samcodes

I wanted this article to convince me, but they really don't acknowledge the reason that everyone currently uses Python - the libraries. Do linear algebra using lightly wrapped C? numpy. What about NLP? spaCy. Implicitly specify a computation graph with high level code? PyTorch. Explicitly specify a computation graph using leaky C++ abstractions? TensorFlow. And using either PyTorch or TensorFlow, you get to interact with CUDA.

For now, if you want a functional language for doing deep learning, IMO that language needs Python interoperability. Long-term, I'm hopeful that GraalVM [0] can provide a way of calling Python from the JVM, but until then, I think the best option is coconut-lang [1], "a functional programming language that compiles to Python." You get pattern-matching, TCO, the pipe operator, and ADTs, all while being one AOT compilation step away from Python.

[0] https://www.graalvm.org/

[1] http://coconut-lang.org/

When I was doing cosmology, the scales we were talking about were absurd - "model this galaxy cluster as a point" scales. So it seems conceivable to me that enough GEODEs, well-distributed, could look uniform on this scale.

I think most recommendation algorithms rely too much on collaborative filtering and not enough on content based filtering.

When I open source code, I use MIT. I have introduced open source strategies at places I work, and guided teams through the process.

Ok now that I've established my OSS cred: I'd be really interested in a license that is effectively MIT, _except_ for the 20 largest software companies, for whom there is no license and any use is restricted (you'd need to clearly define what makes a company a software company, and what it means to be the largest, but that's what lawyers are for). They could pay for a commerical license if the customer demand was high enough.

This seems like it would address the main issue as I see it, that is, I want almost everyone to be able to use my software to build cool shit and make money... But if you're FAANG or some shit, come on. The rules are different.

They should have put some round ears on it...

One time we put these rules in a graph database. It worked better than hard-coding and a rules engine.

Required reading on the subject: https://erikbern.com/2018/05/02/interviewing-is-a-noisy-pred...

I realize that there are problems with it, but I'm a huge fan of take-home problems that are similar to real problems the candidate would face on the job.

Raising children is the primary function of society. Someone going beyond their means in order to raise more, harder to raise children should be celebrated. That we would see any fault is damning evidence that our society is malfunctioning.

Having been through the same, I'm of the opposite opinion. We're just computers made of meat. Some of us weren't optimized to function in society. Any help that works, chemical or not, is welcome.

Have you seen TPOT [0]? AutoML library that uses genetic algorithms to write scikit code for you. So fun.

[0] https://github.com/EpistasisLab/tpot

Well... Time to change careers. A "sex" and "travel" blog called "exploregasm" needs to exist.

Good point, timecard-tracking software and general purpose differentiable computing languages are roughly the same level of complexity.

This is true, but rapidly changing. In addition to fine tuneable language models, you can do deep feature extraction with something like bert-as-service [0] ... You can even fine tune Bert on your days, then use the fine tuned model as a feature extractor.

[0] https://github.com/hanxiao/bert-as-service

This should be completely backend agnostic. I was never a Django person, but you’d just put that script tag in your main template so it loads on every page.

The hard part is that after the convolutions you want a fully connected layer or two, and to get those dimensions right you need to know the input dimensions. But, pytorch is building the graph at runtime, so maybe you could do this...

IME, the naive solution to hierarchical classification - building a classifier for each level of the hierarchy - gets me to ~85% accuracy, compared to ~75% accuracy using a "global" classifier.

The post mentions not getting great results with OpenAI Transformer. I haven't tried that, but using a similar framework, ULM-FiT, I narrowly beat the fasttext benchmark on a 250-class dataset we use internally. I will follow up with how it does on this data set.

But overall it _did_ capture most of the structure. Other than a few adjectives, everything it required was real, and most of the combinations were exotic but good.

However, thanks for sharing the Kazama paper, that looks really interesting.

Someone already mentioned Luna Lang. Another similar project that is further along is isomorf. You get projections into 6 or so syntaxes, and deployment to AWS Lambda.

https://isomorf.io/#!/syntax-picker/tours-ideas

Did you look at Up by Apex for inspiration? Very similar framework, but for Lambda languages, any language that compiles to a binary, or Crystal: https://up.docs.apex.sh/

There are those who get it done RIGHT, and those who get it done RIGHT NOW.

As someone with degrees is math and physics, I used to think this. I don't anymore. I think that those fields offer many opportunities to examine and analyse ideas, and that is important.

Honestly, if the primary value proposition was "fuck your cable company, here's how to get internet without paying them" you could probably motivate some sales even if the price and service were the same.

This is a great deck... but leaving TLA+ out of Formal Methods? It is hands down the most important thing going on there. And it is a combo between academia and industry. Amazon uses it extensively.

Because when someone makes the argument that "ok now discrimination has gotten out of hand" it usually is implied "... Now that I'm the target!" Which would imply the author isn't used to being the target of discrimination, which would imply middle aged white male.

The point I was trying to jokingly bring up is a valid one, I think. How much of what we do leaks identity-related metadata? How much are of the responses we get are responses to that inferred metadata?

I didn't say to ignore his "valid points." I intentionally didn't mention the content. I was trying to point out that, if, as you read the piece, you correctly inferred metadata about the author, (a) it's funny because (b) it means there might be more to the theories he is dissmissing out of hand.

Ok but... While reading that, did you at any point wonder the ethnicity, gender and approximate age of the author? Because I was less than surprised when I saw him at the end of the article.

This is great! I've looked for exactly this in the past.

I TAed at a Cal State and worked in the math tutoring center. If the tests were like other standardized tests I encountered there, they probably were not a very good measurement of anything. And the remedial courses, best I could tell, did not help.

Some people seem to be interpreting this as lowering the bar. That position supposed that these tests are more accurate measurements than grades and SAT scores. I believe you are putting way too much faith in these tests. Also, the plan of "stretch" courses seems like a better one - motivation is a large factor in performance, and students will be more motivated in classes that count towards their degree, and don't tell them they preemptively failed.

K-12 education needs lots of improvement. But the Cal State system can't affect that. This is a change they can make, and it seems like it will help bring the kids who need it up to speed faster.

Looking ahead 10-20 years, I don't see how anyone born from 2015 onwards has any solid concept of reality the way I feel I know it. Between things like this and AR, I feel like "real" vs "simulated" will seem to the AR natives to be a pedantic distinction... Like how my parents still distinguish between "having met someone" and "someone I have talked to on the internet."