The difference: LUX-ZEPLIN, which is underground, is waiting to detect a dark matter particle itself. On the other hand, NGRST seeks to observe the effects of dark matter.

Previously: https://news.ycombinator.com/item?id=42021237 Linux on Apple Silicon with Alyssa Rosenzweig

Comments URL: https://news.ycombinator.com/item?id=49459140

Points: 35

# Comments: 6

For example, with frontend, I can get an LLM to design half a page 80% to my satisfaction with two paragraphs worth of a prompt. "It should look like so, and have a text box here, and room for a demo there".

With ML training or backend or user-facing code, I might instead spend a paragraph thinking out my design intentions for a single function or even a single line, more for myself than the LLM. A harness generates a plan based on that paragraph, which one can then comment on and review the pseudocode it provided, ensuring it aligns with expectations.

Lastly have the LLM output some sort of documentation and "here's what I did" after each change. Your final step is to handwrite (paraphrasing what it gave) into any docs or commit messages, and ensure your commits are small enough to keep this maintainable. Paraphrasing the LLM, rather than the LLM paraphrasing you, is helpful to ensure the commit messages make sense to you three months from now.

Comments URL: https://news.ycombinator.com/item?id=48909137

Points: 2

# Comments: 1

https://x.com/keshavchan/status/1787861946173186062

In my opinion, whether it was actually by Ilya or not is not worthy of debate. Many of them are widely recognized for being good pedagogical resources (e.g. annotated transformer, unreasonable effectiveness of RNNs, understanding LSTM networks), and others are landmark papers which anyone interested in the field would benefit from reading:

- Krizhevsky et al. (2012) introduced AlexNet

- Bahdanau et al. (2014) introduced attention

- He et al. (2015) introduced ResNet

- Vaswani et al. (2017) introduced the Transformer

Other papers are more specialized. Of them, I think Kaplan et al. (2020) by OpenAI is probably most important.

Also useful is Steve Brunton's channel [3]. He has a freely available book [4] co-authored with Nathan Kutz that ties machine learning and control.

[1] https://engineeringmedia.com/ [2] https://www.youtube.com/@BrianBDouglas [3] https://www.youtube.com/@Eigensteve [4] https://news.ycombinator.com/item?id=36374528

In fact, on the second last page of the paper, they discuss this very problem. There is a clear correlation between performance and increasing sequence lengths for the Q-K=V model. While limited to a tight n=3 sample between 512, 1024, 2048 lengths, the degradation decreases from 5.4% to 2.2% as context is increased, suggesting that it is unlikely shorter sequences are the reason K=V performs acceptably.

In addition, replicating Jameson et al. (AIAA 1981-1259) [1], is a worthwhile, more advanced follow up, great if you want to get into serious CFD development eventually.

[1] http://aero-comlab.stanford.edu/Papers/jameson.aiaa.1981-125...

When I started out, I was under the assumption that I had to understand at least the undergraduate real analysis curriculum before I could grasp quantum algorithms. In reality, for the main QC algorithms you see discussed, you don't need to understand completeness; you can just treat a Hilbert space as a finite-dimensional vector space with a complex inner product.

For those unfamiliar with said concepts from linear algebra, there is a playlist [1] often recommended here which discusses them thoroughly.

[1] https://www.youtube.com/playlist?list=PLZHQObOWTQDPD3MizzM2x...