Andrej Karpathy
cs.stanford.edu · Personal blog · on HN since 2007
Andrej Karpathy's Stanford pages host his deep learning projects, image captioning work and PhD thesis, plus other Stanford CS material.
What HN saysHN links the cs.stanford.edu domain for Stanford pages on game theory, CS enrollment history, and research papers, not just Karpathy's own work.
284HN people shared it
201linked it in comments
104stories submitted
2.1Kstory points
2 months agolast shared
2007people per year2026
Most-shared posts
- 1ConvNetJS: Deep Learning in your browser18 people · 88 points · first shared 2014 · 1 comments on HN
- 2Neural Networks That Describe Images17 people · 379 points · first shared 2014 · 60 comments on HN
- 3Tide.pdf6 people · first shared 2023
- 4Axelrod's Tournament5 people · 3 points · first shared 2023 · discussion on HN
- 5The IT Productivity Paradox (2015)4 people · 91 points · first shared 2020 · 55 comments on HN
- 6A History of Capacity Challenges in Computer Science [pdf]4 people · 72 points · first shared 2016 · 5 comments on HN
- 7DenseCap: Fully Convolutional Localization Networks for Dense Captioning4 people · 9 points · first shared 2016 · 1 comments on HN
- 8Connecting images and natural language: Andrej Karpathy's PhD thesis [pdf]4 people · 3 points · first shared 2016 · discussion on HN
- 9Von Neumann and the Development of Game Theory4 people · 2 points · first shared 2014 · discussion on HN
- 10Visual Q-Learning Demo in JavaScript from Karpathy (2013)4 people · 1 points · first shared 2014 · 1 comments on HN
- 11ConvNetJS MNIST demo4 people · 1 points · first shared 2014 · discussion on HN
- 12DB mooc4 people · first shared 2017
- 13Econ hours productivity4 people · first shared 2016
- 14Doing well in your courses: Andrej's advice for success (2013)3 people · 684 points · first shared 2018 · 198 comments on HN
- 15Stratified Synthesis: Automatically Learning the x86-64 Instruction Set (2016) [pdf]3 people · 74 points · first shared 2022 · 7 comments on HN
What HN people wrote when they linked it
uriegas · 2026-01-27
There are some research projects out there that use LLMs for health diagnostics. Here's one: https://cs.stanford.edu/people/jure/pubs/med-pmlr23.pdf
They usually require more data It is not a great idea to diagnose anything with so few information. But in general I am optimistic of the use of LLMs on health.
samspenc · 2024-05-20
1985-1993: these two articles can explain it better than I can
https://cs.stanford.edu/people/eroberts/CSCapacity/CSCapacity-History1… https://cseducators.stackexchange.com/questions/2942/why-did-interest-…
2003-2010: dot-com bust and poor economic growth followed by Great Recession. Even though the dot-com crash happened in 1999-2001, the students who were in CS would pass out a few years later into an economy scarred by both the dot-com bust and 911. And then the Great Recession followed, of course.
goldenjm · 2023-11-06
On 2, you could recreate Axelrod's classic prisoner's dilemma tournament from 1980. Here's a short writeup about it: https://cs.stanford.edu/people/eroberts/courses/soco/projects/1998-99/…
Another game you could try is the "guess 2/3 of the average" game.
You could make the assignment one game like these, or put a few together.
keithplayer · 2023-01-01
I agree and the following is more related to the text you quoted.
I interviewed at GS and as soon as I saw the room of clearly exhausted and joyless developers, I could see they actually valued the appearance of work more than productivity.
The 40-hour work week was introduced for a reason: https://cs.stanford.edu/people/eroberts/cs181/projects/crunchmode/econ…
It is sad that 100 years after Ford made the discovery people still don't get this. Team A consistently working 40 hours per week will output more work than Team B working 60+ hours per week. That's true for 60, 80 and especially 95+ hours.
lioeters · 2022-10-12
Planiverse is one of my all-time favorite books. Some tasty morsels:
The Planiverse: Computer Contact with a Two-Dimensional World - Engineering Designs in Planiverse - https://cs.stanford.edu/people/eroberts/courses/soco/projects/2005-06/…
---
> Dewdney wrote The Planiverse as..an allegory for his search for a reality deeper than that of scientific enquiry, and his subsequent conversion to Sufiism.
We're the ones living in flatland, a planiverse of limited 3D perspective, who yearn for higher dimensions of experience, a glimpse of the beyond. Unfortunately it's not without risk, as this way madness lies - either as a trap of illusion, or at least as a step toward fundamental truth (if any).
jhugo · 2022-10-05
On a whim a few years ago I attended Knuth80[0], a symposium celebrating Knuth's 80th birthday in Piteå, northern Sweden. I traveled there by train from London and spent a few days listening to fascinating lectures and attending the inaugural performance of Fantasia Apocalyptica, performed by Jan Overduin on the wonderful new pipe organ in Piteå. I only had the opportunity to speak briefly with Knuth, but the whole thing (including the location in northern Sweden in winter!) was a wonderful experience, and should there be any further birthday symposiums I would heartily recommend them to anyone!
[0] https://cs.stanford.edu/content/contacting-donald-knuth/news17.html
recursivedoubts · 2022-01-23
This essay contains the famous "Worse is Better" section, made popular by Jamie Zawinski
https://cs.stanford.edu/people/eroberts/courses/cs181/projects/2010-11…
Summarized as:
"When tradeoffs are required in a software system, priority should go to implementation simplicity at the expense of completeness or correctness."
See also the wikipedia article:
https://en.wikipedia.org/wiki/Worse_is_better
An interesting aspect of all this is that Gabriel argues with himself in a series of papers under a false name:
https://dreamsongs.com/Files/worse-is-worse.pdf
A great example of someone looking at both sides of an argument.
j-pb · 2021-10-06
TL;DR
Your intermediary result sets can be significantly bigger than your ouput set. By instead of combining whole relations but performing a depth first search through the variable assignment space, while peforming constraint propagation, you can achieve a runtime that is much better than what classical pairwise joining systems can achieve.
Since the proposed solution for joins in RxDB is an especially naive kind of pairwise join, it will be especially bad.
https://cs.stanford.edu/people/chrismre/papers/paper49.Ngo.pdf
falcor84 · 2021-10-05
There was actually a third excellent course launched at the same time - Jennifer Widom's Introduction to Databases. The original version is no longer available, but there are mini-courses in its stead. Highly recommended - https://cs.stanford.edu/people/widom/DB-mooc.html
latte_machiato · 2019-05-22
It's interesting to see how reinforcement learning is used to assist humans. This is a great combination of HCI and AI. Emma Brunskill at Stanford - https://cs.stanford.edu/people/ebrun - has a series of work using an RL framework to teach humans faster. For example, in this work they developed a POMDP to choose pedagogical activities that will maximize student learning, based on the student's current knowledge.
inutard · 2018-12-17
Great to see this come up again! Here's a simulator I wrote for it a looooong (almost 8 years!) ago: https://cs.stanford.edu/people/paulliu/webapps/pendulum.html
capkutay · 2018-11-29
This has to be a joke? This looks more like 'how to use a k-v store 101 for non programmers'
I don't want to put more content under a terrible post, but the best resource for this material Jennifer Widom's MOOC.
https://cs.stanford.edu/people/widom/DB-mooc.html
nickpsecurity · 2018-11-05
If interested in this topic, here is a survey that tells you about various ways to do automatic, test-case generation:
https://cs.stanford.edu/people/saswat/research/ASTJSS.pdf
My baseline recommendation is using one of every type if you can just run overnight on a server dedicated to verification or testing. I also think hybrid methods have a ton of promise combining static analysis, dynamic analysis, and above methods of test generation. Here's an example:
https://homes.cs.washington.edu/~mernst/pubs/palus-testgen-issta2011.p…
slim · 2018-10-14
Nortel willingly sacrificed its ip to profit from building the great firewall of China
https://cs.stanford.edu/people/eroberts/cs201/projects/2010-11/Freedom…
Joris1225 · 2017-06-10
In Andrej Karpathy's excellent blogpost on RNNs[1], he links to some of the code he used for visualisation[2]. I have done something similar. In general: put your the activations of each cell memory over time through tanh and color it based on that.
[1] http://karpathy.github.io/2015/05/21/rnn-effectiveness/
[2] http://cs.stanford.edu/people/karpathy/viscode.zip
esfandia · 2017-03-13
I love this. It has a Back to Basics, no BS approach to CS that appeals to me. I agree with all the recommendations. A couple of tiny comments:
- I know that learning C is not strictly speaking part of Computer Science, but it is a nice counterpart to SICP, ties in with other topics (such as computer architecture and OS) and should definitely belong to this curriculum. The authors of this site themselves have defended C in another blog post. Like pg would say, all you need is Lisp and C.
- IMO a better option for learning databases is Jennifer Widom's MOOC: http://cs.stanford.edu/people/widom/DB-mooc.html
theparanoid · 2017-03-03
I first heard of her when she wrote about taking a year off and traveling with her family [0].
I've taken 6 months off and would love to do it again.
[0] http://cs.stanford.edu/people/widom/yearoff.html
colah3 · 2016-10-15
t-SNE is used very widely in ML, both to analyze input data and to analyze learned representations.
To me, the canonical example of people using t-SNE this way is visualizing word embeddings, like this: http://metaoptimize.s3.amazonaws.com/cw-embeddings-ACL2010/embeddings-…
Also, some lovely examples of using t-SNE to visualize conv net representations: http://cs.stanford.edu/people/karpathy/cnnembed/
I think people have found it very useful as one of the main tools for understanding what deep models are doing (along with optimization-based feature visualization) and to just check that your model is learning. I haven't really heard about people using it to pick input features, but that's probably mostly because I don't really work with anyone doing feature engineering.
romaniv · 2016-04-13
>Before deep learning people would manually design all these extra features sin(x_1), x_1^2, etc.
It's probably worth pointing out that this is true for ANNs, but there were (and are) other "shallow" classifiers that can handle swiss roll problem without manual parameter encoding. SVMs, for example.
http://cs.stanford.edu/people/karpathy/svmjs/demo/
typon · 2015-12-15
It has been discussed to death here why what they do is evil.
If you want a security expert's opinion, read Bruce Schneier's blog, and if you are inclined to learn more about the ethics, this page is great: http://cs.stanford.edu/people/eroberts/cs201/projects/ethics-of-survei…
deepnet · 2015-12-07
Thumbnailed Nips-2015 papers, sorted by top 100 concepts (a yearly LDA analysis by Stanford's Karpathy).
A great Interface for browsing this years papers (& previous NIPS).
https://cs.stanford.edu/people/karpathy/nips2015/
bdamos · 2015-06-14
This year's CVPR has a lot of great papers.
In case anybody hasn't seen, a full list's available at
http://cs.stanford.edu/people/karpathy/cvpr2015papers/
cs702 · 2015-05-21
Nice. Andrej Karpathy deserves some kind of award for demystifying deep learning and making the subject so accessible to a wider audience. If you're a developer who knows little about the subject and want to learn more, a great starting point is the home page for his ConvNetJS project.[1]
--
[1] http://cs.stanford.edu/people/karpathy/convnetjs/
tegeek · 2014-11-20
This November is pretty heavy for geeks & Singularity lovers (I'm one of them).
First we saw a man-made machine landing/crash-landing 500 millions KM from Earth.
Then HN discussed this [1] which is IMO one of the biggest machine-learning demo I've ever seen (other than Self-driving cars).
And now this futuristic tech which is aiming to eliminate the screens altogether. (As some of its use cases)
If Magic Leap delivers her promise within next 2-3 years, then we are already ahead than Singularity timeline [2].
1. http://cs.stanford.edu/people/karpathy/deepimagesent/
2. http://en.wikipedia.org/wiki/Predictions_made_by_Ray_Kurzweil#2019
poof131 · 2014-08-22
Rain is good. No matter how smart you are, you can still be an idiot sometimes:
"Von Neumann knew that it was only a matter of time before the Soviet Union became a nuclear power. He predicted that were Russia allowed to build a nuclear arsenal, a war against the U.S. would be inevitable. He therefore recommended that the U.S. launch a nuclear strike at Moscow, destroying its enemy and becoming a dominant world power, so as to avoid a more destructive nuclear war later on."
http://cs.stanford.edu/people/eroberts/courses/soco/projects/1998-99/g…
schakrava · 2014-04-01
SEEKING WORK - San Francisco, CA or REMOTE.
I'd love to work on Python+Django projects.
Here's my resume: http://cs.stanford.edu/people/sumanc/Suman_Chakravartula.pdf
github: https://github.com/schakrava
Contact me if you are interested.
karpathy · 2014-02-24
I did a similar analysis to the one posted here and computed a similar heat map to the one you describe, but I marked a submission as successful when it went from new -> front page, not when it hit 100 points. The result is in ~middle of the post and it seems that weekends are best for chances of a story making it to the front.
http://karpathy.ca/myblog/2013/11/27/quantifying-hacker-news-with-50-d…
and the raw ipython notebook with too many details:
http://cs.stanford.edu/people/karpathy/hn_analysis.html
tensor · 2011-11-23
Algorithms are core CS and I would definitely recommend taking them prior to other courses. The best way to know what to take first is to take a look at course prerequisites and the program tracks:
http://cs.stanford.edu/degrees/undergrad/ProgramSheets.shtml
nobody31 · 2011-11-02
Presumably the greatest minds in computer science sat down and decided that silverlight was the optimum technology
Or perhaps they were suspended over a tank of piranha by an evil laughing Bill Gates somewhere in the basement of http://cs.stanford.edu/info/gates
Or perhaps they just out sourced the web design to the cheapest company who happened to use Silverlight.
endtime · 2010-06-06
As elblanco said, any program worth its salt will have most of those as requirements or electives, even tiny ones like Georgetown. You should be able to check the school's CS department's website and see what their curriculum looks like (example: http://cs.stanford.edu/degrees/undergrad/ProgramSheets.shtml). My email is my profile - if you want me to take a look at any specific school I'd be happy to.
More generally, I think the anti-education sentiment you mentioned is unfounded. I'm colored by my experiences, of course, but I don't think I'd be half the engineer I am today without the education I've had, and I'm not a stupid guy.