IMO ini pertanyaan valid sih. AI di medical imaging emang harus dibandingkan sama golden standards yaitu PCR ini. Harus bandingin metrics kya AUC-ROC, precision, recall,... Baru ke efisiensi.
Bahkan ada artikel sendiri buat ini https://t.co/oQo7wO54ok
This is next level insane.
All of this was learnt from predicting the next token.
It is absolutely mind blowing.
And about the question if LLMs have a "world-model":
They have LITERALLY a world map inside them.
Worrying if my job was jeopardized by AI this week or if we’re still good, I read a new paper evaluating #GPT4V - a #GPT4 version handling image and text inputs. It produces *impressive* radiology reports. But let’s delve deeper into some of the results... #radiology#AI
@lynxluna Drama tentang dhh removing TS definition di hotwire, terus (expectedly) di troll primeagen dan th3o.
Sempat juga theo bikin MR yg revert TS removal 😂
Early in my career I broke Facebook Chat (pre Messenger) for all IE users. Couldn't sleep all night afterwards.
However, one of the most interesting issues I caused was when I removed about 300,000 duplicated files in Facebook's monorepo and pointed everything to the remaining copies. I went to sleep and woke up the next day with angry messages on my phone: I made https://t.co/Yj0AscnHTG 1.2 seconds slower to load!
How? Facebook's bundler used machine learning to optimize JS bundles. That bundler was relying on paths for JS files, and it didn't track changes like mine. So when I pushed my change, the bundler was under the impression that all those files were new and basically unbundled them and sent them to users as individual files instead of JS bundles!
Given the scale Facebook was operating at, and some issues in the Python based Mercurial client (which have since been replaced), it took about an hour to merge my PR into main. The revert to make https://t.co/Yj0AscnHTG fast took a three people team 12 hours to complete. Afterwards an intern and I redid the initial work by also patching the bundler's paths correctly.
I caused other large issues, but this one is my best bad one. It worked while testing, could not have been noticed in development since the bundling infrastructure was different, and basically only one person at the company would have known this issue could have been caused.
Wait wait,
Hold on a sec,
Did they just dump 8 Trillion tokens into a 7B model and then casually tweeted the download link??
5,714% of chinchilla's optimal.
Fucking Legends!!! 😂
@rayestu Well, jadi ada single point of failure. Soalnya NIK kita udah (presemubly) open source. Kecuali format penomoran NIK seluruh Indonesia baru dan dijaga kerahasiannya.
Menghitung Untung Jualan Kangkung Hidroponik
15 hari bisa panen
Setiap mulai menyemai kangkung selalu terlintas bagaimana 7 tahun lalu saat masa sulit terselamatkan oleh kangkung.
Sekarang saya mau share hitungan untung jualan kangkung Hidroponik aja.
Thread