The most prestigious law school admissions discussion board in the world.
|
 |
 
|
GPT 5.2 is again so far ahead of everything that its a joke
| irradiated elastic band | 01/12/26 |
| fragrant pit | 01/12/26 |
| twinkling feces associate | 01/12/26 |
| Swollen Center | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| Bistre hairy legs gas station | 01/16/26 |
| Aphrodisiac friendly grandma | 01/12/26 |
| Aromatic gay international law enforcement agency orchestra pit | 01/12/26 |
| Aphrodisiac friendly grandma | 01/12/26 |
| Aromatic gay international law enforcement agency orchestra pit | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| twinkling feces associate | 01/12/26 |
| Aromatic gay international law enforcement agency orchestra pit | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| poppy persian | 01/16/26 |
| Vigorous Principal's Office | 01/12/26 |
| Aromatic gay international law enforcement agency orchestra pit | 01/12/26 |
| Aphrodisiac friendly grandma | 01/12/26 |
| Vigorous Principal's Office | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| Aphrodisiac friendly grandma | 01/12/26 |
| honey-headed underhanded trailer park lettuce | 01/12/26 |
| Vigorous Principal's Office | 01/12/26 |
| flatulent contagious mad-dog skullcap | 01/12/26 |
| twinkling feces associate | 01/12/26 |
| grizzly soul-stirring hall party of the first part | 01/12/26 |
| marvelous bronze business firm coldplay fan | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| Exciting Spot Macaca | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| marvelous bronze business firm coldplay fan | 01/12/26 |
| Garnet fluffy puppy stock car | 01/12/26 |
| naked house-broken ticket booth messiness | 01/12/26 |
| irradiated elastic band | 01/12/26 |
| irradiated elastic band | 01/15/26 |
| pontificating crystalline locus tank | 01/15/26 |
| irradiated elastic band | 01/15/26 |
| Hairraiser rough-skinned preventive strike | 01/15/26 |
| irradiated elastic band | 01/15/26 |
| Hairraiser rough-skinned preventive strike | 01/15/26 |
| irradiated elastic band | 01/15/26 |
| Hairraiser rough-skinned preventive strike | 01/15/26 |
| honey-headed underhanded trailer park lettuce | 01/16/26 |
Poast new message in this thread
Date: January 12th, 2026 2:22 PM Author: irradiated elastic band I was honestly wondering if Claude was going to jump ahead with 5.1 being kind of lackluster. but then 5.2 made a huge leap ahead of opus. and obviously gemini is still way behind both of those (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583695) |
 |
Date: January 12th, 2026 2:23 PM Author: fragrant pit how can you tell? (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583697) |
|
Date: January 12th, 2026 2:24 PM Author: twinkling feces associate (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583701) |
|
Date: January 12th, 2026 3:10 PM Author: Swollen Center (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583849) |
|
Date: January 12th, 2026 4:32 PM Author: irradiated elastic band (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584035) |
|
Date: January 16th, 2026 2:27 AM Author: Bistre hairy legs gas station (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593262) |
Date: January 12th, 2026 2:27 PM Author: Aphrodisiac friendly grandma It's a hilarious/brutal market. There's effectively no lock-in, so I just switch between whoever is good right now. Think they're the current winner, I never need to switch bots because it can't do something. In 4 the python became passable, now in 5.2 it's really good. I never bother coding "by hand" any more. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583708) |
|
Date: January 12th, 2026 2:31 PM Author: Aromatic gay international law enforcement agency orchestra pit I've used ChatGPT a few times to make scripts for Windows, and it always fucks up. Is it better at real coding? (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583728) |
|
Date: January 12th, 2026 2:32 PM Author: Aphrodisiac friendly grandma Windows Scripts? You mean PowerShell? I haven't tried that. It does just fine with python, which isn't surprising. Python is one of the most documented languages on the planet. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583733) |
|
Date: January 12th, 2026 2:37 PM Author: Aromatic gay international law enforcement agency orchestra pit Powershell and just good old batch files for cmd. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583753) |
|
Date: January 12th, 2026 4:42 PM Author: irradiated elastic band i dont believe for a second that it is messing up powershell scripts. maybe a script 3K lines long you glued together to train a model might not work the first go around. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584053) |
|
Date: January 12th, 2026 4:58 PM Author: twinkling feces associate (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584090) |
|
Date: January 12th, 2026 6:07 PM Author: Aromatic gay international law enforcement agency orchestra pit Nothing nearly that long. And it has been things like CLI syntax errors and mistakes in handling errors. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584273) |
|
Date: January 12th, 2026 6:15 PM Author: irradiated elastic band well idk. i dont use powershell I use ubuntu and mac os only. it never fucks up bash and python scripts for me. the older versions used to fuck them up a lot. I haven't seen gpt make a syntax error since like gpt 4.1. maybe once in a while on o3 and o4. but not since it switched to 5. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584292) |
|
Date: January 16th, 2026 2:22 AM Author: poppy persian I've also noticed it's lackluster at powershell (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593260) |
|
Date: January 12th, 2026 2:34 PM Author: Vigorous Principal's Office No it's horrible (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583738) |
|
Date: January 12th, 2026 2:41 PM Author: Aromatic gay international law enforcement agency orchestra pit One of the things I've found most frustrating is that when you're telling it to fix an error (which is annoying enough), it will often change (and break) other things that aren't at issue and were working just fine. It will even do that sometimes when you specifically tell it not to touch anything else. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583773) |
|
Date: January 12th, 2026 2:45 PM Author: Aphrodisiac friendly grandma What are you trying to do and what prompt did you give the bot (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583786) |
|
Date: January 12th, 2026 2:55 PM Author: Vigorous Principal's Office I am coding a discord replacement in assembly language
But it can't even give the right proxmox commands. It always says pct restart instead of pct reboot for instance, It's bad |
|
Date: January 12th, 2026 4:34 PM Author: irradiated elastic band lol yeah no wonder. you are using it to code assembly language. its really good at the popular languages. python is probably its best. also you can't always just give it some abstract prompt and expect it to guess the "right" implementation you have in mind. you need to be feeding it real concrete evidence from outputs etc. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584036) |
|
Date: January 12th, 2026 5:19 PM Author: Aphrodisiac friendly grandma (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584139) |
|
Date: January 12th, 2026 5:20 PM Author: honey-headed underhanded trailer park lettuce were u possessed by NSAM (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584143) |
|
Date: January 12th, 2026 5:50 PM Author: Vigorous Principal's Office I'm flaming online (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584236) |
|
Date: January 12th, 2026 6:09 PM Author: flatulent contagious mad-dog skullcap (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584278) |
|
Date: January 12th, 2026 3:01 PM Author: twinkling feces associate (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583832) |
|
Date: January 12th, 2026 3:10 PM Author: grizzly soul-stirring hall party of the first part (bizarro world where opus 4.5 doesn't exist) (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49583847) |
|
Date: January 12th, 2026 4:34 PM Author: marvelous bronze business firm coldplay fan (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584037) |
|
Date: January 12th, 2026 4:35 PM Author: irradiated elastic band its much better than opus 4.5. opus 4.5 just comes up with all of this stuff that literally doesn't matter. like give it something thats broken and it will give you 100 edge cases that matter in 1 in a billion situations and tell you to fix them. gpt 5.2 gives you useful stuff every time. opus 4.5 can be ok though if you have gpt 5.2 to keep it in check but still nowhere even close to as good at coding and math as gpt 5.2. also gpt 5.2 is much more conservative about conclusions. opus 4.5 will be super super certain about an answer when its just guessing. the guess might be correct like 1 in 5 times but the times its hallucinating add up. there are probably things opus is better at like writing and business intelligence. but im 100% sure gpt 5.2 is substantially better at coding and math because the gap is big. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584039) |
|
Date: January 12th, 2026 4:54 PM Author: Exciting Spot Macaca ChatGPT is smarter likexo autist
Claude is more polished like legally blonde or whatever |
|
Date: January 12th, 2026 5:46 PM Author: irradiated elastic band yeah i agree with that (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584221) |
|
Date: January 12th, 2026 8:03 PM Author: marvelous bronze business firm coldplay fan (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584733) |
|
Date: January 12th, 2026 4:43 PM Author: Garnet fluffy puppy stock car (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584058) |
|
Date: January 12th, 2026 5:04 PM Author: naked house-broken ticket booth messiness What a Joke (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584107) |
|
Date: January 12th, 2026 5:46 PM Author: irradiated elastic band (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49584223) |
|
Date: January 15th, 2026 10:37 PM Author: irradiated elastic band yeah turns out my intuition was correct. even though claude slightly edges out gpt 5.2 on the SWE i think by .8%, GPT scores SIGNIFICANTLY HIGHER on abstract reasoning benchmarks (54.2% vs. 37.6%) for Opus 4.5 making it better at generalizing to solve problems outside its training data. and also Blocker-Severity Vulnerabilities: GPT 5.2 High achieved a best-in-class security posture with only 16 blocker vulnerabilities per million lines of code (MLOC). Claude Comparison: By contrast, Claude Opus 4.5 Thinking generated 44 blockers per MLOC—nearly 3x as many—while Claude Sonnet 4.5 registered a high of 198 blockers. In deep-reasoning evaluations, GPT 5.2 has demonstrated a significant lead in identifying "blocker-severity" issues: GPT 5.2: Identified 13 out of 15 critical system-level errors (such as subtle race conditions and complex handler registration bugs). Claude Opus 4.5: Identified only 5 out of 15 of these same high-level architectural errors, often missing the deeper root causes despite fixing the surface-level bugs.
Also GPT 5.2 scores significantly higher on a newer version of the SWE called SWE pro which is less python centric. 55.6% for 5.2 with 43.3 for Claude. |
|
Date: January 15th, 2026 10:47 PM Author: pontificating crystalline locus tank With "the best LLM" use case is everything. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593074) |
|
Date: January 15th, 2026 10:50 PM Author: irradiated elastic band to me the generalized reasoning capability is the most important part. in terms of more specific "use case" stuff its just about being clever enough to make the model give you useful outputs and clever ways to cross-check and verify without putting in too much effort
I want an AGI level superhuman general reasoning machine. Not a leading expert in x gay field |
|
Date: January 15th, 2026 10:55 PM Author: Hairraiser rough-skinned preventive strike It changes on a weekly basis lately
Opus 4.5 was kicking everyone's ass since December to the point it became personally ranked #1 for the first time ever, but only for a few weeks. Anthropic must be selectively throttling max subs because some days it's noticeably weaker. On Sunday nights it is so much smarter and better at outputs and handling large contexts that it feels like a different AI
Gpt 5.2 needs more conscious prompting to extract the desired output and its responses are dense and less legible; but there is no question that it bitch slaps opus 4.5 on more complicated reasoning tasks as of THIS WEEK
Next week might be a different story |
|
Date: January 15th, 2026 11:00 PM Author: irradiated elastic band this makes it sound like you think reasoning tasks themselves change in difficulty from week to week, but maybe you just were lazy about what you wrote? The thing that changes is that the next model claude releases could be better than 5.2. I guess what you mean is that Opus 5 might be better than gpt 5.2 which I agree with. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593093) |
|
Date: January 15th, 2026 11:13 PM Author: Hairraiser rough-skinned preventive strike I have a multi-model validation process that's a ghetto version of mainlining's mahchine. It's the same exact prompts that I'm comparing outputs from. There was a brief period that 4.5 still outperformed 5.2 (and Sunday nights) (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593110) |
|
Date: January 15th, 2026 11:18 PM Author: irradiated elastic band You are confusing different output from the same prompt as the model changing from week to week. The outputs are not going to be the same every time. Outputs are stochastic. The model itself is the same until they release an update. (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593113) |
|
Date: January 15th, 2026 11:42 PM Author: Hairraiser rough-skinned preventive strike I would say it's more fluid than week to week
And I never said that models change in between releases |
|
Date: January 16th, 2026 6:58 AM Author: honey-headed underhanded trailer park lettuce I rike a GRM (http://www.autoadmit.com/thread.php?thread_id=5821130&forum_id=2\u0026mark_id=5310915#49593313) |