Thanks for posting this, very timely as I'm also playing around with pgvector for semantic search. I saw that you ended up trimming inputs longer than 8K tokens. Have you looked into chunking (breaking input into smaller chunks and doing vector search on the chunks)? Embedding models I'm playing with have a max of 512 tokens, so chunking is pretty much a must. Choosing a chunking strategy seems to be a deep rabbit hole of its own.
I was a Slashdot regular for around 10 years. Then somewhere around 2008/9 I realized that almost all of the stories appearing on the /. front page had featured on this new site called Hacker News days earlier, minus the random trolling and "frist post" spam. I briefly used both and ended up switching to HN entirely. Still my daily source of tech news.
Yeah, from my perspective this was just a dog and pony show. Trump and Vance got to look big and strong (their perception anyways) next to Zelensky, and he got to look like he's willing to continue the conversation. Doubt anyone actually expected progress out of this.
I would lose a healthy dose of respect for you if you trotted one of these out in a conversation with me in an attempt to convince me of something. As another comment mentions, argument from authority is literally a logical fallacy (https://en.m.wikipedia.org/wiki/Argument_from_authority). Make an argument based on knowledge and facts, and be ready to accept opposing viewpoints, rather than browbeating the other party into submission. There are times when a decision is already made and disagree-and-commit is desirable, but there are more tactful approaches than just throwing your seniority around.
Oh that's interesting. Explains the other common diminutive of Vladimir - Volodya. And yeah, Vlad is short for Vladislav.
My favorite set of Russian diminutives that makes zero logical sense to a non-native is: Alexander (formal) = Sasha (informal) = Shura/Shurik (very informal).
The rates are believable, but this seems somewhat different from what the thread is about - compensation in tokens rather than equity. I suspect many an engineer who's OK with crypto work in general will balk at this arrangement.
EDIT: I reread the original comment. It technically makes sense, but still I think answers a different question than the one posed about the compensation model.
The suspicious part is that ChatGPT did not have any kind of answer when the parent poster checked before posting. Yet soon after my reply it suddenly not only had the answer, but also used many of the same words and had the exact same timestamp. A little too close to be a coincidence, IMO.
Not just Google. A couple weeks back I saw pretty strong evidence of OpenAI ingesting an HN comment of mine and regurgitating it in ChatGPT within 10-15 minutes of posting.
Once I had to track down an issue where very rarely, with no discernible pattern the web app would produce garbled PDFs. Turned out this happened when an admin account remotely connected to the app server, which caused a reset of default screen resolution, which messed up the PDF library that relied on a specific resolution (it was HTML-to-PDF conversion). Happened rarely and randomly because there were multiple web servers and they were occasionally restarted which would fix the problem until next time.
Another fun problem I dealt with was when I was moving my employer's codebase from Subversion to Mercurial version control ages ago. Everything looked good, except a directory named CVS (after the pharmacy, a customer) was missing. Was banging my head on the table before realizing that the default .hgignore file instructed Mercurial to ignore all contents of .*/CVS (another old version control system).
I can believe it with dogs, but my cat of almost 20 years has never indicated any understanding of or interest in human speech. She sometimes reacts to tone of voice, sometimes to specific noises (like treat baggie crumbling when opened), but not any particular words. Half the time she could care less about the tone or noises as well. That seems pretty in line with my general understanding of cat "psychology".
I really don't understand the product decisions at Spotify. Some really basic functionality is missing like shuffling across all tracks in my library or treating Liked Songs as a playlist but not really (can't share with others). I keep using it for its discovery features, but even that is so much worse than ten years ago. Back then I would easily keep half of Discover Weekly. These days I'm happy if I find one or two songs I like. They feel like a company on auto-pilot.
Without going into specifics, trying to classify a body of freeform text (or rather, many such bodies) along a bunch of dimensions. Imagine a fiction story that you're trying to answer questions about such as "what is the genre?" and "is the protagonist male or female, adult or child?", etc, where I'm instructing the LLM to return values out of a specific range only (aka, enums). OpenAI's new-ish structured outputs feature enforces answers to a user-specified JSON schema, which is awesome, but I'm finding that the actual answers it comes up with are sub-par. There are all kinds of hacks out there like creating additional fields to let it reason out the answers before constraining them to an enum elsewhere, but I have yet to make it work well for me. There are other approaches I haven't yet tried that require much more effort or cost (more expensive models, fine-tuning custom models). It feels like it should work, so the random poking and prodding continues...
I've been playing around with data extraction using OpenAI's structured outputs feature last week. Perhaps this is due to me still being fairly new to this world, but man, has it been an exercise in frustration. My general takeaway so far is that LLMs are fantastic for creative writing, and are a much harder beast to tame when you're looking for exact, reproducible answers. Even when those answers should be the most obvious ones. My engineering brain does not appreciate all the prompt engineering hacks, the hopeful JSON schema tweaking, the "let's try this other random thing just to see if it works" approach to making progress. The worst has been getting to an OK point with one of the fields, only to find out later that giving directions for another field blew away the earlier progress. I'm sure I must be missing something since there are products that are doing this, but so far I have little to show for hours spent other than growing knowledge of LLM quirks and subtle behaviors.
EDIT: Looks like the parent post entered the second chance pool (https://news.ycombinator.com/item?id=26998308). Not only did all the timestamps get rewritten, but apparently I can now edit this day-old comment :) Interestingly, it does show the correct timestamp when editing. Off-topic, but thought it's interesting behavior worth mentioning.
+1 for One Night Ultimate Werewolf. Removing the need for a narrator role and keeping the game short enough that the antagonistic behaviors don't have a chance to develop works wonders. Friend and couple fights after Mafia are real. Werewolf is pretty kid-friendly too.
Wow, haven't seen this before, thank you. Amazing that the writing can still be read by a modern reader (that said, really helps to know what it's supposed to say though). The note I found most relatable is the one with greetings to his classmate.
Say I generate embeddings for a bunch of articles. Given the query "articles about San Francisco that don't mention cars" would cosine similarity uprank or downrank the car mentions? Assuming exclusions aren't handled well, what techniques might I use to support them?
Just a couple of months ago you posted https://news.ycombinator.com/item?id=41983089, which makes it pretty clear (together with your earlier submissions/comments) that you're not anywhere close to 18 or that this is your first app.
Rands Leadership Slack is a fantastic and supportive community targeting Engineering and Product leaders primarily, but not exclusively. At around 35K members right now, it has channels on just about every topic. It's invite-only, but super easy to get invited.
Haha, you're very welcome. The ChatGPT thing is odd in how close the answer was to my comment, including specific words I used. Makes me wonder about the likelihood of it being scraped and processed by OpenAI in the 10-15 minutes between posting the comment and trying it in ChatGPT. Or perhaps our brains are just LLMs after all :)
Oh hey, that totally sounded familiar :) Pretty sure this is from Konyok Gorbunok (The Little Humpbacked Horse), a 1975 Soviet cartoon based on a famous Russian fairy tale. The bit you're describing is at 1:07:30 of https://youtu.be/fKc22eSL1gA.
This doesn't quite fit several of the points you remember (very much in line with the post!), so perhaps it was some other edition of that same story.
EDIT: so, I just plugged your description into ChatGPT and it gave the exact same answer, including an identical timestamp! Weird.