HNHacker News
TopNewBestAskShowJobs

fzliu

4,973 karma · joined October 20, 2014

Applied ML @ Voyage AI (https://voyageai.com).

I love chess, deadlifting, and Elden Ring.

Twitter: https://x.com/frankzliu

LinkedIn: https://www.linkedin.com/in/fzliu

submissionscomments
fzliu··on New and Improved Embedding Model for OpenAI
Solid work OpenAI, though I'd definitely like to see some more benchmarks on a wider variety of datasets in addition to the ones listed in the post. Regardless, it's good to see embeddings becoming more and more mainstream and easier to leverage out-of-box. We tried image embeddings many moons ago (2015) with AlexNet trained across a custom dataset, but we still had to add quite a few custom roles post-inference.

A large selling point for ada-002 embeddings seems to be the reduced dimensionality. While lower-dimensional embeddings definitely help performance, I would say it's still highly dependent on the index that's being used. Graph- and tree-based indexes will benefit less than ones based on IVF (https://zilliz.com/blog/vector-index), as they do fewer overall distance computations during query time, but the speedup should still be significant.

Still been meaning to try semantic search across Wikipedia via text embeddings. Will definitely play around with OpenAI + Milvus (https://github.com/milvus-io/milvus).

fzliu··on Leaving Spotify for freer pastures
Is your music recommendation system open source? Would be down to check it out and learn a thing or two from it.

On the topic of vector search, I'm fairly certain that Spotify still uses Annoy (https://github.com/spotify/annoy). Like Faiss, it's a great library but not quite a database, which would ideally have features like replication (https://milvus.io/docs/replica.md), caching, and access control, to name a few.

fzliu··on Notion AI – waiting list signup
Thanks for the shout-out! For folks interested in playing around with vector and/or hybrid search: Milvus is open-source (https://github.com/milvus-io/milvus).
fzliu··on Similarity search and deduplication at scale
I'm surprised to see that ML-based semantic search is barely touched on in this article. There's a strong focus on entity matching, but an arguably more powerful way to conduct similarity search is to leverage embedding vectors from trained models.

A great upside to this approach is that it works for a variety of different types of unstructured data (images, video, molecular structures, geospatial data, etc), not just text. The rise of multimodal models such as CLIP (https://openai.com/blog/clip) makes this even more relevant today. Combine it with a vector database such as Milvus (https://milvus.io) and you'll be able to do this at scale with very minimal effort.

fzliu··on Ask HN: Data Scientists, what libraries do you use for timeseries forecasting?
PyTorch for recurrent nets (TensorFlow would work too).
fzliu··on Sonic: Fast, lightweight and schema-less search backend
+ Milvus (https://github.com/milvus-io/milvus) for large scale similarity/semantic search.
fzliu··on Vectors are over, hashes are the future
Hashes are fine, but to say that "vectors are over" is just plain nonsense. We continue to see vectors as a core part of production systems for entity representation and recommendation (example: https://slack.engineering/recommend-api) and within models themselves (example: multimodal and diffusion models). For folks into metrics, we're building a vector database specifically for storing, indexing, and searching across massive quantities of vectors (https://github.com/milvus-io/milvus), and we've seen close to exponential growth in terms of total downloads.

Vectors are just getting started.

fzliu··on A CTO should be technical
I'm going to put this out on a limb - CEOs should also be technical.
fzliu··on TimescaleDB 2.7 vs. PostgreSQL 14
Seconded. I understand that Timescale extends Postgres, but I'd still like to see performance comparisons relative to other time-series databases such as InfluxDB or even TiDB.

Shameless self-plug: we're also building database technology, but for embedding vectors (https://milvus.io) rather than time-series data - when doing query performance comparisons, we get several orders of magnitude of performance improvement over traditional databases. It's an unfair comparison, so we generally avoid it.

fzliu··on ZincSearch – lightweight alternative to Elasticsearch written in Go
As an Yahoo internal project, Vespa actually started out very similar to Elastic in that it was meant for fast text search. The team then added ANN search capabilities later on.

Vector search infrastructure is overall still fairly new (https://milvus.io/docs/text_search_engine.md), but up and coming for sure. Disclaimer: I'm a part of the Milvus community.

fzliu··on Measuring the popularity of different vector databases
Great article. Vector databases are still somewhat niche, but there's been rapid growth in interest (I'm a part of the Milvus community). https://db-engines.com also provides a holistic popularity metric (Milvus and Weaviate are on there), but I'm not exactly sure how it's calculated.

One quick note - FAISS is an ANN library rather than a vector database. Vector databases support a multitude of other features that you'd see in a traditional database such as caching, replication, horizontal scalability, etc. I've also found Elastic's ANN search functionality a bit on the slow side, likely due to its original architecture being focused on more general text search.

fzliu··on Pollen’s Collapse: “$200M Raised” but Staff Unpaid
Collapse implies that it was sudden and unavoidable. All of this was the result of an executive team that didn't want to sit down and have tough conversations with staff.

As the article mentions, tons of other VC-funded companies have gone bankrupt before, but those bankruptcies did not end in disaster like Pollen did.

fzliu··on Pollen’s Collapse: “$200M Raised” but Staff Unpaid
This is not a collapse. It's hubris, poor financial management, and a complete breach of trust.
fzliu··on Algolia Acquires Search.io
I'd like to add that vector search works not just for natural language, but also for a variety of other types of unstructured data as well. Images, video, user profiles, and pretty much anything else that can be vectorized. Here's an example of image search: https://milvus.io/docs/image_similarity_search.md
fzliu··on Transformers seem to mimic parts of the brain
Saying that neural networks are "similar to" or that they "mimic" the human brain can be misleading. Today's architectures are the byproduct of years of research and countless GPU-hours dedicated to training and testing architecture variants. Many neuroscience-based architectures that mimic the brain better than transformers end up performing much worse.

The Quanta article is overall pretty reasonable, but I've unfortunately seen other news outlets regurgitate this kind of blanket statement for the better part of a decade. The very first models were perhaps 100% inspired by the brain, but today's ML research more or less follow a "whatever works best" principle.

fzliu··on Cheating at chess with a computer for my shoes
Good catch. I was searching for a game without Nf3, but this is effectively a transposition.

One minor point folks aren't mentioning that I think works in Neimann's favor - Carlsen is known for enjoying the Catalan, and the opening in the game was essentially a Nimzo/Catalan hybrid. This fact makes Neimann's supposed pre-game prep a bit more palatable. Everything else (postgame interview, past history, Twitter antics) seems to work against Neimann though.

My question is - if Neimann did cheat, how? Doesn't seem like something that can be done easily in an OTB tournament.

fzliu··on Cheating at chess with a computer for my shoes
Seconded re: Neimann's analysis. Specifically, he stated in a post-game interview that Carlsen had played the same Nimzo-Indian g3 line in a prior tournament, even though that never happened.

For the sake of the sport, I hope Neimann didn't cheat, but I doubt we'll ever know for sure.

fzliu··on Computing VIN Checksums
I'm always amazed by the lack of tech saviness in a lot of "traditional" industries.
fzliu··on Solr’s Dense Vector Search for indexing and searching dense numerical vectors
There's no one answer to this, but I'd say that anything past 10k vectors would benefit greatly from a vector database. A vector DB will abstract away the building of a vector index along with other core database features such as caching, failover, replication, horizontal scaling, etc. Milvus (https://milvus.io) is open-source and always my go-to choice for this (disclaimer: I'm a part of the Milvus community). An added bonus of Milvus is that it supports GPU-accelerated indexing and search in addition to batched queries and inserts.

All of this assumes you're okay with a bit of imprecision - vector search with modern indexes is inherently probabilistic, e.g. your recall may not be 100%, but it will be close. Using a flat indexing strategy is still an option, but you lose a lot of the speedup that comes with a vector database.

fzliu··on Ask HN: I want to dive into how to make search engines
This depends mostly on what kind of search engine you're trying to build. I unfortunately won't be able to point you towards courses, but there are tons of great resources online to help you get started.

Search engines are a fairly broad topic, and a lot of it depends on the _type of data_ that you want to build a search engine for. If you're looking towards more traditional, Google/Yahoo-like search, Elasticsearch's learning center (https://www.elastic.co/learn) has quite a few good resources that can point you in the right direction. Many enterprise search solutions are built on top of Apache Lucene (including Elasticsearch), and videos/blogs discussing Lucene's architecture (https://www.endava.com/en/blog/Engineering/2021/Elasticsearc...) is a great starting point as well.

Opposite text/web search is _unstructured data_ search, i.e. searching across images, video, audio, etc... based on their semantics (https://www.deepset.ai/blog/semantic-search-with-milvus-know...). Work in this space has been ongoing for decades, but an emerging way of doing this is via a _vector database_ (https://frankzliu.com/blog/a-gentle-introduction-to-vector-d...) such as Zilliz Cloud (https://zilliz.com/cloud) or Milvus (https://milvus.io/). The idea here is to turn the type of data that you're searching for into an high-dimensional vector called an embedding, and to perform nearest-neighbor search on the embeddings themselves.

Disclaimer: I'm a part of the Zilliz team and Milvus community.

fzliu··on Non-Machine Learning Image Matching with a Vector DB
Cool to see a vector search application that leverages traditional image features. I'm curious to know which of these methods performed best on cropped images, which remain somewhat of a challenge for traditional classification models (contrastive models trained specifically with image crops tend to work much better).

Also for the vector database itself, have you considered spinning up your own open-source alternative such as Milvus (https://github.com/milvus-io/milvus), or were you only considering managed services?

fzliu··on Vector search just got up to 10x faster and vertically scalable
On the topic of publishing, these two papers from the Milvus community may be of interest to some folks.

SIGMOD'21 - https://www.cs.purdue.edu/homes/csjgwang/pubs/SIGMOD21_Milvu... This paper talks about the vector database vertical (compute core and user-facing API)

VLDB'22 - https://arxiv.org/pdf/2206.13843.pdf This paper discusses the development of a cloud-native vector database.

Reading through them should help folks understand where the novelty and difficulty in developing an full-fledged vector database comes from.

Disclaimer: I'm a member of the Milvus community.

fzliu··on Vector search just got up to 10x faster and vertically scalable
Larger disk space does help effortlessly scale storage in single-node systems, but I agree with you that shared nothing (and/or shared something) is a necessary step for extracting maximum performance on a larger machine. When it comes to distributed architectures, shared nothingness is important as well. The decoupling of storage from compute and stateless from stateful helps minimize resource allocation when it comes to billion-scale vector storage, indexing, and search. Milvus 2.0 implements this type of architecture - here's a link to our VLDB 2022 paper, if you're interested: https://arxiv.org/abs/2206.13843
fzliu··on Vector search just got up to 10x faster and vertically scalable
Just to clarify, Milvus is much more than a wrapper around FAISS. Our vector search component called Knowhere (https://github.com/milvus-io/knowhere) utilizes FAISS and Annoy and will soon include ScaNN, DiskANN, and in-house vector indexes as well. Milvus uses Knowhere as the compute engine, and implements a variety of database functions such as horizontal scaling, caching, replication, failover, and object storage on top of Knowhere. If you're interested, I recommend checking out our architecture page (https://milvus.io/docs/architecture_overview.md).

[EDIT]: Forgot to mention - Milvus development began in 2018 was open sourced in 2019.

fzliu··on WeWork's former CEO has a new startup, reportedly valued at more than $1B
This is the single largest investment into any company by a16z. This is just insane, especially given that Flow does not yet have a solid product.
fzliu··on 'Too many employees, but few work': Pichai, Zuckerberg sound the alarm
> There are plenty of people who would love to be dead weight just to float around in a company.

Very true. It's an unfortunate occurrence at many larger companies (not just in tech).

fzliu··on The semantic web is dead – Long live the semantic web
I'm always surprised when articles like these don't talk about the meaning behind the original phrase. The semantic web came to be because there existed a need for computers to understand the contents of webpages, which were invariably human-generated.

We now have a huge set of tools for that under the broader AI/ML umbrella - ML is obviously imperfect, but its cautious utilization across various industries is, to me, a step in the right direction. There's simply no need to pigeonhole ourselves into a "semantic web" data model that might not fit a particular topic or application.

I personally think that embeddings (https://milvus.io/docs/v1.1.0/vector.md) will eventually take over semantic search applications. NLP, for better or worse, have seen great success naively throwing text into massive pre-trained models (I say naively because a lot of these models still think Obama or Trump is president). We've also made great progress unifying the architecture of NLP and CV models via transformer architectures, and we're now seeing lots of CV applications follow suit.

fzliu··on Ask HN: Who is hiring? (August 2022)
Zilliz (https://zilliz.com) | Remote and On-site | San Francisco, CA

Zilliz is built by the engineers and scientists who created Milvus (https://milvus.io), an open-source vector database for production-ready AI. On a mission to unleash data insights with AI, we build next-generation database technologies to help organizations create AI/ML applications stunningly easy and fast. By taking the burden of complex data infrastructure management off our users, we are committed to bringing the power of AI to every corporation, every organization, and every individual.

We've raised significant funding (we're just about to announce our latest round), and are hiring for multiple positions:

- Head of Marketing: https://boards.greenhouse.io/zilliz/jobs/4057464005

- Head of Developer Relations: https://boards.greenhouse.io/zilliz/jobs/4031247005

- Senior Software Engineer: https://boards.greenhouse.io/zilliz/jobs/4057484005

- Machine Learning Platform Engineer: https://boards.greenhouse.io/zilliz/jobs/4057480005

- Technology Evangelist: https://boards.greenhouse.io/zilliz/jobs/4031246005

- Developer Community Manager: https://boards.greenhouse.io/zilliz/jobs/4057465005

Interested in learning more? Feel free to get in touch with us at careers@zilliz.com.

fzliu··on Should I do startup or research in my 20s?
I'm fairly certain that cognitive abilities do not decline across the board upon reaching 30. Ability to learn probably does, but other aspects of intelligence such as memory and emotional intelligence probably do not peak until much later in life.
fzliu··on Should I do startup or research in my 20s?
Do whichever you think will allow you to make the biggest impact.
← PreviousPage 5 of 6Next →