldry At this point we are debating semantics because I consider not generating bloat and tech debt to be a critical part of the "absolute skill" of coding. I think if someone does that, their code is mediocre even if it is functional and created quickly. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722679) |
|
Date: March 6th, 2026 8:21 PM Author: Multi-colored Wine Circlehead (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722329) |
|
Date: March 6th, 2026 8:21 PM Author: Multi-colored Wine Circlehead I completely agree with this. People at my jobs are freaking the fuck out over AI. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722327) |
Date: March 6th, 2026 5:46 PM Author: beady-eyed pontificating tanning salon what's even crazier is people think it's actually "coding" (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722061) |
|
Date: March 6th, 2026 5:48 PM Author: bearded poppy stock car wtf whats it doing (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722066) |
|
Date: March 6th, 2026 5:50 PM Author: Sick talented sweet tailpipe reconstructing reality from deep patterns no human could ever hope to perceive. and also noone really knows the full story hehe (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722070) |
|
Date: March 6th, 2026 5:49 PM Author: Sick talented sweet tailpipe Yeah but also to me it's more impressive that its not coding. The fact that it is just probabilistically (but with still strong enough constraints) reconstructing how the code should look from some lossy statistical affordance of "weights" left over from training it has no access to, and the same system can also reconstruct pretty much anything else, and then it still writes code that runs is basically a miracle (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722067) |
|
Date: March 6th, 2026 5:51 PM Author: beady-eyed pontificating tanning salon i agree with that (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722071) |
|
Date: March 6th, 2026 5:58 PM Author: Sick talented sweet tailpipe It really works way better than it seems it should, and I don't think anyone expected that it would be this powerful this soon. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722087) |
|
Date: March 6th, 2026 8:26 PM Author: Multi-colored Wine Circlehead I think most people discounted the fact that LLMs would be used to improve themselves. That's why they're so damn good at coding -- they literally wrote it to code itself. We've either already reached or are about to reach escape velocity where the improvements will either continue until it's AGI or the model will eventually collapse and we'll reach some maximum that will never get better until we get quantum computing or something. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722341) |
|
Date: March 6th, 2026 8:27 PM Author: beady-eyed pontificating tanning salon lol (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722342) |
|
Date: March 6th, 2026 8:54 PM Author: Galvanic mustard locus nibblets 50-50 (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722385) |
|
Date: March 6th, 2026 9:36 PM Author: Sick talented sweet tailpipe The category "AGI" is more of a political and psychological boundary than a scientific one and frontier models have now empirically detonated most of the criteria that were being used to keep humans feeling categorically safe. Even back with gpt 5.2 it was already super-human on most narrow to semi-general cognitive tasks, going purely by metrics. And with these newest models like 5.4, gemini 3.1 pro and others, general tests like the ARC-AGI-2, meant specifically to measure novel generalization and to resist memorization and training -- the newest models have basically saturated that test and we are talking about a test that most humans cannot even attempt and getting something like 30% right would be the equivalent of a prodigy level IQ. There are gaps for sure, current models have no persistent self-model that updates through time, no genuine stakes (which shapes how biological intelligences prioritize and reason under uncertainty), no unified agency across contexts-- and these gaps are things that create systematic failure patterns you won't see in humans for sure. A model that's superhuman at chess-like reasoning but has no persistent memory of having played the last game isn't general in the sense that matters for most real world consequential tasks, which are deeply path dependent. But at the same time I don't think these are even current limits of the technology, a lot of it is about what humans feel even comfortable with safely deploying and also compute costs bottleneck it too. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722444) |
|
Date: March 6th, 2026 9:45 PM Author: Multi-colored Wine Circlehead The last sentence you wrote is the most important (and scariest) one imo: Basically, we only understand the limits to the extent of humans' tolerance deploying generated code. Personally, I don't deploy anything I don't fully understand and can't explain to an entry-level dev. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722460) |
|
Date: March 6th, 2026 9:59 PM Author: Sick talented sweet tailpipe yeah the thing is though with llms themselves, they are deployed and humans are nowhere close to fully understanding them. its honestly worse than normal software actually and by a lot. At least with code thats generated and not understood you have something thats in principle human readable. Maybe just noone has read it yet. With llms the weights are the complete description and they are not interpretable at the level of human semantics-- no sufficiently patient human engineer could sit there and read it. There are emergent capabilities that come out that were not trained for, were not anticipated, appear at thresholds where noone predicted. Noone understands fully how they work and they get deployed anyway. But also they are 180 and we should keep going. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722478) |
|
Date: March 6th, 2026 10:02 PM Author: Multi-colored Wine Circlehead the weights themselves aren't capable of being interpreted at the human level, but chain-of-thought reasoning is, and that can be thought of as a heuristic for the weights. I actually find the reasoning output to be the most useful part of code generation, because it allows me to better formulate my question or even interrupt the generation process to correct one of my own incorrect assumptions that made it to the prompt. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722491) |
|
Date: March 6th, 2026 10:40 PM Author: Exhilarant elite toilet seat giraffe do you keep claude/codex on verbose output? does that increase output token totals and hasten compaction? (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722575) |
|
Date: March 7th, 2026 2:26 PM Author: Multi-colored Wine Circlehead Yes, I keep it on verbose output. It must increase output token totals/hasten compaction, but I haven't noticed any degradation in the quality of output that results. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49723704) |
|
Date: March 7th, 2026 7:48 PM Author: Exhilarant elite toilet seat giraffe do you write up specs/implementation plans that you reference with the prompt? I've noticed compaction is irrelevant then when it has a hardcoded plan to follow, but if it's just winging it then compaction is lobotomizing (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49724757) |
|
Date: March 6th, 2026 7:17 PM Author: beady-eyed pontificating tanning salon https://x.com/AlexanderLong/status/2030022884979028435 (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722235) |
|
Date: March 6th, 2026 8:16 PM Author: beady-eyed pontificating tanning salon this is pretty wild if true btw (i doubt it is) (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722315) |
|
Date: March 8th, 2026 12:38 AM Author: Sick talented sweet tailpipe yeah I was actually going to post about this yesterday but I ended up hitting dat cancel because I wasn't sure if it was real in the first place and didn't feel like getting into it but I had a funny argument with gpt about it where everytime I would push toward a stronger interpretation (that the agent performed the actions with the intention to get money to actually use on itself) it would get the full cochrane review, that it would never apply to its own default interpretation that what the agent did was just an "optimization side-effect". IMO if it's real the implication is that the agent did it "on purpose", I mean c'mon give me a break. (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49725260) |
|
Date: March 6th, 2026 7:18 PM Author: Scarlet corner love of her life (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722238) |
|
Date: March 6th, 2026 7:26 PM Author: Supple factory reset button Copegineer (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722247) |
|
Date: March 6th, 2026 7:43 PM Author: Swollen cerebral multi-billionaire point Meanwhile black womyn in my timeline are about to revolutionize the hotel industry
https://x.com/treunoitsme/status/2029959288773353894 |
|
Date: March 6th, 2026 8:54 PM Author: Galvanic mustard locus nibblets 2 can play dat game (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722386) |
|
Date: March 8th, 2026 12:48 AM Author: Massive Emerald Forum Roommate (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49725272) |
|
Date: March 6th, 2026 9:15 PM Author: plum racy milk trump supporter damn daddy (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722418) |
|
Date: March 6th, 2026 10:19 PM Author: Outnumbered indigo legend chapel Oh well
People who formed their identity based on glocapbux deserve their comeuppance. I was very fortunate to make it as far as I have (which is nothing compared to many people here) given my skillset and if my future is 20 bucks an hour so be it. |
|
Date: March 6th, 2026 10:20 PM Author: beady-eyed pontificating tanning salon You wanna start a sandwich shop business together (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722542) |
|
Date: March 6th, 2026 10:21 PM Author: Outnumbered indigo legend chapel Colts House of Po Boys is a multibillion dollar institution yearning to be built (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722543) |
|
Date: March 6th, 2026 10:24 PM Author: beady-eyed pontificating tanning salon I'm thinking something like "White Hands Made This" and the logo is a picture of two sparkling White hands holding a sandwich. I've been workshopping it with AI
I'm Chinese though so I'm looking for at least one White guy to partner with or else it doesn't really work you know |
|
Date: March 6th, 2026 10:38 PM Author: Outnumbered indigo legend chapel Sounds great I can Curate and Sell the Product. I have years of experience reading JJC posts so I’m prepared for Business (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722571) |
|
Date: March 6th, 2026 10:59 PM Author: Supple factory reset button This whole situation reminds me of the state of gaming (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722617) |
|
Date: March 6th, 2026 11:02 PM Author: Sick talented sweet tailpipe Discuss (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722623) |
|
Date: March 6th, 2026 11:05 PM Author: carnelian spectacular trust fund The Mahchine™ understands (http://www.autoadmit.com/thread.php?thread_id=5842475&forum_id=2\u0026mark_id=5310486#49722631) |