AI / machine learning engineer / founder of @unboxresearch; coding in Python, C, js, Lua, Obj-C; math for fun. @tylern@mastodon.social

Berkeley, CA
Joined March 2009
I just published a fun article on one of my favorite lesser-known algorithmic ideas: locality-sensitive hashing. Fuzzy search in constant time! unboxresearch.com/articles/l…
3
69
3
279
Summary and Analysis: So Good They Can’t Ignore You, open.substack.com/pub/tylern…
1
4
413
8/ ExtrapolaTED Generate extremely well-researched TED talks on demand RAG powered by 2 million Arxiv papers and 6 million Wikipedia pages, and 3 million photos 🥉Third place @tylerneylon @roberthdly @ashvardanian @wordware_ai
1
3
17
3,001
Wordle 798 3/6 ⬜⬜⬜⬜⬜ ⬜⬜⬜⬜⬜ 🟩🟩🟩🟩🟩 The stealth approach. You sneak up on the word.
1
4
461
If you want to learn about LLMs beyond blog posts and tutorial videos: I'm teaching a professional LLM workshop on Thursday. It's a one-day course on setting up and customizing self-hosted LLMs. unboxworkshops.com/self-host…
1
1
2
559
At the end of the day, you'll know how to download, run, integrate data into, and fine-tune an LLM on a computer you own or rent. I worked with Gabriel Bianconi to create the course I wish I could have taken. We worked through a lot of accidental complexity so you don't have to.
1
264
New github gist: a js function that uses a partial order to fully sort any array. By "partial order," in this case I mean a comparison function that can sometimes say "I don't know;" as long as it doesn't contradict itself, we can still sort the array. gist.github.com/tylerneylon/…
1
1
400
I'm wrote this so I can efficiently choose the drawing order for shapes being rendered in a 3d scene. GPUs handle this with z-buffering, but I'm aiming to convert a scene to a vector-based, pdf-friendly set of ordered drawing primitives. A different kind of output.
1
287
I'm thrilled to be co-teaching a GPT / LLM workshop with the amazing @GabrielBianconi. I'd love to hear from you: Which sub-topic should we focus on first?
33%Self-hosting LLMs
33%Productionalizing LLMs
33%Custom info/DBs for LLMs
3 votes • Final results
2
440
I launched a machine learning newsletter! It's called Learn & Burn. I work with an amazing team at Unbox Research to publish a weekly summary of AI/ML news around arxiv papers. We explain, clearly and quickly, new ideas as they're shared by researchers. learnandburn.substack.com/
3
286
Feeling a bit judged by my AI pair programmer right now. (I start typing TODO, copilot doesn't hold back.) 😆
1
4
597
Q's about vector compression: 1. Are there well-known general techniques here? Not for images/audio/text, but for preserving relative cosine similarity? 2. Relatedly, is PCA known to preserve relative cosine similarities? (More context below)
1
1
1
How SVD helps: Let D be a wide matrix (D is m×n, m<n). the cols of D are the vecs we'll compress. Let D = UΣVᵀ be the reduced SVD, so U is m×k. (k < m) Let C = UᵀD, which is k×n; its cols are the smaller vecs. ‖dⁱ-dʲ‖ ≈ ‖Ucⁱ-Ucʲ‖ = ‖U(cⁱ-cʲ)‖ = ‖cⁱ-cʲ‖.
1
In that eqn, the ≈ is a property of the SVD. The SVD ensures that the input matrix D is approximated by the reconstruction UΣVᵀ in terms of sums of element distances (aka Frobenius norm). And UC is the same as UΣVᵀ. So the column dⁱ should be close to the column Ucⁱ.
This is one of my favorite open problems in math. Cool to see progress on it!
One of the best known open problems in combinatorics is the union-closed conjecture, which states that if you have a finite collection X of sets such that if A and B belong to X then so does the union of A and B, then at least one element of X belongs to at least half of them. 1/
2
Coding approaches, ranked by runtime efficiency from slowest to fastest: * Python * JavaScript * C / C++ * Fortran * CPU assembly * Code for a GPU/TPU * Design custom hardware * Be John Carmack
1
1
To be fair to Python: It's one of my favorite languages, I use it constantly, and it makes good use of *my* time, which is typically my bottleneck :)
1
2