jbellis

14 ms·

211.
by jbellis 2y ago
o3-mini-medium is slower than claude but comparable in quality. o3-mini-high is even slower, but better.
212.
by jbellis 2y ago
Especially with swe-verified, I thought that was the whole point of that dataset
213.
by jbellis 2y ago
Caffeine is a gem. Does what it claims, no drama, no scope creep, just works. I've used it in anger multiple times, most notably in Apache Cassandra and DataStax Astra, where it handles massive workloads invisibly, just like you'd
214.
Lessons Learned Writing LLMap(spyced.blogspot.com)
1 points by jbellis 2y ago|0 comments
215.
by jbellis 2y ago
I'm interested. I submitted my email to your landing page form.
216.
by jbellis 2y ago
Looks great, thanks for training this! - Can I fine tune it with SentenceTransformers? - I see ColBERT in the benchmarks, is there an answerai-colbert-small-v2 coming soon?
217.
by jbellis 2y ago
an order of magnitude less*
218.
by jbellis 2y ago
On the subject of scifi roguelikes, I've sunk 700+ hours into Jupiter Hell. I've been playing roguelikes since the original Rogue in the 80s and I think JH is the GOAT: skill tree that rewards experimentation, learning curve that
219.
by jbellis 2y ago
Actual SPANN or janky "inspired by SPANN" IVF with HNSW in front? Only real SPANN (with SPTAG, and partitioning designed to work with SPTAG) delivers good results. A superficial read of the paper LOOKS like you can achieve similar
220.
by jbellis 2y ago
High performance has never been a reason to use couchdb.
221.
by jbellis 2y ago
OOMs even in quantized mode on a 3090. What's a better option for personal use? > torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 43.71 GiB. GPU 0 has a total capacity of 24.00 GiB of which 20.74 GiB is free.
222.
A DIY photographer built his own full-frame camera and open-sourced the project(dpreview.com)
162 points by jbellis 2y ago|30 comments
223.
by jbellis 2y ago
True, but none of the slow methods in the article involve this.
224.
by jbellis 2y ago
Isn't this where "NPUs" are going now?
225.
by jbellis 2y ago
"Food deserts" are a demand problem, not a supply problem. https://www.nyu.edu/about/news-publications/news/2019/decemb...
226.
by jbellis 2y ago
Have you tried it? pgvector performance falls off a cliff once you can't cache in ram. Vector search isn't like "normal" workloads that follow a nice pareto distribution.
227.
by jbellis 2y ago
One of the few apps that could single handedly make an iPad pro interesting. :(
228.
by jbellis 2y ago
Technically it does run on windows, you just can't build the entire dataset without adding the sharding code mentioned. Set divisor=100 in config.properties and it will happily build an index over 1% of the dataset.
229.
by jbellis 2y ago
Unless you're counting your network access cost, the dataset of embeddings is in fact free and TFA includes instructions on how to download them for free.
230.
by jbellis 2y ago
> I’m not sure why we would ever do anything with segmentation according to that analysis if it were correct. What's your alternative when you can't build an index larger than C?
231.
by jbellis 2y ago
JVector (the index used in TFA) is available as a service with a friendly API from DataStax. https://www.datastax.com/products/datastax-astra [article author, I work on JVector and Astra]
232.
by jbellis 2y ago
I guess I'm kind of a CTO emeritus now -- I mostly write code, by choice. https://github.com/jbellis
233.
by jbellis 2y ago
I was trying to make the point that the dominant factor becomes linear instead of logarithmic, but more accurately it's O(S log N) = O(N log N) because S (number of segments) is proportional to N (number of vectors).
234.
by jbellis 2y ago
I'm baffled that so many people fixate on the estimated cost and miss the fact that it's a public dataset . As in, free.
235.
by jbellis 2y ago
I was trying to make the point that the dominant factor becomes linear instead of logarithmic, but more accurately it's O(S log N) = O(N log N) because S is proportional to N.
236.
by jbellis 2y ago
ColBERT gives you best of both worlds. https://arxiv.org/abs/2004.12832 https://thenewstack.io/overcoming-the-limits-of-rag-with-col...
237.
by jbellis 2y ago
It's split by language. TFA builds an index on the English language subset.
238.
by jbellis 2y ago
Oops, HN maxed out the free Cohere API key it was using. Fixed.
239.
by jbellis 2y ago
It's been a while since I read the source to pgvector but at the time it was a straightforward HNSW implementation that implicitly assumes your index fits in page cache. Once that's not true, your search performance will fall off
240.
by jbellis 2y ago
Probably, but I'm not aware of any Java library that provides that functionality.

More ›