kingcauchy

I wonder if we'll see a new sort of "role" in the training (user, system, assistant) for unstrusted sources, I'm a little surprised we haven't already. In fact it would probably make sense to have an arbitrary number of entity roles and to be able to configure the chat calls with truth values. Interesting article though.

That being said AI is not code, it's a statistical algorithm with non-determinism baked in. You can write code to run them but it's nothing without the evolution of the model weights from the training process. And you can absolutely make the model weights better aligned with intent.

How much of the apology was written by Claude? How much of the release note process was written by Claude? Will they have better prompts going forward to make sure Claude doesn't write upsetting things into the release notes for devs like silent nerfing? Spooky times.

« Trust us, we’re doing this for the good of humanity » (fills pockets with stock value and externalities from data center polloution) « No seriously trust us , at least we’re not Sam Altman »

Update: « Oh and we’re the only ones who will stop AI from turning into SkyNet and eating your babies, you just have to pay us to make sure we invent SkyNet first »

It's also hard to imagine them not doing this with any of the products they're building. "You can't use Claude to build an agent because that competes with Claude Code, you can't use Claude to build a design tool because that competes with Claude Design, you can't use Claude to build an email tool because that competes with Cowork."

"We won't use this product to spy or build weapons but you'll have to trust us, but we're also going to intentionally lie to you when you break our terms of service but trust us."

The silently never telling you is so insidious on top of it being ridiculous given how they trained the model in the first place. We do distributed model training for embedder/reranker models and I'd deeply resonate that this article's message exactly for our company. We couldn't trust the model in the first place, but now the model is intentionally burning our money if we asked it the wrong question, on top of being deeply expensive in the first place. If we did find evidence of being incorrectly nerfed, we'd never be able to reach a human to let them know. Too many reverse incentives with Anthropic, maybe they're about AI security but that doesn't make them ethical to consumers (i.e. humans).

I work with you so I'm bias but I think this is neat.

Thanks for the feedback!

In regards to contention, the answer is definitely dependent on how you host. We've had a lot of experience running different ML workloads and from an SRE perspective we knew you'd need a variety of different styles of hosting the models depending on read/write patterns of your usage. Termite and the proxy service/operator allow for all styles of model loading, either preloading and compiling to prevent cold starts or lazy loading to protect memory, with different pooling strategies and caching strategies for bundling multiple models running in the same Termite container.

If a heavy indexing job is running on a CPU only single-node deployment, it won't be using Raft (no replication). If it's running with GPU it doesn't share resources with the DB anyways really significantly there. If it's running distributed, also no issue with contention really.

Let us know if you have any other questions!

I'd be super interested to here more about what you all do in this space, currently Antfly (and Termite) doesn't handle custom content types explicitly because we've mostly focused on supporting the "classic" ones (application/pdf, image/png, image/jp2, e.g.) but we've had to build out a lot of the support for these things as custom support into the system. For instance I chose jsonschema for the schema so users could do exactly what you're suggesting, custom content types indexed differently. The ML side of things also has to know how to support them (i.e. does a pdf get rendered ocr then embedded or text extraction on some fallback). Would love to here about what you all do and the types of media you make searchable today!

https://github.com/antflydb/antfly/pull/8

Upon another look it looks like we were actually missing the pause lock for the backfill operation too during a shard split though, I also went ahead and added it to batch for good measure although that case should be caught by the manager! Thank you for the report!

I guessed that neo4j gets this question a lot so I went to see if they had a good blog post and this one is pretty good https://neo4j.com/blog/cypher-and-gql/10-things-you-can-do-w...

Possibly, Amazon and Google also made the ability for smaller startup based DB companies to go that route with things like ValKey and OpenSearch. LLMs have made it super easy to transpile the ideas into whatever programming language you please though, you just have to put in the time.

Nope! Awesome you’re poking around though. I’m currently working on deterministic simulation testing and a feature set to allow pausing of index backfills but it’s not fully implemented yet, stay tuned!

Great question! I think the fundamentally hard problem with distributed systems (at least for me!) comes down to the complicated distributed state machines you have to manage rather than the memory management problems. I think async rust gets in my way with respect to these problems more than it helps (especially when it comes to raft or paxos). That being said with the new async Zig, I’ve been excitedly implementing a swappable backend for the core database that I hope will be a nice marriage of performance and ergonomics.

Fascinating! We settled on Quic with Protobuf because it was more performant in our testing than the gRPC when coupled with the backoff, failure cases (node startup ordering server/client connections), and to not be coupled with the gRPC library versions in Go, which has bitten us a number of times when dealing with dependency management when you're trying to juggle k8s, etcd, and google dependencies in the same Go project. Plus the performance bottleneck in most of the use cases we're specializing in are on the embedding/ml side of things.

Thanks for the links! I hadn't seem yamux before!

I've added a specific example for that using the go-sdk https://github.com/antflydb/antfly/pull/5 here!

I can't speak for everyone, knowledge graphs are the "new hotness" of the ai space (RAG and MCP are seeing a lull in their hype cycles I guess). But I've used graphs professionally for a long time to connect relationships that SQL normal forms have trouble expressing non-recursively. E.g. I used graphs to define identity relationships between data sources hierarchically, and then had a another graph relationship on top of that to define connections between those identities, user at one level and organizations at the next. Graphs as indexes allow you to express arbitrary relationships between data to allow for more efficient lookups by a database. Some folks use it to express conceptual relationship between data for AI now, so if I have a bunch of images stored in google drive, I might want to abstract the concept of pets and pets have relationship with a human etc. then my database queries for looking up all pictures related to the dog-pets owned by some human becomes a tractable search instead of a scan of the corpus!

There's some examples in the quickstart on the website but I'll add an explicit e2e example case for that too. Otherwise the tests for that are a little lower level in the code! I'll add the RSF (merging of the two lists) example for that too!! Thanks for the feedback.

Thanks! Fixed that up!

Oh thanks for the 404 on the verify link (I abstracted out the auth OIDC for cross domain login and must have missed a path).

Yes good call, I tried to start that on the website with a react-flows based architectural flow chart a little bit but it's a bit high level, and not consumable directly in github markdown files but I'll work on that!

That's exactly the direction I've been working on, the reranking, embedders and chunkers are all plugable and the schema design (using jsonschema for our "schema-ish" approach allows for fine-grained index backend hints for individual data types etc.) I'll work on getting a good architecture doc up today and tomorrow!

Definitely open to working with you on supporting even better tooling for this as I imagine many different "styles" of migration will be necessary.

The number 1 supported migration path for users though is one of my personal favorite features of antfly which is the linear merge api, which allows you to incrementally reconcile an external pageable datasource with antfly at the pace you want while also getting the benefit of batching! We support index templates just like ES and the ability to change you schema and antfly manages the full-text reindex for you. If you're looking at migrating your embeddings in Elastic or another vectordb we can also support that! Let us know :)

Yes we do use SIMD heavily! https://github.com/ajroetker/go-highway I also added SME support for Darwin for most algorithms. We use it in the full-text index, all over the vector indexes and heavily for the ml inference we do in go especially.

Not strictly google but microsoft/bing too, here's the top ones from my notes:

https://arxiv.org/abs/2410.14452 spfresh, https://arxiv.org/abs/2111.08566 spann, https://arxiv.org/abs/2405.12497 rabitq, https://arxiv.org/abs/2509.06046 diskann,

I have a variety of blogs that I used too and reference implementations!

It's a Rabit[Q]uantized Hierchical Balanced Clustering algorithm we use for the vector index and we use a chunked segment index for the sparse index if you're curious! Happy to discuss more!

I built this for myself because I hated running a large ElasticSearch instance at work and wanted something that would autoscale and something that allowed for reindexing data. I also had a lot of experience running a large BigTable/Elasticsearch custom graph database I thought could be unified into a single database to cut costs. Started adding an embedding index for fun based on some Google papers and now here we are!

Thanks for the awesome questions!!

Exactly the use case I built it for! I wanted a world where you could build your indexes and the query planner could just be smart enough to use them in a single query. I've not quite nailed down the agentic query planner side 100% (it's getting there), but the JSON query DSL allows you to pipeline, join, fuse all the full-text, semantic, graph, reranking, pruning (score/token pruning) all in one query.

The CLI is my primary development tool with antfly, I am definitely looking for feedback on what people would like to see there, it's a little chonky with the flags --pruner e.g. requires writing the JSON for the config because I didn't want users to have to memorize 1000 subflags. It's definitely a first class citizen.

With respect to "Termite + single binary" that's exactly right, Termite handles chunking, multimodal chunking, embeddings (sparse + dense), reranking, fused chunking/embedding models, and we're excitedly getting more support for a variety of onnx based llms/ner models to help with data extraction use cases (functiongemma/gliner2/etc) so you don't have to setup 10 different services for testing vs deployment.

We run Antfly ourselves for our https://platform.searchaf.com (cheeky search AntFly) Algolia style search product in a distributed setup, and some users run Antfly in single node with large instances (more at the Postgres size datasets with millions of documents vs. large multitenant depoys). But we really wanted to build something with a more seamless experience of going back and forth between a distributed vs single node instance than elasticsearch or postgres can offer.

Hope that helps! Let me know if I can help you with anything!