Is your model fast?
— Chip Huyen (@chipro) December 23, 2020
No
Is it cheap?
No
Does it at least solve our problem?
No
…
But it’s StAtE oF tHe ArT
Is your model fast?
— Chip Huyen (@chipro) December 23, 2020
No
Is it cheap?
No
Does it at least solve our problem?
No
…
But it’s StAtE oF tHe ArT
ML community: * create algorithms that optimize for a single objective *
— Chip Huyen (@chipro) December 20, 2020
Companies: * use ML to optimize for user engagement, which learns to favor extreme content since it gets the most attention *
ML community: “Why are people on Twitter getting so extreme?”
The process:
— Nihilist Data Scientist (@nihilist_ds) October 2, 2020
1. Data collection
2. Data cleaning
3. Visualization
4. More data cleaning
5. More visualization
6. Even more data cleaning
7. One last attempt at a decent plot
.
.
1076. This was a stupid idea, delete all code & forget it ever happened.#datascience #rstats #pydata
"Theory is when you know something, but it doesn’t work. Practice is when something works, but you don’t know why. Programmers combine theory and practice: Nothing works and they don’t know why." - Unknown pic.twitter.com/rwy1rbNcl8
— MIT CSAIL (@MIT_CSAIL) September 26, 2020
i like how the first step to use "TensorFlow Probability on JAX" is to "uninstall TensorFlow from this Colab entirely" https://t.co/HSbJrBslgN
— Kyunghyun Cho (@kchonyc) September 17, 2020
Omg just heard about this new cool framework numpy I think I'm gonna use it to replace tensorflow https://t.co/ux15zVNXkk
— Chip Huyen (@chipro) September 16, 2020
From the paper: "Do ImageNet Classifiers Generalize to ImageNet?" by @UCBerkeley https://t.co/h7WVmSDP0r
— Vladimir Iglovikov (@viglovikov) September 8, 2020
"Authors ordered alphabetically. Ben did none of the work."
🤣 pic.twitter.com/yRpjzfERXo
The progress in ML algorithms for tabular data has been somewhat stagnant in recent years. I blame the Titanic dataset - that problem has been worked out to death. We need a Titanic 2 dataset. Build a ship 10X in size that hits a 20X sized iceberg. Jumpstart the new research.
— Bojan Tunguz (@tunguz) August 24, 2020
Dev 1: lol holy shit I screwed up so bad
— Chris Albon (@chrisalbon) August 16, 2020
Dev 2: Hahaha nothing close to me, I am basically a technical debt machine
Dev 3: I have literally given talks on mistakes I’ve made in our codebase
CEO: Uh... should we fire these people?
CTO: Dear god no, we are lucky to have them https://t.co/n1qilom2L3
It is unreasonable to expect that Data Scientists should just be doing Data Science at their jobs. You should also be able to do software engineering, data engineering, electrical engineering, nuclear physics, optometry, pharmacology, urology, accounting, etc. #HorribleCareerTip
— Bojan Tunguz (@tunguz) August 15, 2020
An important question has now been answered: What do we call a male #Karen?
— Randy Olson (@randal_olson) August 14, 2020
According to an analysis of 70 years of baby names in the US, the answer is “Terry.”#dataviz source: https://t.co/mbjo0jAspy pic.twitter.com/8N08V6Uo4K
Want to get into ML? It’s too late — focus on your progeny! Select embryos for ML potential, play Andrew Ng lectures in utero, & force your toddler to study, withholding sleep & play time. For good measure, plagiarize tweets & start their blog. https://t.co/vhhpxOOWzC
— Zachary Lipton (@zacharylipton) August 14, 2020