The most prestigious law school admissions discussion board in the world.
|
 |
 
|
reports are of 60% yield for TSMC 2nm node
Poast new message in this thread
Date: June 20th, 2025 1:57 PM Author: razzle range we should see 2nm devices in latter half of 2025. latest NVIDIA GPUs are on 5nm, for reference.
middle eastern shitholes lobbing bombs at their useless buildings while Taiwan manufactures the Future.
get ready for faster Screens. |
Date: June 20th, 2025 2:03 PM Author: excitant feces wow (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035179) |
|
Date: June 20th, 2025 2:04 PM Author: razzmatazz violet brunch 180
Should I wait then? Im about to buy a ThreadRipper 7985WX |
 |
Date: June 20th, 2025 2:05 PM Author: razzmatazz violet brunch Missed the part about this being GPU. Still buying a 5090. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035186) |
|
Date: June 20th, 2025 2:10 PM Author: Cruel-hearted Background Story Piazza are you gonna pay $3000k or whatever the ebay cost is or murder someone IRL to get one at msrp? (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035209) |
|
Date: June 20th, 2025 2:14 PM Author: Exhilarant out-of-control chapel gunner You can buy 4x 3060 ti's for $1000 and have more VRAM than a 5090. Newer software can pool it and give you 48gb. The speed bump you get from the 5090 is more than offset by the 32gb cap. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035221) |
|
Date: June 20th, 2025 3:03 PM Author: glassy temple Yeah just run games off 4 gpus at once, good shit you chink retard (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035356) |
|
Date: June 21st, 2025 1:00 AM Author: Exhilarant out-of-control chapel gunner What can't run on a 3060 ti? (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49036663) |
|
Date: June 20th, 2025 2:07 PM Author: razzle range might want to wait but for a workstation CPU it isn't going to make a huge difference like with mobile/embedded imo (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035192) |
|
Date: June 20th, 2025 2:04 PM Author: Exhilarant out-of-control chapel gunner Apple bought Nvidia's entire production capacity for 3nm silicon for the first two years after it debuted. Nvidia still can't get 3nm. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035185) |
|
Date: June 20th, 2025 2:07 PM Author: razzle range (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035197) |
|
Date: June 20th, 2025 2:06 PM Author: Bateful Box Office Dysfunction kikes are annoying as fuck. but israel contributes bigly to global semiconductor research and development. taiwan isn't doing this on their own (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035188) |
|
Date: June 20th, 2025 2:08 PM Author: Exhilarant out-of-control chapel gunner (it is 1992) (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035199) |
|
Date: June 20th, 2025 2:49 PM Author: Khaki shrine (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035322) |
|
Date: June 21st, 2025 12:45 AM Author: sexy gaping (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49036652) |
|
Date: June 20th, 2025 2:29 PM Author: Soul-stirring Indecent Dopamine Theater there are several other improvements coming in the near future - higher memory bandwidth, more memory on the chips, better cooling. the time to train frontier-scale language models is likely to go from months to weeks by the late 2020s. we won't have to deal with the current slow pace of AI development where we only see meaningful model improvements every 2-3 months. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035265) |
|
Date: June 20th, 2025 2:30 PM Author: excitant feces it definitely hasn't been hardware that has slowed down frontier model improvements (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035266) |
|
Date: June 20th, 2025 2:37 PM Author: Soul-stirring Indecent Dopamine Theater i don't really buy the story that progress has slowed much. GPT 4.5 sucked, but the last 6 months has seen fairly significant model improvements. even comparing the first generation of the Gemini 2.5 pro in late March to the current version shows pretty large improvements on most benchmarks. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035281) |
|
Date: June 20th, 2025 2:44 PM Author: excitant feces the benchmarks are made up and meaningless
the models haven't become better at anything except some kinds of coding (only because they've been fine-tuned to do that specifically) |
|
Date: June 20th, 2025 2:55 PM Author: Soul-stirring Indecent Dopamine Theater Well, if you completely dismiss the benchmarks then there is really nowhere to go with the argument. It turns into a feeling based argument on model performance. People are notoriously bad judges of this - look how many people still insist a couple weeks after a model is released that it became dumber.
This argument does look strained though because 1) model developers know training specifically on benchmarks and releasing weak models that don’t actually perform would eliminate their credibility 2) the benchmarks are increasingly broad and model performance across different domains have very similar rank orderings. When people release new ones that the models haven’t seen, this remains true. Performance should be much more uneven if they are benchmark fitting. |
|
Date: June 20th, 2025 4:53 PM Author: Exhilarant out-of-control chapel gunner Look at AI image generation. No improvements in ages and the models aren't even that big. They still have the same crippling limitations, e.g. skin either looks plastic or everything in the background gets distorted. That's why the most realistic images have to simulate shallow focus and blur out everything else. No one seems to know how to fix basic shit like that. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035616) |
|
Date: June 20th, 2025 3:14 PM Author: Cruel-hearted Background Story Piazza the research/tool calling loops in recent models are much more capable than a year ago. this only became possible once the base models got good enough to not rapidly compound errors across iterations--there's a reason nobody was doing it two+ years ago. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035380) |
|
Date: June 20th, 2025 4:51 PM Author: vibrant fluffy address preventive strike (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035612) |
|
Date: June 20th, 2025 4:52 PM Author: excitant feces https://x.com/ylecun/status/1935108028891861393 (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035614) |
|
Date: June 20th, 2025 4:58 PM Author: Cruel-hearted Background Story Piazza ai couldn't code for 5 minutes two years ago. the whole point is the models have finally crossed a threshold where their error margin is small enough that they don't immediately compound to the point of failure, at which point relatively short-acting agents becomes viable. and as they keep improving the viable window of work duration will go up (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035634) |
|
Date: June 20th, 2025 4:59 PM Author: Exhilarant out-of-control chapel gunner Nothing mission critical will ever get turned over to AI in the private sector. Governments might dabble it with but the oil and gas industry won't. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035638) |
|
Date: June 20th, 2025 5:01 PM Author: Khaki shrine (not a Boeing exec) (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035645) |
|
Date: June 20th, 2025 5:00 PM Author: excitant feces i agree with your first post above but i'm much more skeptical of this one. i don't think they will get much better at this with more/better hardware and compute (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035641) |
|
Date: June 20th, 2025 5:09 PM Author: Soul-stirring Indecent Dopamine Theater better hardware and more compute and has been the primary way models have been improved over the last decade. it's not just about being able to train on more data, with more parameters, with more aggressive regularization, etc. it allows researchers to try out more ideas and at larger scale. algorithmic improvements are extremely dependent on being able to rapidly try out ideas. agent based coders or approaches like AlphaEvolve make compute even more important for algorithmic improvement. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035668) |
|
Date: June 20th, 2025 5:19 PM Author: Cruel-hearted Background Story Piazza it'll certainly level off at some point. but it's crossed the threshold where it's clear that with a human still in the loop current models could lead to pretty significant job loss after a fair amount of effort tailoring agentic workflows. it's just up in the air as to whether we progress to total chaos ai-ization from here or whether we sort of stagnate at a point which still requires significant numbers of humans in the loop, even if it is far fewer than required before. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035690) |
|
Date: June 20th, 2025 5:43 PM Author: excitant feces oh yeah even if AI never gets any better than it is now (and it definitely will), it will still swallow up tons of jobs
there's just going to be a lag time to adopt and implement it |
|
Date: June 20th, 2025 2:32 PM Author: ivory pisswyrm dragon seems like only yesterday that 9nm was the frontier and 6nm was the unattainable dream of the distant future (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035272) |
|
Date: June 20th, 2025 4:57 PM Author: Exhilarant out-of-control chapel gunner If backside power delivery happens it's going to reset the clock. Right now everyone is doing shithybrid designs. (http://www.autoadmit.com/thread.php?thread_id=5740903&forum_id=2#49035630) |