Hacker Newsnew | past | comments | ask | show | jobs | submit | jascha_eng's commentslogin

This is legitimately pretty impressive. I think the rule of thumb is now, go with postgres(pgvector) for vector search until it breaks, then go with turbopuffer.

seems like a good rule of thumb to me! though i would perhaps lump "cost" into the "until it breaks" equation. even with decent perf, pg_vector's economics can be much worse, especially in multi-tenant scenarios where you need many small indexes (this is true of any vector db that builds indexes primarily on RAM/SSD)

It was not a fire sale I'm pretty sure. Langfuse has been consistently growing, they publish some stats about sdk usage etc so you can look that up.

They also say in the announcement that they had a term sheet for a good series a.

I think the team just took the chance to exit early before the llm hype crashes down. There is also a question of how big this market really is they mostly do observability for chatbots but there are only so many of those and with other players like openais tracing, pydantic logfire, posthog etc they become more a feature than a product of its own. Without a great distribution system they would eventually fall behind I think.

2 years to a decent exit (probably 100m cash out or so with a good chunk being Clickhouse shares) seems like a good idea rather than betting on that story to continue forever.


I don’t know about that. I looked at them a couple of months back for prompt management and they were pretty behind in terms of features. Went with PromptLayer

say more about what you were looking for that promptlayer had/langfuse didnt?

Ohh. Reply from Shawn. I love your work. As far as I recall, we were not looking for a net of features but specifically a git like API that could manage and version the prompts. Meta data tagging, Jinja2, release labels and easy rollback. Add that up with Rest, Typescripts and Python support and it worked pretty well. Langfuse seemed way better at tracing though.

hi! :) thanks for supporting my work. ok yea that makes sense - i was a langfuse only user but sounds like i might want to check out promptlayer (tbh i was until recently of the opinion that you should always check in your prompts into git... flipped recently)

Took a few days off to continue working on Kviklet a PR like review/approval tool for SQL queries/statements.

Recently added role sync for Saml, oidc and ldap. These auth protocols are complicated though.

https://github.com/kviklet/kviklet


Can't read this every paragraph ends with it's not x it's y. Just give me the prompt so I can read the real insights you have and not the generated fluff.

I randomly skipped to five different paragraphs and each one ended with a "!x but y" logical statement, just formatted differently most of the time. Crazy how you can't unsee it.

A sibling [dead] comment to mine is a rebuttal to "just post the prompt", where it itself was expanded to several paragraphs that each say nearly nothing, including this gem:

> "That’s not a critique of the writing. It’s a diagnosis"

I miss when people just typed their thoughts concisely and hit send without passing it to an inflater. I'd maybe have a chance of understanding the sibling comment's point.


Now it is a tell but eventually people may natutally start speaking like this!!

This isn't mind control, just language evolution quiety nudged by AI. ;)


I've rather suspected that people are subconsciously adapting their language patterns to those that they hear over and over, and with AI content so prolific online now, it's natural that people are being programmed more and more with those patterns.

We trained a model on human language which is now in the business of itself retaining human language.


Yeah it's becoming increasingly obvious now. The moment I see this "contrast framing" I stop reading.

I read this as 'contrast farming' and like the term better.

That's not just contrast framing. It's contrast farming.

This made me think of:

- use an LLM to compress a blog article into a singular prompt

- Run it through against all the major LLMs to have them expand it back out again

- Diff the original against the generated versions in terms of content/ideas

- Spit out an "entropy ranking".


Again I'm the author.. and I actually wrote this.

Maybe I need to stop reading AI posts and thinking that it's good writing? But will take the feedback on board and try not to over analyse each paragraph - it was a lot messier upon first draft but maybe ready better I guess.


I appreciated the essay, GPTisms aside. The core concept is one that I've felt for a long while, but you articulated it more cleanly than I've been able to. I'd explained it as into "the tasks are replaced, the job is not", but I like your distinction better.

Appreciate it! Glad you enjoyed the read.

Planning on writing a lot more this year.


Thanks for saving me from reading it myself.

You really learn how bad everyone's taste is, just how low the bar can be when it comes to the written word. LLMs really can teach us something about ourselves!

we just need to send the article back to the LLM to get it synthesized /s

https://github.com/timescale/pg_textsearch

BM25 index for postgres, mostly written by a single (very smart) guy and Claude Code.


Written by a nobody of course:

> T.J. Green is a Senior Staff Engineer at Tiger Data, creators of TimescaleDB, where he is the implementor of pg_textsearch, a new Postgres extension for high-performance BM25 ranked text search. He brings deep expertise in database systems internals and indexing methods to this project. At Tiger Data, he has also contributed to pgvectorscale, the company's vector search extension for Postgres.

The guy could have written it eyes closed.


Yes TJ is very experienced and smart. But no he would not have been able to build this in a few months by himself without AI. In fact he uses Claude so much that he goes over the max subscription and spends thousands of dollars each month on tokens to continue delivering this fast.

AI is a big multiplier for experienced folks like him.


Incredible.

If I get production ready bm25 and incremental view maintenance in postgres, it really will do everything I care about.

Especially with the browser/client synch maintenance stuff people are working on.


I feel like we should pin this or something as a canonical counter to the now excruciatingly tired claim:

“Generative software code doesn’t work”


I like this prediction and it would be a good fit. Vercel can also monetize existing traffic much more broadly than tailwind can with just tailwind plus.


They do however often ask friends and family to upvote and leave comments. Like "we have been very happy using XYZ" which is against HN rules but not that strictly enforced. I feel like it is extra-lax towards YC companies but maybe I'm imagining things.


On the contrary, we particularly tell YC founders not to do this—mostly privately, but there's a public version of it at https://news.ycombinator.com/yli.html (scroll down to "Comments" and see the part in bold. That's me trying to scare them.)

We do tend to be more lenient when there's no evidence of organized manipulation, just friends/fans/users trying to be helpful and not realizing that it's actually unhelpful. What dedicated HN users tend not to realize is that such casual commenters usually have no idea how HN is supposed to work.

But this leniency isn't YC-specific. We're actually less lax when it comes to YC startups, for several reasons.


Okay I have heard different things from some friends that are YC founders and saw them do it in practice (e.g. posting about their HN post on LinkedIn).

I'm not going to out people here but maybe it helps you to know that not everyone plays by the rules. Tbf I also understand that this is just really hard to enforce.


This is a great way to get your posts buried. We have to warn well-wishers not to do this (autonomously) before we post anything major, from bitter experience.


This sounds cool but I feel like I need to often run the code in one way or another when verifying what Claude does. Otherwise it feels like driving blind. Claude Code already has the web version which I could use from my phone and fair it can't run scripts etc which limits the output quality. But if I can't verify what it did it also limits how long I can let it run before I need my laptop eventually.

Ofc if you have demo deployments etc on branches that you could open on mobile it works for longer.

Another issue is that I often need to sit down and think about the next prompt going back and forth with the agent on a plan. Try out other product features, do other research before I even know what exactly to build. Often doing some sample implementations with Claude code and click around these days. Doing this on a phone feels... limiting.

I also can't stand the constant context switching. Doing multiple feature in parallel already feels dumb because every time I come from feature B to A or worse from feature G to E it takes me some time to adjust to where I was, what Claude last did and how to proceed from here. Doing more tasks than 2 max. 3 in parallel often ends up slowing me down. Now you add ordering coffee and small talk to the mix and I definitely can't effectively prompt without rereading all history for minutes before sending the next prompt. At which point I might have also opened up my laptop.

Ofc if you truly vibe code and just add feature on feature and pray nothing breaks, the validation overhead and bar for quality goes down a lot so it works a lot better but the output is also just slop by then.

I typed this on my phone and it took 20 minutes, a laptop might have been faster.


I found wisprflow (paid app) to do a great job at smart-dictating my notes / emails.


We have a lot of happy customers that moved from rds to tiger cloud if you think pgvectorscale is interesting to you and you don't want to self host pg.

But yes big cloud providers move slow in adopting extensions.


No


Combined with our other search extension for full text search these two extensions make postgres a really capable hybrid search engine: https://github.com/timescale/pg_textsearch


I'm not how you'd combine the two; care to give us a quick outline ?


We have docs on how to do hybrid search here: https://www.tigerdata.com/docs/use-timescale/latest/extensio...

Essentially you combine the pgvector score and the bm25 score to hopefully get better results.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: