jbarrow

7 ms·

31.
Turning my laptop into a Search Relevance Judge with local LLMs(softwaredoug.com)
3 points by jbarrow 2y ago|0 comments
32.
by jbarrow 2y ago
This trick is also the way to teach adults, if you're teaching or learning to ride! For children, there are companies that sell progressive balance bikes[1][2], that start off as balance bikes but can be converted to pedal bikes later.
33.
Question Answering is a Format; When is it Useful?(arxiv.org)
1 points by jbarrow 2y ago|0 comments
34.
by jbarrow 2y ago
Research community is definitely not agreed on this, and there are a number of different camps. Broadly, 2 perspectives from the NLP community: The 2020 Bender and Koller paper[1] that argues that meaning is not learnable from form, and LLM
35.
Do LVLMs Understand Charts?(arxiv.org)
1 points by jbarrow 2y ago|0 comments
36.
by jbarrow 2y ago
Went through the structured outputs chapter, and honestly feels like a good resource! Covered a lot of the gotchas and current research into whether or not structured outputs hurts (e.g. one of the linked papers: https://arxiv.or
37.
by jbarrow 2y ago
There's a company doing that for stovetops, which I found really interesting ( https://www.impulselabs.com )! Unfortunately, when training on a desktop it's _relatively_ continuous power draw, and can go on for days. :&#
38.
by jbarrow 2y ago
Primarily concerned about the memory bandwidth for training. Though I think I've been able to max out my M2 when using the MacBook's integrated memory with MLX, so maybe that won't be an issue.
39.
by jbarrow 2y ago
I’m really curious what training is going to be like on it, though. If it’s good, then absolutely! :) But it seems more aimed at inference from what I’ve read?
40.
Show HN: Tinyhnsw – The Littlest Vector Database(github.com)
17 points by jbarrow 2y ago|0 comments
41.
by jbarrow 2y ago
Yes, though how you do it depends on what you're doing. I do a lot of training of encoders, multimodal, and vision models, which are typically small enough to fit on a single GPU; multiple GPUs enables data parallelism, where the data
42.
by jbarrow 2y ago
The increasing TDP trend is going crazy for the top-tier consumer cards: 3090 - 350W 3090 Ti - 450W 4090 - 450W 5090 - 575W 3x3090 (1050W) is less than 2x5090 (1150W), plus you get 72GB of VRAM instead of 64GB, if you can find a motherboard
43.
Google Gemini 101 – Object Detection with Vision and Structured Outputs(notes.penpusher.app)
1 points by jbarrow 2y ago|0 comments
44.
by jbarrow 2y ago
> did we exclude drivers who were not eligible to drive I don't think we can discount people who aren't fit/eligible to drive _if they still drove_. They might not have a _legal_ right to be on the road, but they are still
45.
by jbarrow 3y ago
I put together a slow (but readable!) HNSW implementation in python to really understand how it works: https://github.com/jbarrow/tinyhnsw Indexing time isn't great, but query time is surprisingly good for it bein
46.
by jbarrow 3y ago
Typically not. Google as an example: the transformer paper (Vaswani et al., 2017) was arxiv'd in June of 2017, and NeurIPS (the conference in which it was published) was in December of that year; BERT (Devlin et al., 2019) was similarl
47.
by jbarrow 3y ago
I'm not certain how many other researchers would be willing to swap away from PyTorch (or tf or jax) to yet another framework. Amazon pushed MXNet for a while, but the only time I used it was while interning there. From what I understa
48.
by jbarrow 3y ago
This explains a lot; I’ve been thinking for years (ever since being diagnosed with a chronic health condition) about writing an essay likening the American insurance system to Kafka’s _The Trial_. The bureaucracy is absolutely soul-crushing
49.
by jbarrow 3y ago
The order of magnitude of suggested pricing is really interesting: $0.001/word is significantly more expensive than, say, OpenAI's pricing of GPT-3.5-turbo ($0.002/1k tokens, ~750 words, so ~$0.000003/word, assuming I go
50.
OpenFlamingo v2: New Models and Enhanced Training Setup(laion.ai)
6 points by jbarrow 3y ago|0 comments
51.
Multi Hash Embeddings in SpaCy(arxiv-vanity.com)
1 points by jbarrow 4y ago|0 comments
52.
by jbarrow 6y ago
In terms of a Kindle-like laptop for writing, I’ve seen people enjoy the Freewrite Traveler (though it’s not for me): [1] https://getfreewrite.com/products/freewrite-traveler
53.
by jbarrow 12y ago
Actually, that's exactly how they speed up neural networks, especially deep neural networks. Andrew Ng showed that they could run the Google Brain on COTS GPUs for about $21,000 [1]. You can experiment with this yourself using a packag
54.
by jbarrow 12y ago
It's reasonably fast, as it uses HMatrix for linear algebra -- HMatrix relies on BLAS rather than native Haskell for all the matrix and vector math.
55.
Show HN: LambdaNet – A functional neural network library written in Haskell(github.com)
56 points by jbarrow 12y ago|7 comments
56.
by jbarrow 12y ago
In terms of understanding neural networks, I generally point people to the same resource: http://neuralnetworksanddeeplearning.com Reading that will give you a good idea of the capabilities and limitations of neural networks, as
57.
Coding for Lawyers(codingforlawyers.com)
91 points by jbarrow 12y ago|43 comments
58.
by jbarrow 12y ago
Using Safari, which isn't supported, but I wanted to say kudos to the author for adding a video for unsupported browsers. That was definitely a neat feature that many WebGL demos overlook!
59.
by jbarrow 12y ago
You can preorder now and they'll ship soon. The consensus is that they're a little pricey, though: [1] http://openbci.com
60.
by jbarrow 12y ago
I think this is an interesting proof of concept; it seems like the exact opposite of WhatsApp, which is based on the concept that data is cheaper than SMS. I know that SMS plans are often "unlimited" in the US, which is why (at le

More ›